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.0001946 | Leishmaniasis Direct Agglutination Test: Using Pictorials as Training Materials to Reduce Inter-Reader Variability and Improve Accuracy | The Direct Agglutination Test (DAT) has a high diagnostic accuracy and remains, in some geographical areas, part of the diagnostic algorithm for Visceral Leishmaniasis (VL). However, subjective interpretation of results introduces potential for inter-reader variation. We report an assessment of inter-laboratory agreement and propose a pictorial-based approach to standardize reading of the DAT.
In preparation for a comparative evaluation of immunochromatographic diagnostics for VL, a proficiency panel of 15 well-characterized sera, DAT-antigen from a single batch and common protocol was sent to nine laboratories in Latin-America, East-Africa and Asia. Agreement (i.e., equal titre or within 1 titer) with the reading by the reference laboratory was computed. Due to significant inter-laboratory disagreement on-site refresher training was provided to all technicians performing DAT. Photos of training plates were made, and end-titres agreed upon by experienced users of DAT within the Visceral-Leishmaniasis Laboratory-Network (VL-LN).
Pre-training, concordance in DAT results with reference laboratories was only 50%, although agreement on negative sera was high (94%). After refresher training concordance increased to 84%; agreement on negative controls increased to 98%. Variance in readings significantly decreased after training from 3.3 titres to an average of 1.0 titre (two-sample Wilcoxon rank-sum (Mann-Whitney) test (z = −3,624 and p = 0.0003)).
The most probable explanation for disagreement was subjective endpoint reading. Using pictorials as training materials may be a useful tool to reduce disparity in results and promote more standardized reading of DAT, without compromising diagnostic sensitivity.
| Until the 1990's accurate Visceral Leishmaniasis (VL) diagnosis necessitated parasitological confirmation by microscopy or culture of the blood, bone-marrow, lymph nodes or spleen. These techniques are invasive and splenic aspirates are associated with a risk of serious bleeding. This has led to the development of non-invasive serological tests such as the direct agglutination test (DAT). During infection with VL, circulating antibodies are produced against the surface antigens of the invading parasites. The DAT detects antibodies to L. donovani s.l. in the blood or serum of those infected by means of direct agglutination. In the absence of antibodies to Leishmania the DAT antigen accumulates at the bottom of the plate to form a dark blue spot. If antibodies to Leishmania are present then the antigen forms a pale blue film over the well constituting a positive result. Here, we report on shared experiences of six endemic countries using DAT to characterize performance panel samples. There was considerable inter-reader variability and in order to standardize the reading of DAT we developed and implemented pictorial training aids. After refresher training, agreement between readers increased; the pictorial aids and recommendations for using DAT are available in this article.
| Up until the 1990's accurate visceral leishmaniasis (VL) diagnosis necessitated parasitological confirmation by microscopy or culture of the blood, bone-marrow, lymph nodes or spleen [1]. Microscopic detection of parasites in clinical material from the spleen is still considered the reference standard; however, splenic aspirates are associated with risk of serious bleeding and should only be carried out in settings with access to blood transfusion and surgical services. The invasiveness and potentially fatal complications associated with splenic aspiration has spurred the development of non-invasive serological tests such as direct agglutination test (DAT) [2] over 25 years ago and in the past decade, lateral flow immuno-chromatographic tests (ICT), commonly referred to as rapid diagnostic tests (RDTs). RDTs have now been adopted widely, in the Indian subcontinent [3], but in other endemic regions, DAT is part of the diagnostic algorithm or is used for epidemiological surveys due to variable sensitivity of RDTs [2], [4].
The DAT, in its present form, is a freeze dried suspension of trypsin-treated fixed and stained culture of L. donovani promastigotes [5]; liquid formulations of DAT are also manufactured locally. During infection with VL, circulating antibodies are produced against the surface antigens of the invading parasites. The DAT detects antibodies to L. donovani s.l. in the blood or serum of those infected by means of direct agglutination. In the absence of antibodies to Leishmania the DAT antigen accumulates at the bottom of the plate to form a dark blue spot. If antibodies to Leishmania are present then the antigen forms a pale blue film over the well and this constitutes a positive result.
DAT requires moderate technical expertise, and laboratory equipment and reagents, including calibrated pipettes, micro-titre plates, multiple reagents and a toxic solution (chemical 2-beta Mercapto-ethanol (2-ME)) [2]. Furthermore, despite very good accuracy, inter-observer discrepancy in routine DAT serology readings is common [6]–[8].
Prompted by shared experiences of six endemic countries using DAT to characterize performance panel samples, we report an assessment of DAT inter-reader variability. It was noted that the inter-laboratory agreement of DAT titres on a panel of 15 sera was low. Here, our objective was to standardize the reading of DAT by developing and implementing pictorial training aids.
Nine laboratories from three global endemic regions were involved in a WHO/TDR-sponsored evaluation of VL RDTs; namely Asia (n = 4), South America (n = 2) and Eastern Africa (n = 3) (Table 1). The Institute of Tropical Medicine, Antwerp, Belgium (ITM) assembled a proficiency panel including sera from 10 VL confirmed patients including a range of DAT titres, and 5 VL negative patients, one healthy endemic control and others who harbored potentially cross-reacting, infections, including Chagas disease, tuberculosis, malaria and leprosy. Prior to shipping, each sample within the panel was assigned a random numerical code that varied from centre to centre.
All samples were left over from samples that had been taken as part of research projects conducted between 1978 and 2000 at the Institute for Tropical Medicine (ITM) Antwerp, Belgium. The samples were anonymised and kept stored for future use for scientific purposes. In the studies conducted since 2000 explicit consent was asked for storage and future use of left overs of the samples that were taken. In the older studies no explicit mention was made of future use of the stored left overs though a general informed consent was asked. However, it was not possible to trace back the study participants in the studies preceding 2000 and to ask them for informed consent for storage and use of left over samples
The proficiency panel was tested blindly using the DAT assay (KIT-Biomedical Research, Lot 0904) in each of the nine evaluation laboratories (Table 1) and both reference laboratories (KIT and ITM). Results were returned electronically to ITM using a standard recording form. All microtitre plates used in the procedure were provided by reference laboratories (Greiner 651101 100). The DAT was performed as described previously [2].
Due to significant discordance in end-titres between all laboratories, photographs of DAT plates with 10 VL positive serum samples and 5 VL negative serum samples were prepared by the reference laboratories (following joint agreement on end titres) and were used as pictorial training aids. Refresher DAT training was given by staff of KIT and Banaras Hindu University (BHU) to all participating laboratories (Table 1). Trainers assessed equipment and compliance with the DAT SOP, including preparation of reagents. Subsequently, the proficiency panel was repeated in the presence of the trainer. End titres were read independently by two separate technicians and the trainer. When readers did not agree on the end titre they came to a common conclusion after joint discussion. Combined results of the readers were sent to ITM and decoded by a study team member not involved in the refresher training; results of the trainer were not taken into consideration unless the results of the readers were significantly different from those of the reference laboratories and the test was repeated. Disagreement was defined as greater than one titre above or below those of the reference laboratories [6]. Variance in results before and after refresher training was compared with a two-sample Wilcoxon rank-sum (Mann-Whitney) test.
Despite having received the same panels, batch of DAT, microtitre plates and protocol, overall DAT results concordance (agreement within one titre) with the reference laboratories was only 50%. Agreement on negative controls was very good (94%). Using a cut off of 1∶1600 serum dilution, the pre-training sensitivity and specificity were 79% and 94%, respectively.
Refresher training was initiated due to the large differences in DAT reading between participating laboratories. Here, photographs of DAT plates were used as training aids, where end-titres had been agreed upon by the reference laboratories. During refresher training the trainers did not identify any faulty or inappropriate equipment, nor did they witness any non-compliance with the DAT SOP. After refresher training the concordance (agreement with one DAT titre) increased to 84% with the reference laboratories. The agreement on negative controls increased to 98%. Average variance in results before refresher training was 3.3 titres; this improved to an average variance of 1.0 titre reading (the accepted limit) after refresher training. A non-parametric test was used to test for significant differences before and after training using a two-sample Wilcoxon rank-sum (Mann-Whitney) which showed significant difference (z = −3,624 and p = 0.0003). Post-training the sensitivity increased to 97% and the specificity to 100% (cut off values 1∶1600).
Overall, the refresher training increased the operator performance of the DAT in this small proficiency panel (Table 2). After refresher training a cut-off point of 1∶1,600 (serum dilution) gave 97% sensitivity (CI: 91.6–99.0%) and 100% specificity.
Further, pictorial guides (Figures 1, 2, and 3) of DAT training plates reflecting consensus end titres by several experts in the VL-LN, with many years of experience in using DAT as a diagnostic tool, are now available. Further recommendations to be taken into account for completion of the DAT assay are highlighted in Table 3 and data per laboratory pre-training and post-training with pictorial aids can be seen in Supporting Information S1.
The DAT assay has been used as a diagnostic tool for more than 25 years, it is robust, reliable has a high clinical accuracy and can be performed in laboratories with minimal equipment. However, the subjective manner in which the result (end-titre) of the test is read means that inter-reader variation in titre reading can be an issue. Preparations for a multicenter evaluation of RDTs unexpectedly uncovered a significant discordance in DAT results among reference and evaluation centres; this presented an opportunity to address discordance and create an international, consensus-based protocol and training materials to strengthen standardized reading of the DAT for VL diagnosis without compromising diagnostic accuracy.
The reasons for all of the discrepancies between the different laboratories is not fully understood however, it was noted that technicians were generally competent in the DAT procedure, particularly those who used it as part of their routine diagnostic algorithm. It was not possible to test the saline that the laboratories previously used in testing, but it is possible that the origin and quality of the saline solutions used as a diluent for the DAT antigen did affect performance, generating false positive precipitation in negative sera wells. The most consistent problem identified in the laboratories can be attributed to the subjective manner in which the end titre of the DAT test is typically read. Some readers record the end-titre when 50% of agglutination of the well has occurred (as occurs with other agglutination tests), whilst other readers record the end-titre where the whole well has agglutinated and there is no difference between a negative control well (antigen plus saline) and the sample well. Even though 1 titre difference in reading is considered acceptable [6], the discrepancy and variance in results reported here was far greater. Training plates developed by the reference laboratories proved to be extremely helpful in illustrating the end titre. Positive sample wells were defined by any reaction in the test well in comparison to the negative control well; this ensured that the high sensitivity of the DAT was not compromised. Figure 2 shows the end-titre as agreed by the VL-LN; a follow up plate can be used to test users before revealing the results as seen in figure 3. High quality photos in figures 2 and 3 are also available by request (contact corresponding author) for use as reference training material for future DAT users.
Since slight variations in readings between different DAT antigen batches may occur it is advised that the same batch of DAT should be used within one project or epidemic to decrease variability in results. If this is not possible then it is recommended to keep reference sera in order to assess this lot-to-lot variation, this should not be more than one titre difference. In addition, it is important that all users of the DAT specify the type of dilution used, i.e. serum dilution (starting 1∶100) or antigen plus serum dilution (starting 1∶200). It is likely that cut-off values are different between endemic areas and even during epidemic cycles. Local guidance as to appropriate cut-off values is essential.
The problems uncovered during a multicenter DAT proficiency testing scheme are potentially relevant to other DAT users. To reduce inter-reader variability and increase accuracy, photos of training plates were made, and end-titres were agreed upon firstly by the reference laboratories and subsequently by experienced users of DAT within the VL-LN. These photos can be used to promote a standardized approach to interpreting DAT without compromising sensitivity. Protocols and photos can be requested for training and quality control purposes by two of the major manufacturers of the assay, KIT and ITM. High sensitivity and specificity can be achieved with this reliable and robust diagnostic tool, and we hope that provision of good training materials can increase the usefulness of DAT.
|
10.1371/journal.pcbi.1002322 | Evidence for Sequential and Increasing Activation of Replication Origins along Replication Timing Gradients in the Human Genome | Genome-wide replication timing studies have suggested that mammalian chromosomes consist of megabase-scale domains of coordinated origin firing separated by large originless transition regions. Here, we report a quantitative genome-wide analysis of DNA replication kinetics in several human cell types that contradicts this view. DNA combing in HeLa cells sorted into four temporal compartments of S phase shows that replication origins are spaced at 40 kb intervals and fire as small clusters whose synchrony increases during S phase and that replication fork velocity (mean 0.7 kb/min, maximum 2.0 kb/min) remains constant and narrowly distributed through S phase. However, multi-scale analysis of a genome-wide replication timing profile shows a broad distribution of replication timing gradients with practically no regions larger than 100 kb replicating at less than 2 kb/min. Therefore, HeLa cells lack large regions of unidirectional fork progression. Temporal transition regions are replicated by sequential activation of origins at a rate that increases during S phase and replication timing gradients are set by the delay and the spacing between successive origin firings rather than by the velocity of single forks. Activation of internal origins in a specific temporal transition region is directly demonstrated by DNA combing of the IGH locus in HeLa cells. Analysis of published origin maps in HeLa cells and published replication timing and DNA combing data in several other cell types corroborate these findings, with the interesting exception of embryonic stem cells where regions of unidirectional fork progression seem more abundant. These results can be explained if origins fire independently of each other but under the control of long-range chromatin structure, or if replication forks progressing from early origins stimulate initiation in nearby unreplicated DNA. These findings shed a new light on the replication timing program of mammalian genomes and provide a general model for their replication kinetics.
| Eukaryotic chromosomes replicate from multiple replication origins that fire at different times in S phase. The mechanisms that specify origin position and firing time and coordinate origins to ensure complete genome duplication are unclear. Previous studies proposed either that origins are arranged in temporally coordinated groups or fire independently of each other in a stochastic manner. Here, we have performed a quantitative analysis of human genome replication kinetics using a combination of DNA combing, which reveals local patterns of origin firing and replication fork progression on single DNA molecules, and massive sequencing of newly replicated DNA, which reveals the population-averaged replication timing profile of the entire genome. We show that origins are activated synchronously in large regions of uniform replication timing but more gradually in temporal transition regions and that the rate of origin firing increases as replication progresses. Large regions of unidirectional fork progression are abundant in embryonic stem cells but rare in differentiated cells. We propose a model in which replication forks progressing from early origins stimulate initiation in nearby unreplicated DNA in a manner that explains the shape of the replication timing profile. These results provide a fundamental insight into the temporal regulation of mammalian genome replication.
| Eukaryotic chromosomes replicate from multiple replication origins that fire at different times in S phase [1]–[3]. In the yeast S. cerevisiae, microarray analysis of replicating DNA isolated from cells progressing synchronously through S phase first demonstrated that each region of the genome replicates at a reproducible mean time [4]. Similar findings have been reported for other eukaryotes including mammals [5]–[14]. The reproducible replication time might be interpreted to reflect a deterministic replication timing program, with replication origins located at specific positions firing at specific times in S phase. However, other methods had revealed that origins are often inefficient, firing in only a fraction of cells and being passively replicated by a fork emanating from another origin in other cells [15], [16]. Furthermore, single-molecule analyses of chromosomal replication intermediates showed that both time and order of origin firing are extremely variable so that no two cells use the same pattern of origin firing [17], [18]. These results favored a stochastic model for chromosomal replication where origins fire independently of each other and the mean replication time of each region is an ensemble average that only reflects the variable firing efficiencies of the surrounding origins [19]. Numerical simulations suggested that such models are compatible with the existing replication time course and origin efficiency data in yeast [20]–[22].
On the other hand, studies performed mostly in metazoan cells suggested that replicons are arranged in functional groups [23]. DNA fiber techniques revealed that adjacent origins are organized as clusters that often fire at similar times [24]–[30]. Intra-nuclear labeling of replication sites revealed discrete sites, or replication foci, that appear to contain multiple adjacent replicons and to correspond to stable structural units of both interphase and mitotic chromosomes [27], [31]–[34]. Furthermore, foci that replicate during consecutive time intervals are often spatially adjacent in nuclei and correspond to adjacent replicon clusters along chromosomes [35]–[40]. Therefore, origin clusters may correspond to stable structural entities that become available for efficient replication initiation at specific times in a sequence that depends on their order along the chromosomes. A study of the mouse immunoglobulin heavy chain region revealed a 0.4 Mb temporal transition region (TTR) that connects an early and a late replicating domain and is replicated by a single fork progressing in a unidirectional manner [41]–[43]. Studies of genome-wide replication profiles suggested that the dichotomy between 0.2–2 Mb domains containing multiple synchronous origins and 0.1–0.6 Mb originless TTRs that replicate in a unidirectional manner is a general feature of mammalian chromosome organization [8], [9], [11], [14], but the possibility that there is a gradual activation of origins in TTRs has also been considered [44].
Here we have performed a quantitative analysis of DNA replication kinetics using a combination of DNA combing data, genome-wide replication timing data and origin mapping data generated in this work or in previous studies in several human cell lines, as summarized in Table 1. We find that a large fraction of TTRs replicate at an apparent speed compatible with unidirectional progression of a single fork in embryonic stem cells. However, in differentiated cells or in cancer cells, most if not all TTRs replicate significantly faster than predicted by unidirectional progression of a single fork. Origins are activated synchronously in regions of uniform replication timing and more gradually in TTRs. We discuss how these findings may be reconciled with a stochastic model for replication timing. We propose an alternative domino model for origin activation in which replication forks progressing from early origins stimulate initiation in nearby unreplicated DNA and the space/time intervals between consecutive initiations explain the observed range of apparent replication speeds.
We used DNA combing [45], [46] to measure replicon size and replication fork progression rate in HeLa cells at different stages of S phase (Figure 1). Asynchronously growing cells were pulsed with the halogenated nucleotide IdU for 20 min followed by CldU for another 20 min, fixed and sorted into four temporal compartments of S phase (S1, S2, S3 and S4) according to total DNA content. DNA was stretched on coverslips by combing and total DNA was stained in red with an anti-DNA antibody. The replicative labels were revealed in blue (IdU) and green (CldU) using appropriate antibodies. The blue-to-green transitions show the position and orientation of mobile forks at the time CldU was added (Figure 1A).
Replication fork velocities were determined by measuring the length of CldU or IdU tracts that could be unambiguously assigned to the progression of a single fork during an entire 20 min labeling interval. Fork velocities were narrowly distributed around a mean of 0.68 kb/min, with almost no values >2 kb/min, and did not change throughout S phase (Figure 1B).
The global density of replication forks (total number of forks divided by total length of DNA examined, corrected for contamination by non-replicating G1 or G2/M cells and for replicated genome fraction; see Material and Methods) increased through S phase (from 2.64 to 3.88, 4.55 and 5.4 forks per Mb in S1, S2, S3 and S4, respectively; Table 2). The measured replication fork densities and velocities were used to calculate the time required to duplicate the entire genome (see Material and Methods). The result (6 h 18 min) was consistent with the length of S phase independently measured from the cell doubling time and the fraction of the sorted cells in S phase (22 h×1/3 = 7 h 20 min), corroborating the fork density and velocity measurements.
The inverse of the global fork density is the global fork-to-fork distance (FTFD). The global FTFD decreased from 379 kb to 258, 220 and 185 kb in S1, S2, S3 and S4, respectively. However, the local FTFDs measured on single DNA fibers containing forks were much smaller (mean ∼19 kb) and did not decrease so much during S phase (from 22.0 kb in S1 to 17.0 kb in S4; Figure 1C; Table 2). Furthermore, the mean intra-fiber inter-origin distances and inter-termini distances were commensurate with the intra-fiber FTFDs, i.e. both were in the 35–42 kb range at all stages of S phase (Figure 1D, E; Table 2). Thus, replicons were much shorter than global FTFDs would suggest. The discrepancy between local and global FTFDs might be attributed to the finite fiber size, which prevents measurement of large FTFDs, but actually results from the fact that origins are activated as clusters that fire at different times in S phase. Thus, only 10–20% of all fibers showed replication forks at any stage of S phase but among these many showed several forks (Figure 1F). To assess the clustering of replication forks, we compared the distribution of the number of forks per fiber with that generated in a simulation that assumed random initiation and a fiber size distribution and global fork density identical to the experimental samples. The observed distributions were significantly (P<10−4) different from the simulation, with a lack of fibers with one fork (whole S-phase average, 6.3% vs. 10.6%) and an excess of fibers with ≥2 forks (8.5% vs. 2.2%). This demonstrates a clustering of origin firing.
We next examined whether the global fork density increased because more origin clusters fired or because more origins per cluster fired during S phase. We found that the number of forks per fork-containing fiber (2.33, 2.54, 2.94 and 3.31 forks per fiber in S1, S2, S3, and S4, respectively; Figure 1F; Table 2), and the IdU-labeled fraction of fork-containing fibers (30.1% , 44.3%, 41.9% and 52.0%; Figure 1G; Table 2) increased throughout S phase. Thus, more origins per cluster fired as S phase progressed. The distances between origin clusters are generally too large to be measured, because they exceed the mean fiber size. Although such distances cannot be individually measured, their mean can be computed from the statistics of fibers with and without forks (by dividing the total length of DNA minus the sum of intra-fiber FTFDs by the number of fork-containing fibers, assuming at most one cluster per fiber). Note that intercluster segments mainly consist in unreplicated DNA in early S phase and already replicated DNA in late S phase and that the total DNA length used in our calculations is corrected for the extent of DNA replication (see Material and Methods). The mean intercluster distance decreased from 772 kb in S1 to 484, 465 and 501 kb in S2, S3 and S4. Thus, inter-cluster distances were reduced as S phase progressed from S1 to S2 but did not change thereafter. This reduction was too large to be explained by the increase in cluster size. Therefore, the number of active clusters increased from S1 to S2.
To further evaluate the tightness of origin synchrony we reasoned that the consecutive IdU/CldU labeling scheme allows us to distinguish origins that fired before (type 1) or after (type 2) CldU addition. Type 1 origins are flanked by two divergent blue-to-green transitions whereas type 2 origins give rise to doubly-labeled, isolated tracks. For example, most origins shown on Figure 1A fired before CldU addition but the leftmost one fired after CldU addition. We first noticed that when inter-origin distances were plotted separately for type 1 and type 2 origins (not shown), their distributions were not markedly different from those shown on Figure 1D, where all origins were taken into account. This suggested that type 1 and type 2 origins were not frequently interspersed with each other. We then selected fibers containing more than one origin and found that adjacent origins were significantly more frequently of the same type than if randomly interspersed (254 type1/type1; 60 type2/type2; 99 type1/type2; P<10−4, chi-square test of homogeneity). Thus, adjacent origins tended to fire within 20 min of each other. Together these observations suggested that a wave of initiations propagates on the DNA molecule.
In conclusion, DNA combing showed that in HeLa cells i) replication origins are spaced at mean ∼40 kb intervals; ii) adjacent origins fire within 20 min of each other, resulting in a spatial clustering of replication forks; iii) replication fork velocity (∼0.68 kb/min) does not change during S phase; iv) the global fork density increases during S phase, because more replicon clusters and more origins within clusters become active as S phase progresses. Therefore, the global rate of DNA replication increases during S phase due to increasing origin synchrony.
Our conclusion that fork speed is constant through S phase contrasts with earlier reports of changes in fork speed during S phase [47], [48]. However, in these studies, chemicals or serum starvation were used to synchronize cells, which may affect nucleotide pools and replication fork progression, whereas the retroactive (FACS) synchronization we used does not perturb the cell cycle. Furthermore, these studies used less precise techniques than DNA combing to spread DNA fibers, and some of the track length changes interpreted as changes in fork progression may in fact have resulted from changes in the synchrony of adjacent origins and consequent merging of forks. We have minimized such potential artifacts thanks to the use of two short, consecutive labeling pulses and the better resolution of DNA combing, which allowed us to demonstrate an increase in adjacent origin synchrony during S phase.
The fork speed (∼0.7 kb/min) and interorigin distance (∼40 kb) we found are somewhat, though not much, lower than usually reported in other human cell lines (typically 1.0–2.0 kb/min and 100–200 kb) [49]. Small interorigin distances (57 kb) and slow forks (0.37 kb/min) have also been found by DNA combing in K562 leukemic cells [50]. Small interorigin distances were also reported using another DNA fiber technique both in U2OS osteocarcinoma cells (50 kb) and in nontransformed MRC5 cells (42.5 kb) [51]. More intriguingly, our estimates also differ from those reported by other investigators in HeLa cells (fork rates of 0.59–1.37 kb/min [52] and 1.7±0.3 kb/min [27] and interorigin distances of 144±66 kb [27]). In yet another HeLa clone (data not shown) we observed slightly larger replicons (50 kb) and faster forks (1.0 kb /min) than in this work. Thus, clonal variation as well as differences in labeling scheme, DNA fiber technique and track choice probably explain these differences.
Such clonal variation is consistent with the possibility that the cancerous nature and genetic or epigenetic instability of HeLa cells influence origin activity and fork progression and their response to a number of physiological and pathological stimuli [53]–[55]. Indeed, recent work showed that forced expression of oncogenes in primary keratinocytes can slow down replication fork progression and trigger activation of dormant origins due to decreased nucleotide pools [56]. However, in another study, no change in origin spacing and fork velocity could be observed between primary keratinocytes and a keratinocyte-derived tumour cell line [57]. Thus, it remains possible that the fork speed and origin spacing observed in HeLa cells just reflect some physiological tissue variation range.
We generated a high-resolution, genome-wide replication timing profile in HeLa cells as described previously [12] with minor modifications detailed in the Material and Methods. Briefly, HeLa cells were pulsed with BrdU, sorted into four temporal compartments of the S phase and nascent DNA was immunoprecipitated with anti-BrdU antibodies and sequenced using the Illumina technology to yield a total of 50 million reads that mapped uniquely to the human genome sequence. The abundance of sequence reads along the genome was computed every 10 kb in a 100 kb sliding window in each S phase compartment allowing to cover 90% of the genome. The resulting profile was used to compute in each window the fraction of the S phase at which 50% of the DNA was replicated (S50, [12]). Using the FACS DNA fluorescence histogram to extract the proportion of cells at different stages in S phase and the temporal profile of the rate of DNA synthesis, we calculated the profile of DNA content as a function of the time spent by a cell in S phase. The S50 values were then used to deduce the time (TR50) at which a defined genome region had replicated in 50% of the cells (see Material and Methods). A biological replicate showed excellent reproducibility of the TR50 (Pearson R = 0.97, P<10−16). The average of the two TR50 determinations was used for subsequent analyses.
The genome-wide TR50 histogram (Figure 2A) showed a continuum of replication times with no dearth of replicating regions in mid-S phase and an increasing number of replicating regions during S phase. This is consistent with the increase in global fork density observed by DNA combing (Figure 2B) and the one of global rate DNA replication observed by flow cytometry analysis (Figure 2C). This is also consistent with the observed dip in the FACS histogram of DNA content from S1 to S3, due to cells moving faster through this DNA content (Figure S1A). The expected dip in S4 was not observed but this was due to the spreading of the adjacent G2 peak. The continuous dip from S1 to S4 was indeed visible in the post sort control, where the DNA content of sorted S1 to S4 cells was reexamined in a second round of sorting (Figure S1B).
The TR50 profile along the genome showed a landscape of peaks and valleys interspersed with flat domains of uniform replication time (Figure 3A shows an exemplary 15 Mb chromosomal segment; see Figure S2 for a whole-genome profile). The slope of replication timing profiles has often been taken as a measure of replication fork velocity. However, since replication timing profiles are population averages, this is only true for regions in which forks progress in the same direction in all cells. Here we demonstrate (see Material and Methods), as first proposed by de Moura et al [22] in a recent analysis of yeast replication timing profiles, that the derivative of the replication timing, dt/dx, depends not only on the fork speed, v, but also on the local proportion of rightward (R) and leftward (L) moving forks in the cell population, such that dt/dx = (R−L)/v. The apparent replication speed is defined here as the inverse of this derivative, dx/dt. Note that the equality dx/dt = v/(R−L) implies that the sign of the apparent replication speed indicates the predominant direction of replication progression and that in flat domains of uniform replication time (infinite apparent replication speed), forks move equally in both directions.
We performed a multiscale analysis of the apparent replication speed genome wide, using the continuous wavelet transform, a robust method to obtain a well defined and numerically stable measurement of the local slope of the timing profile at any scale of observation (Figure 3B; Figure S2). The replication speed modulus, |dx/dt|, critically depended on the measured segment scale dx. At very large scales (>2 Mb), the entirety of the genome appeared to replicate at >10 kb/min. At smaller scales, a differentiation of the genome into smaller and slower replicating segments was observed, revealing finer details of the replication profile. The landscape of replication speeds stabilized below the 100 kb scale, as expected from the spatial resolution of the profile. At this scale, a broad distribution of replication speeds was observed in the HeLa cell genome (Figure 3C), with 1% of 100 kb segments replicating at an apparent speed ≤2 kb/min, 53% in the 2–10 kb/min range, and 46% at >10 kb/min. We noted that the speed distribution was shifted toward higher speeds for S1 and S4 compared to S2 and S3 fractions (Figure 3D).
The observed range of apparent replication speeds cannot be explained by the range of single fork velocities measured by DNA combing in the same cells. The mean and maximum fork velocities are 0.68 kb/min and 2.0 kb/min, whereas 99% of the genome replicates at an apparent speed >2 kb/min. The possibility that regions with the slowest apparent replication speed are specifically replicated by the fastest forks seems unlikely since fork velocities at single loci usually show the same degree of heterogeneity as in the bulk genome (e.g. [58], and see below our data on the IGH locus). These results imply that in HeLa cells, |R−L|<1, i.e. replication forks move in both directions, in most of the genome and that the proportion of right and left forks varies widely along the genome. There is a complete gradation between regions where forks progress predominantly (if not exclusively) in one direction (steep timing gradient, apparent speed ≤vmax), and regions where they progress equally in both directions (flat timing gradient, high apparent speed). The apparent speeds must therefore reflect the statistics of origin activation around and within the timing gradients. Essentially similar results were obtained for several other cell lines (see below).
To address the mechanism by which different proportions of rightward and leftward moving forks are established in different parts of the genome in HeLa cells, we segmented the whole genome into constant timing regions (CTRs) replicating at >10 kb/min and timing transition regions (TTRs) replicating at ≤10 kb/min and analyzed them separately. Figure 4A–F shows the size distribution, genome coverage, TR50 and apparent replication speed of CTRs and TTRs defined at 100 kb (blue), 200 kb (green) and 500 kb (red) scales. At the 100 kb scale, the whole genome was segmented into 7548 CTRs and 7504 TTRs (Figure 3A; Figure S2). All CTRs were ≤2 Mb and 53.8%<100 kb (Figure 4A), with CTRs>100 kb covering 34.2% of the genome (Figure 4C). All TTRs were ≤900 kb and 64.4%<200 kb (Figure 4B), with TTRs>200 kb covering 32.4% of the genome (Figure 4D). At larger scales, as expected, the mean size of both CTRs and TTRs increased and the genome fraction covered by CTRs increased at the expense of TTRs. The TR50 distribution of CTRs was relatively insensitive to scale (Figure 4E) and was similar to that of the whole genome, but the apparent replication speed of TTRs increased with scale (Figure 4F). The small oscillations in the TR50 distribution of CTRs are an artifact of the finite number of S phase fractions, which we have not attempted to correct. The proportion of CTRs was higher in S1 (48%) and S4 (56%) than in S2 (28%) and S3 (32%), consistent with the fastest distribution of speeds in these two S phase compartments (Figure 3D).
One possible mechanism for explaining why an equal proportion of rightward and leftward moving forks replicate a CTR is that it does not contain origins and is passively replicated from an outside origin that is activated equally often on its right or its left side (Figure S3A). Given a mean fork velocity of 0.68 kb/min (40 kb/h) this mechanism could only apply to short enough CTRs (<300 kb) to replicate within a 7–8 h S phase in HeLa cells. At the 100 kb scale, CTRs<300 kb and >300 kb cover 19.7% and 21.2% of the genome, respectively (Figure 4C). This mechanism predicts that i) the edges of small CTRs would replicate asynchronously in non-adjacent S-phase compartments whereas their centers would replicate synchronously in mid-S phase; ii) that small CTRs lying between 150 and 300 kb would replicate rather in mid-S phase. A previous study of Hela cells replication timing determined that about 20% of the ENCODE regions present a pan-S replication profile [59]. However, we reported that in HeLa cells only 7.4% of all genomic sequences replicate with such a pan-S profile [12]. Although a significant correlation was observed between these two studies (Pearson, R = 0.77, P<10−15), the differences may result from the use of microarray hybridization and cell synchronisation by drug treatment in the first study vs. massive sequencing and no drug treatment in our study. Furthermore, we have found that the TR50 distribution of CTRs spans the entire S phase whatever their size (data not shown), inconsistent with the mechanism proposed above.
Alternatively, CTRs might consist of regions in which multiple origins are synchronously activated (Figure S3B). This mechanism would result in an equal number of forks moving in both directions whatever the size and the replication time of the CTR. The fact that the TR50 distribution of CTRs spans the entire S phase whatever their size suggests that all long CTRs and most small CTRs replicate during defined intervals of S phase by synchronous firing of multiple replication origins. The small-scale changes in fork polarity around individual origins are not seen due to the small replicon size and/or to the use of different potential origins in different cells, which effectively smooth replication timing gradients across multiple replicons.
Our demonstration that the apparent replication speed is equal to v/(R−L) (assuming that v is locally constant), implies that in TTRs replication forks move predominantly but perhaps not exclusively in one direction. To further investigate this we analyzed TTRs individually. We found that the temporal transitions were directly proportional to the length of the TTRs (Figure 5A). Even at the smallest scale analyzed (100 kb), only 24 out of these 7504 transitions were compatible with the progression of a single fork even at maximum rate (vmax = 2 kb/min) and together they only covered 0.13% of the genome. None of them was >250 kb, as expected from the maximum distance that a single fork can travel during S phase. Therefore, systematic unidirectional replication of large regions is not observed in HeLa cells. Replication forks instead appear to move in both a major and a minor direction in most TTRs. One potential explanation is that some TTRs support no internal initiation and are replicated from alternative origins located on either side of the TTR and used in unequal fractions of the cells (Figure S3C). As discussed for CTRs, this mechanism could only apply to TTRs<300 kb and would predict asynchronous replication of their edges, for which we did not find convincing evidence.
Alternatively, multiple origins could fire in a progressive manner along the TTRs (Figure S3D). The mean replication progression rate along TTRs was 3.63 kb/min, 5 times the mean progression rate of single forks (Figure 5A). This suggests that on average 2–3 adjacent replicons simultaneously operated along the gradient or, in other words, that on average adjacent origins spaced at ∼36 kb intervals were consecutively activated at ∼10 min intervals. Faster (slower) apparent speeds may result from shorter (larger) space and/or time intervals between adjacent initiations. This mechanism not only explains why replication progresses faster than single forks in TTRs but also why a higher proportion of forks move downstream than upstream the gradient, because when a new origin fires, the upstream moving fork will rapidly merge with the converging fork emanating from the upstream origin, whereas the downstream moving fork will progress for some distance before the next downstream origin fires. According to this mechanism, the faster distribution of speeds in late S phase is due to an increased synchrony of origin firings, consistent with the DNA combing results.
A visual inspection of the replication timing profile suggested that the slope of a large fraction of the TTRs tended to flatten with distance from their early edge. To asses this point, we selected TTRs>400 kb and measured the apparent replication speed at different positions along the slope. It was found that the apparent replication speed increased for about two thirds of the TTRs (Figure 5B). Furthermore, the distribution of apparent replication speeds along the TTRs was shifted to higher values at increasing distances from the early edge of the TTR (Figure 5C). These results suggest that forks move more and more in both directions along the TTRs as S phase progresses. These results are consistent with the DNA combing data showing that origins fire in an increasingly synchronous manner as S phase progresses.
The recent availability of high-resolution replication timing data in six other human cell lines (BG02, a human embryonic stem cell line; K562, a chronic myelogenous leukemia cell line; BJ, normal fibroblasts; GM06990, TL010, and H0287, lymphoblastoid cell lines) [13] prompted us to carry out a similar multiscale analysis of their apparent replication speeds. As shown in Figure 6, the distributions of replication speeds at the 100 kb scale were quite similar to HeLa cells with 3% (BJ) and <1% (other cells) of apparent speeds ≤2 kb/min, except for BG02 cells where a higher proportion of speeds ≤2 kb/min was observed (14.3%). Note that in the absence of published measurements of S phase length in these cell lines we have assumed a uniform S phase length of 8 h, typical of most mammalian cell lines. These distributions would be shifted toward proportionately faster (slower) speeds if S phase turned out to be shorter (longer). It is interesting to note that apparent speed distributions were much more similar among cell lines than single fork speeds and, by inference, origin activation patterns. This is consistent with a number of observations suggesting that replication timing is a more conserved feature among cell types than replication origin distribution [60]. The difference between BG02 and the other cell lines presumably reflects the previously described smaller replication domain size and higher density of timing transition regions in embryonic stem cells than in differentiated cells [8], [14].
As in HeLa cells, the observed ranges of apparent replication speeds in these cells cannot be explained by the range of single fork velocities measured by DNA combing in identical or comparable cells. In K562 cells, mean and max fork velocities are 0.37 kb/min and 1.0 kb/min [50] whereas >99.9% of the genome replicates at apparent speed >1.0 kb/min. To our knowledge, replication fork velocities have not been measured in the five other cell lines. However, mean and max fork velocities have been estimated to 1.73 and 2.9 kb/min in MRC5 fibroblasts (M. Debatisse, pers. comm.) and to 2.06 and 4.4 kb/min in JEFF lymphoblastoid cells [58]. Taking these values as reasonable estimates for BJ fibroblasts and for GM06990, TL010, and H0287 lymphoblastoid cells, respectively, it appears that 99.5–99.8% and 76–85% of the genome replicate faster than the mean and max fork velocity, respectively, in all those cell lines. Furthermore, mean fork velocities of 1.53–2.49 kb/min have been found in H9 and H14 embryonic stem cells [61]. Assuming that mean and max velocity in BG02 embryonic stem cells are 2.0 kb/min and 4.0 kb/min, respectively, we find that 85.7% and 61.9% of the genome replicate faster than these respective speeds. Thus, a higher proportion of the genome replicates at an apparent speed compatible with unidirectional progression of a single fork in BG02 cells.
To further investigate this we analyzed the TTRs of these six cell lines individually (Figure 7). The number of TTRs is about 2-fold higher in BG02 embryonic stem cells than in the differentiated cells (numbers in Figure 7 legend). Interestingly, a large fraction of the BG02 TTRs replicated at an apparent speed compatible with unidirectional progression of a single fork (Figure 7 A; average apparent speed 2.34 kb/min, mean fork velocity 2.0 kb/min). In all the other cell lines (Figure 7 B–F), however, the TTRs replicated faster than in BG02 (average apparent speed ranging from 3.24 kb/min to 4.21 kb/min, green lines), and faster than the mean fork velocity (compare dots with orange dashed lines). The discrepancy was most pronounced in K562 cells (Figure 7 B), where no TTR replicated slower than the fastest single forks (vmax = 1.0 kb/min, purple dashed line). In BJ fibroblasts (Figure 7 C) and in the three lymphoblastoïd cell lines (Figure 7 D–F), however, many TTRs replicated at an intermediate speed between the mean and max fork velocity (orange and purple dashed lines, respectively). The possibility that the slowest TTRs are specifically replicated by the fastest forks cannot be formally discounted but seems unlikely, as explained above. Therefore, most of the TTRs in these five cell lines replicate faster than by a single unidirectional fork. In other words, internal initiation in TTRs is more frequent in differentiated cells than in BG02 stem cells. Furthermore, in cancerous cells K562 and HeLa, replication forks progress more slowly and this likely triggers additional origin activation in TTRs.
Importantly, if domains of constant replication time were separated by timing transition regions of uniform and slow replication speed [8], [14], a biphasic distribution of apparent replication speeds should have been observed. This was not the case in any of the cell lines investigated. We found instead that the apparent replication speed, dx/dt, has a continuous and wide-range distribution significantly faster than the known range of fork velocities, v, in the vast majority of the genome. This implies that in all these cell lines, the statistics of origin activation creates throughout the genome a complete gradation in the predominance with which forks move in a preferred direction.
Our findings appear to contradict earlier views of genome-wide replication timing in human and mouse cells, which proposed a strict dichotomy between large (0.2–2.0 Mb) CTRs containing multiple synchronous origins and smaller (0.1–0.6 Mb) TTRs with slopes consistent with unidirectional replication fork progression [8], [9], [11], [14].
In the study by Desprat et al [9], the profiles were generated from the <2-fold copy number difference between S and G1 cells, which resulted in a low signal-to-noise ratio, and TTRs were defined as regions >250 kb in which the slope did not differ by more than 0.1 kb/min over their entire lengths. Such TTRs had slopes consistent with unidirectional fork progression (0.8–3.5 kb/min) but they only encompassed 5–8% of the genome. In three other studies [8], [11], [14], the profiles were generated from the abundance ratio of newly replicated DNA in different fractions of S phase and were segmented into CTRs and TTRs using a clustering algorithm. In all three cases, the resulting TTRs again only encompassed a small fraction of the genome (<10%). Hiratani et al [8] and Ryba et al [14], who used only two fractions of S phase, found slopes consistent with unidirectional fork progression (0.8–3.5 kb/min), but Farkash-Amar et al [11], who used up to seven fractions of S phase, found faster slopes (1.5–6.5 kb/min). As can be seen in Figure S3 in Hiratani et al [8], having only two S phase fractions creates an essentially biphasic distribution of replication times, an artifact that is much attenuated by the use of four to six S phase fractions (Figure 4E).
The profiles we analysed in this work were generated from four [12] or six [13] fractions of S phase, allowing us to discern replication timing differences within regions that were merged as a single replication timing domain in previous studies. Furthermore, we determined the full distributions of apparent speeds before any segmentation of the genome. These distributions were continuous, not biphasic, which implies that any segmentation in CTRs and TTRs entails a degree of arbitrariness. When we delineated CTRs and TTRs as contiguous regions which replicate faster (resp. slower) than 10 kb/min at a 100 kb scale, the genome was partitioned in two nearly equal halves. However, to obtain a set of TTRs that encompass <10% of the genome, we would need to set the threshold at ∼3 kb/min. Interestingly, the size range (0.1–0.5 Mb) and mean replication speed (2.3 kb/min) of such TTRs would be similar to those reported in the other studies, yet mostly incompatible with unidirectional fork progression given the fork speed measured by DNA combing in HeLa cells (Figure 5A, and data not shown). In none of the previous studies was the speed of replication forks directly measured on single DNA molecules in the same cells. We therefore believe that the rigid dichotomy reported in these studies overlooked the existence of a broad range of timing transition slopes, due to insufficient temporal resolution and/or to the use of a segmentation algorithm, and needs to be replaced with a more nuanced picture of DNA replication kinetics.
Although we do not exclude passive (but bidirectional) replication of TTRs<300 kb in HeLa cells, our data show that the mean replication progression rate along most of the genome is remarkably high, meaning that most TTRs are preferentially replicated by the progressive firing of multiple origins in most cells of a population. This is also the case for K562 cells. Nevertheless, we observed a higher proportion of apparent replication speeds consistent with unidirectional progression of a single fork in BG02 stem cells, and, to a smaller extent, in fibroblasts and lymphoblastoid cell lines in which replicons are longer and replication forks move faster than in HeLa cells.
In the study by Desprat et al [9], the notion that TTRs are originless regions that replicate by unidirectional fork progression was strongly supported by a single molecule analysis of the human IGH locus. This experiment unambiguously demonstrated that most forks progress unidirectionally in this transition region in human ES cells, in agreement with ample evidence for unidirectional replication of the homologous locus in mouse ES cells and T lymphocytes [41], [42]. This behavior is cell-type dependent, however, since abundant initiation events were detected in the same region during early and late stages of mouse B cell development [42].
We found that in HeLa cells the IGH locus is included in a 440 kb TTR whose apparent replication speed is 3.77 kb/min, inconsistent with unidirectional replication and significantly faster than reported by Desprat et al [9] in other cells (Figure 8A). We used DNA combing to determine the replication mode of this region (Figure 8 and Figure S4). We observed 43 initiation events on 25 DNA fibers evenly spread over a >700 kb region including the three restriction fragments studied by Desprat et al [9]. Only two out of these 26 fibers were found to contain a single fork. We also found that replication fork velocities (1.48±0.21 kb/min, N = 38) and inter-origin distances (46.0±5.1 kb, N = 20) in this region were approximately similar to that of the bulk genome. These results unambiguously demonstrate that this TTR replicates by progressive activation of multiple replication origins in HeLa cells and confirm the validity of our multiscale analysis of apparent replication speeds in predicting regions that cannot replicate by unidirectional progression of a single fork. Together with the results of Desprat et al [9], they also confirm that the replication mode of the human IGH locus can change according to cell type, as previously reported for the mouse IgH locus [42]. Since HeLa cells are derived from an adenocarcinoma, they show that replication origins in this region can be activated in non-B cells, although it is not clear if this results from a normal developmental program or from the tumoral nature of HeLa cells.
To further check our predictions of bidirectionally replicating CTRs and TTRs, we took advantage of the recent work of Letessier et al [58], who used DNA combing in fibroblasts and lymphoblastoid cells to reveal cell-type specific replication initiation programs at the FRA3B chromosome fragile site.
Analysis of the replication timing data of Hansen et al [13] shows that in BJ fibroblasts, the FRA3B region is embedded into a late-replicating 0.9 Mb CTR that is predicted to replicate by synchronous initiations (Figure 9A). The DNA combing results of Letessier et al. [58] entirely confirm this prediction, showing initiation and termination events evenly distributed all along this locus in MRC5 fibroblasts (Figure 9B).
In GM06990 lymphoblastoid cells, the FRA3B region lies within a V-shaped replication timing trough formed by two converging TTRs, each about 1 Mb in length (Figure 9C). Both TTRs replicate at an apparent speed of 6–8 kb/min, inconsistent with the mean (1.87 kb/min) and max (3.2 kb/min) velocity of single forks measured within this locus by DNA combing in JEFF lymphoblastoid cells [58]. Each of these two TTRs is therefore predicted to contain forks moving in both directions. A single fork moving at 2 kb/min could replicate up to 1 Mb of DNA within an 8 h S phase. Therefore, in principle, each TTR could be replicated without internal initiation if it is traversed by a single fork that is initiated two-thirds of the time on its early edge and one-third of the time on its late edge, since the resulting apparent speed would be v/|R−L| = 2/0.33 = 6 kb/min (Figure S3C). However, this scenario would predict that the edges of these TTRs would replicate either very early or very late in S phase, which is not supported by the timing data of Hansen et al [13] (see Figure 3 in Letessier et al [58]). The alternative hypothesis is that these TTRs replicate by internal initiations (Figure S3D). The data of Letessier et al [58] in JEFF cells indeed show initiations over the early and middle parts of each TTR, although initiations are excluded from a 700 kb region that corresponds to the late edges of both TTRs. Forks nevertheless are found to move in both directions in this 700 kb originless region [58], consistent with our predictions (Figure 9D). Another interpretation of these data would be that the edge of the early CTR is different in individual cells but the TTR is unidirectional in all cells, thus termination occurs at different points in the TTR. However, in order to quantitatively explain the discrepancy between the TTR slope and fork velocity, the position of this edge should differ by up to 1–2 Mb in different cells, which is again not supported by the timing data [13] [58]. Therefore, these results again confirm the validity of our analysis of apparent replication speeds in predicting regions that cannot replicate by unidirectional progression of a single fork.
Mesner et al [62] have recently provided a reliable map of replication origins in 1% of the human genome in HeLa and GM06990 cells, using a novel replication-bubble trapping procedure to prepare nearly pure origin libraries that were hybridized to encyclopedia of DNA elements [ENCODE] microarrays [63]. We compared the coverage of CTRs and TTRs by replication bubbles within these regions in HeLa cells (Figure 10A, and Figure S5). Most CTRs and TTRs contained replication bubbles, consistent with our proposal that most of the genome replicates by internal initiations (see e.g. region ENm001, Figure 10B), and we found no major difference in replication bubble coverage in TTRs (22%) vs. CTRs (29%). We noted however a higher bubble coverage in early replicating regions (Figure 10C). This was surprising because we found by DNA combing that interorigin distances did not change during S phase. Similar results were obtained with bubbles mapped in GM06990 cells (data not shown). A potential explanation for this discrepancy is that early bubbles are more efficiently trapped, perhaps because they are more efficient and less delocalised than late ones. Early bubbles may also have a longer dwell time than late ones because they are less synchronous and slower to merge with neighboring bubbles. This interpretation would imply that an even larger fraction of the genome than found by Mesner et al [62] can support a significant level of delocalised, replication initiation of low efficiency.
We also compared the coverage of CTRs and TTRs by short RNA-primed, nascent DNA strands purified by λ exonuclease digestion (λ-SNS) by Cadoret et al [64]. Although there is only modest concordance between bubble and λ-SNS maps [62], we again found no major difference in λ-SNS coverage (Figure 10D and Figure S5) in TTRs (1.05%) vs. CTRs (1.71%) and a higher λ-SNS coverage in early replicating regions (Figure 10E). It is expected that λ-SNS peaks are less efficiently detected if initiation is more random in late replicating regions.
A general correlation between replication timing, chromatin openness and transcriptional activity has been reported [12], [65]–[67]. To examine this in further detail, we analyzed the distribution along TTRs of an experimental marker (DNase I hypersensitive sites determined in HeLa S3 cells) and a DNA sequence marker (CpG islands) of open and transcriptionally active chromatin available genome-wide. We observed that the average coverages are maximum at TTRs early border and steadily decrease when going towards the late border (Figure 11A, B). These gradients of open chromatin marker distribution are unchanged when considering small, intermediate and large TTR size classes, suggesting there exists a characteristic scale for the change of chromatin state along TTRs. These results raise the possibility that there is a direct link between the gradients of origin firing time and a gradient in chromatin openness along the TTRs.
The mechanisms that regulate the timing of replication are unknown. A simple model to account for our data is that origins have different relative firing probabilities and fire stochastically, and that the firing probability of all origins increases during S phase. Thus, efficient origins are likely to fire during early S phase and weak origins are unlikely to fire early but become more likely to fire during late S phase [19]. The firing probability of origins may be specified by chromatin structure, since there is a general correlation between replication timing and chromatin openness [12], [65]–[67] (Figure 12A). Consistent with this model, we show here that markers of open chromatin are correlated with early replication throughout TTRs (Figure 11), and we have previously reported that origin firing probability increases during S phase in a wide range of eukaryotes including human [21], [68]. Furthermore, both the combing data and the distributions of apparent replication velocities at different stages of S phase provide evidences for increasing origin firing during S phase. One observation, however, argues against a purely uniform and uncorrelated stochastic model: origin firings are temporally and/or spatially correlated. It is possible that neighbor origins fire independently of each other but are nevertheless temporally correlated because their timing is set by some underlying chromatin features that change over a characteristic distance longer than individual replicons.
An attractive alternative mechanism to explain the progressive activation of neighboring origins along the TTRs is that active forks stimulate nearby initiation in unreplicated DNA. As discussed elsewhere [21], [69], [70], forks may stimulate initiation due to changes in DNA supercoiling in front of the fork or to association of chromatin remodellers or origin triggering factors with replication fork proteins. Early studies of replication foci labelled by two consecutive pulses showed that the intranuclear distance between consecutively replicated domains increased linearly with the time interval between the labels [36]. Studies on the dynamics of PCNA assembly at replication foci indicated that once replication is completed at a given site, a new replication focus assembles de novo at a neighboring site, consistent with a domino effect in activation of neighboring origins [38], [39]. A more recent study of S phase progression in HeLa cells suggested that replication foci that lie side-by-side in the nuclei are replicated in consecutive intervals of S phase because of their genetic continuity along the chromosomal fiber and that a “next-in-line” principle defines the efficiency with which origins are activated once S phase has begun [40]. In this work, we have quantitatively analyzed the speed of the replication wave progression and shown that it is consistent with a cascade of origin activation along TTRs as predicted by a domino model for origin activation. Thus, replication would first initiate in efficient zones of variable size specified by an open chromatin structure [67], followed by progressive activation of flanking origins in less open chromatin due to the approach of an incoming fork (Figure 12B). This model explains why adjacent origins tend to fire synchronously, why replication progresses faster than a single fork and why origins embedded in closed chromatin do not fire in early S phase but fire efficiently when the replication wave reaches them. With an increasing rate of origin firing during S phase [21], [68], this domino model can further explain why the apparent speed of replication increases along replication timing gradients, and predicts a progressive change in replication fork polarity along these gradients.
Works from several groups suggest that activation of one origin within a potential initiation zone suppresses rather than activates the activation of immediately surrounding origins [71]–[73]. However, the range of this negative origin interference is limited to distances smaller than the typical interorigin distance and is not incompatible with positive origin interference acting over larger distances [71], [74]. Data on origin spacing and synchrony in Xenopus egg extracts are indeed consistent with a mechanism whereby loop formation between a potential origin and an approaching fork suppresses initiation at very close spacing and enhances initiation at a larger, characteristic distance [29], [71], [74], [75].
In favor of a role of fork progression in controlling sequential origin activation, a recent study in yeast has shown that mutants deficient in chromatin remodeling activities located at replication forks specifically delay the replication of late replicating domains [76]. On the other hand, a study with mammalian cells has shown that exposure of aphidicolin-arrested cells to checkpoint inhibitors results in initiation of replication at successively later-replicating domains in the absence of detectable elongation of replication forks [77]. This suggests that fork elongation is not strictly required for at least the global aspect of temporal origin activation, but does not prove that it has no role in this process. Furthermore, it is possible that only the earliest origins are activated in successive large-scale replication domains, and that secondary origins within a domain require activation by replication forks.
In this work, we have performed a quantitative analysis of human genome replication in cells sorted into four or six stages of S phase, using DNA combing, mathematical analysis of replication timing profiles generated by massive sequencing of newly replicated DNA, and bioinformatic analysis of replication origin maps and chromatin structure data. The results show that i) replication origins fire in a correlated manner and at an increasing rate during S phase, ii) the apparent speed of replication progression throughout the genome depends on both the velocity of single forks and the proportion of rightward and leftward moving forks in the cell population, and ultimately reflects the pattern of origin firings along replication timing gradients rather than the unidirectional progression of a single fork. The correlation between adjacent origin firings may be due to their common chromatin environment or to a stimulation of origin firing by approaching forks. Further analyses and mathematical modelling of replication timing profiles are underway to explore these issues.
Asynchronously growing HeLa cells were labeled for 20 min with 25 µM IdU, washed with 1× PBS, and labeled for another 20 min with 25 µM CIdU. At the end of the labeling period, cells were harvested by trypsinisation, centrifuged at 500 g for 10 min at 4°C, washed in ice-cold 1× PBS, centrifuged again and fixed in 80% ethanol in 1× PBS. The fixed cells were centrifuged at 500 g for 5 min and resuspended in 1× PBS, 0.2 mg/mL RNaseA, 67 µg/mL propidium iodide at a final concentration of 2.106 cells/mL. Cells were sorted in four replication temporal compartments S1, S2, S3, and S4 based on their DNA content. DNA was extracted after encapsulation of cells in low-melting point agarose blocks at 60.000 genome equivalents per block (e.g. 60.000 cells for S1 and 30.000 for S4) and combed on silanised coverslips as described [78]. To detect the DNA molecules and the IdU and CldU labels, combed DNA was denatured in 50% formamide, 2× SSC for 10 min at 80°C. Coverslips were blocked in a humid chamber for 30 min at 37°C in antibody dilution buffer (1.5% blocking reagent (Roche), 0.05% Tween 20 in 1× PBS). The following sequential incubations were performed: (1) CldU detection: 1/20 rat anti-BrdU (Abcys) 1 hour, 1/25 chicken anti-rat Alexa Fluor 488 20 min, 1/25 goat anti-chicken Alexa Fluor 488 20 min. (2) IdU detection: 1/5 mouse anti-BrdU (Becton Dickinson) 1 hour, 1/200 rabbit anti-mouse Alexa Fluor 350 20 min, 1/25 goat anti-rabbit Alexa Fluor 350 20 min. (3) Total DNA detection : 1/25 mouse anti-human DNA (Millipore) 1 h, 1/25 goat anti-mouse Alexa Fluor 594 1 h. Coverslips were mounted in phenylenediamine and stored at −20°C before analysis. Incubations were at 37°C (except for the first step of incubations 1 and 2, at room temperature) in a humid chamber and washes between successive antibodies were three times in 1× PBS for a total of 15 min (anti-BrdU and anti-DNA antibodies) or 9 min (secondary antibodies). Coverslips were scanned using an Olympus IX81 or a Nikon Ti inverted microscope with a 100× objective, both connected to a CoolSNAP HQ CCD camera (Photometrics) run by MetaMorph version 6.3r7 (Molecular Devices). Fluorescent signals were analyzed with ImageJ software (Rasband, W.S., ImageJ, U. S. National Institutes of Health, Bethesda, Maryland, USA, http://rsb.info.nih.gov/ij/, 1997–2009.) and Adobe Photoshop 9.0.2 software. Data were inserted in an Excel® (Microsoft®) spread sheet and analyzed using R (http://www.r-project.org).
Several arguments suggest that IdU/CldU labeling had minimal effect on the rate of replication. First, it has been shown previously that the range of BrdU concentrations used for DNA combing does not affect the growth of yeast cells [79]. Second, the concentrations of IdU and CldU we used (25 µM) are among the lowest employed in numerous comparable studies (25–100 µM). Third, the rate of fork progression calculated from the IdU or IdU+CldU tracks was the same (data not shown), suggesting that a doubling of the total analog concentration did not affect fork progression.
Asynchronously growing HeLa cells were labeled with IdU and CldU and sorted and DNA was combed as described above except that cells were sorted in a single S phase compartment. Two sets of 3 fosmids (G248P83284G6, G248P8783H11, G248P81611C11) and (G248P86652F6, G248P87335H11, G248P81864F6) were biotinylated by random priming (Bioprime labeling system, Invitrogen) and were independently hybridized to the combed DNA as described [72]. Fosmids were selected on http://genome.ucsc.edu and distributed by http://bacpac.chori.org. Antibody incubation, washes and slide mounting for IdU/CldU and DNA detection were performed as described above with the following changes: DNA detection was coupled to FISH detection by sequential incubation (for 20 min each) with: 1/25 mouse anti-human DNA and 1/25 Alexa Fluor 594 conjugated Streptavidin, 1/50 biotinylated anti-streptavidin and 1/25 mouse anti-human DNA, 1/25 Alexa Fluor 594 conjugated Streptavidin and 1/25 goat anti-mouse Alexa Fluor 647, 1/50 biotinylated anti-streptavidin and 1/25 goat anti-mouse Alexa Fluor 647, 1/25 Alexa Fluor 594 conjugated Streptavidin and 1/25 chicken anti-goat Alexa Fluor 647.
We studied S-phase progression by flow cytometry analysis based on DNA content and IdU/CldU incorporation as described previously [80] with minor modifications. Here asynchronous HeLa cells were pulse-labeled with 25 µM IdU for 20 min and 25 µM CldU for another 20 min. DNA was stained with 1/5 mouse anti-BrdU antibody (Becton Dickinson) then with 1/25 rabbit anti-mouse Alexa Fluor 488 (invitrogen). Samples were analyzed on a CyAn ADP LX (Beckman Coulter).
To correct DNA content observed in DNA combing, we used flow cytometry analysis to estimate the percentage of IdU/CldU negative cells (i.e. fluorescence≤15) in S1–S4 fractions (Figure 2C). We measured at 31.5%, 9.05%, 4.65% and 15.1% the non replicative cells for S1–S4 fractions. Amount of DNA from these fractions have been subtracted from the total DNA length measured by combing. A second correction has been applied in order to remove previously synthesized DNA for each fraction. To this end, the mean DNA content for each fraction was estimated using FACS profiles (1, 1.15, 1.34, 1.57, 1.82 and 2 respectively for G1, S1 to S4 and G2) and then used to divide the previously corrected total DNA length.
The time required to duplicate the entire genome was estimated as the sum of times spent in S1, S2, S3 and S4 phases and corrected by the proportion of the genome that is replicated during these phases. S1–S4 lengths were individually calculated using the following equation: TSi = QSi/VGSi where QSi is the quantity of DNA synthesised, VGSi is the global progression of DNA replication. QSi was calculated as: QSi = 3.109×PSi with PSi determined by FACS profiles and corresponding to the proportion of the genome that is replicated in each S1–S4 compartment (respectively 15, 19, 22 and 20%). VGSi was calculated using the following equation VGSi = NFSi×VSi where NFSi is the quantity of forks in S1–S4 phases (i.e. forks density×genome length) and VSi is the replication forks velocity. The time required to duplicate the entire genome was computed as T = ΣTsi/ΣQsi.
In one cell cycle, the timing profile around an active origin of replication has a typical inverted V shape corresponding to a local minimum (timing increases when going downward). Downstream of the origin of replication, loci are replicated by forks coming from their left, therefore R = 1 and L = 0, and in that region the timing profile has a positive derivative dt/dx = 1/v where v is the replication fork velocity. Respectively, upstream of the origin of replication, L = 1 and R = 0 and the timing profile has a negative derivative dt/dx = −1/v. Therefore, the derivative of the timing profile of a single cell is given by dt/dx = (R−L)/v. It can be shown that this result still holds when considering the average over a cell population: the average fork polarity is provided by the derivative of the average timing profile given a constant fork velocity. Given the finite resolution of the experimental average timing profile, we defined the apparent speed at scale X kb as the inverse of the slope of the timing profile computed at that scale. This apparent velocity is equal to the fork velocity divided by the average fork polarity over that scale.
For HeLa cells, we previously generated a profile of S50, the fraction of S phase at which 50% of the DNA is replicated in a defined genome region, using massively parallel sequencing of BrdU-labeled nascent DNA from sorted cells in S1, S2, S3, S4 [12]. To verify the DNA content of sorted cells for DNA combing and replication timing experiments, 105 BrdU-labelled, sorted cells in S1, S2, S3, S4 and 105 sorted cells with a DNA content ranging from G1 to G2 were re-stained with propidium iodide and their DNA content examined by FACS (Figure S1). This control showed that the sorted cells had the expected DNA content.
The enrichment of sequence read densities relatively to background was computed along the genome for each of the four compartments of the S phase and S50 values were computed by linear interpolation of enrichment values [12]. TR50, the time at which a defined genome region had replicated in 50% of the cells was then deduced from S50 values as follows.
The FACS DNA fluorescence histogram was analysed using a modified version of the method developed by Bertuzzi et al. [81]. We assumed that in average all cells whose DNA content at time t is x, synthesize their DNA with the same rate φ(x) that we approximated with the sum of six Gaussian functions. The fraction of cells in S phase (equation (16) in [81]) with a DNA content x is given by:where θ1 is the fraction of cells in G1 phase measured by integration of the peak of the FACS fluorescence histogram corresponding to the cells in G1; the term represents the time spent by a cell in S phase (equation (7) in [81]). Using the fundamental equation of cytofluorimetry [82] and the expression for ñ(x) we fitted the fluorescence histogram utilising a simplex algorithm and extracted the profile of φ(x). Using the expression for time as a function of DNA content, , we obtained the overall fraction of the genome that has replicated at time t in S phase. This was used to convert S50 into TR50.
For the other cell lines, we determined a profile of S50 using Repli-Seq tags for 6 FACS fractions that were obtained from the authors [13]. For a given cell line and for each S-phase fraction, we computed the tag densities in 100 Kb windows, and following the authors [13] the tag densities were normalized to the same genome-wide sequence tag counts for each fraction. We performed a second normalization so that at each genomic position, the sum over S-phase fractions be one. To filter out the noise which could critically bias mean timing profile estimate, we proceeded as follow. We noticed that the genome-wide distribution of the normalized tag density presents a mode at 0.01<m<0.08 (mainly noise) and a long tail up to 1 (mainly corresponding to the replication signal). For each S-phase fraction we set to 0 the normalized tag density <4 m, and re-normalized at each genomic position by the sum over S-phase fractions. The mean replication timing profile computed on these denoised tag densities superimposed on the original one, but was much less noisy. We directly converted S50 into TR50 assuming an S phase length of 8 h and a linear mapping between DNA content and S phase progression.
The apparent replication speed of a locus intuitively corresponds to the inverse of the slope of the replication timing profile. In fact the timing profile is noisy so that its derivative is strictly speaking not defined. We used the continuous wavelet transform (WT), a powerful framework for the robust estimation of signal variations over any length scales [83], [84], to obtain a well defined and numerically stable measurement of the local slope of the timing profile at any scale of observation. This allowed us to construct the space-scale map of apparent replication speeds (Figure 3B and Figure S2B). Using this map, CTRs and TTRs were delineated as the contiguous regions where the speed is above (resp. below) a constant threshold (10 kb/min) at a given scale of observation (100, 200 and 500 kb) (Figure 4).
Sequence and annotation data were retrieved from the Genome Browsers of the University of California Santa Cruz (UCSC) [85]. Analyses were performed using the human genome assembly of March 2006 (NCBI36 or hg18). We used CpG islands (CGIs) annotation provided in UCSC table “cpgIslandExt”. As previously done, we computed 1 kb-enlarged CGI coverage as an hypomethylation marker [67].
We used the DNaseI sensitivity measured genome-wide in HeLa S3 cell line using the Digital DNase I methodology [44], [45]. Data corresponding to Release 3 (Jan 2010) of the ENCODE UW DNaseI HS track, were downloaded from the UCSC FTP site: ftp://hgdownload.cse.ucsc.edu/goldenPath/hg18/encodeDCC/wgEncodeUwDnaseSeq/. We plotted the coverage by DNase Hypersentive Sites (DHSs) identified as signal peaks at a false discovery rate threshold of 0.5% within hypersensitive zones delineated using the HotSpot algorithm (“wgEncodeUwDnaseSeqPeaks” tables).
The coordinates of replication-bubble trapping fragments of HeLa and GM06990 cells within ENCODE regions were obtained from the authors [62].
|
10.1371/journal.pcbi.1003576 | Effects of Specular Highlights on Perceived Surface Convexity | Shading is known to produce vivid perceptions of depth. However, the influence of specular highlights on perceived shape is unclear: some studies have shown that highlights improve quantitative shape perception while others have shown no effect. Here we ask how specular highlights combine with Lambertian shading cues to determine perceived surface curvature, and to what degree this is based upon a coherent model of the scene geometry. Observers viewed ambiguous convex/concave shaded surfaces, with or without highlights. We show that the presence/absence of specular highlights has an effect on qualitative shape, their presence biasing perception toward convex interpretations of ambiguous shaded objects. We also find that the alignment of a highlight with the Lambertian shading modulates its effect on perceived shape; misaligned highlights are less likely to be perceived as specularities, and thus have less effect on shape perception. Increasing the depth of the surface or the slant of the illuminant also modulated the effect of the highlight, increasing the bias toward convexity. The effect of highlights on perceived shape can be understood probabilistically in terms of scene geometry: for deeper objects and/or highly slanted illuminants, highlights will occur on convex but not concave surfaces, due to occlusion of the illuminant. Given uncertainty about the exact object depth and illuminant direction, the presence of a highlight increases the probability that the surface is convex.
| A primary goal of the human visual system is to reconstruct the three-dimensional structure of the environment from two-dimensional retinal images. This process is under-determined: an infinite number of combinations of shape, material properties and illumination conditions could give rise to any single image. Rather than determining the true three-dimensional scene in a deductive manner, the visual system must make its ‘best guess’ based on the image, probabilistic models of image formation, and the stored probability of various scene configurations. For example, the visual system appears to assume that convex surfaces are more common than concave ones, biasing perception toward convex surfaces when the image is ambiguous. Here we identify a new probabilistic cue for surface shape: a shape with a visible specular highlight is more likely to be convex than one without. Highlights occur when light is reflected in a mirror-like way from glossy surfaces such as polished marble or metal. Due to the geometry of reflection, however, highlights are more likely to be occluded on concave objects. We show that the human visual system makes use of this constraint: shape perception is biased toward convex surfaces when highlights are apparent.
| Shading can produce striking impressions of 3D shape. However, recovering shape from shading is far from straightforward; luminance variations in the image are determined not only by the object's shape but also by its reflectance and the illumination conditions. To estimate shape from shading, the visual system biases judgements toward more common scenes, for example, light sources that are roughly overhead (e.g. [1], [2]) and surfaces with homogenous reflectance, at least in the absence of hue variation [3]. Here we explore an additional regularity that the visual system appears to exploit in estimating surface shape: that specular highlights suggest convex, rather than concave curvature. We test this proposal psychophysically and show why, given certain assumptions, this bias is rational: it reflects a higher likelihood of observing a specular reflection from a convex object.
It is well known that a specular highlight can change the perception of surface material, making a matte object look glossy (Figure 1a). However, the effect of specular highlights on shape perception has received little attention. Specular highlights do carry shape information, tending to ‘cling’ to regions of high curvature [4]–[6], and observers can use the structure of specular highlights alone (e.g. on perfectly mirrored surfaces) to estimate curvature magnitude [7]. Several studies have compared monocular shape perception across matte and specular surfaces to assess the role of specular highlights in quantitative shape estimation. Whilst some studies found that specular highlights increased perceived depth [8]–[10] or improved shape discrimination [11], others have found no effect of surface specularity on shape judgements [12]–[14]. Ho, Landy and Maloney [15] found that the glossiness and bumpiness of a surface are somewhat confusable, even under binocular viewing: increasing surface depth increases perceived glossiness and vice versa.
When a glossy object is rotated, specular highlights glide across the object's surface, rather than being fixed to it like texture. This motion provides information that observers exploit to judge both gloss [16], [17] and shape [11], [18], [19]. Similarly, under binocular viewing, the disparity of specular highlights holds information not only about the magnitude of surface curvature but also its sign; for simple convex objects, specular highlights are stereoscopically behind the surface, for concave they are generally in front. The visual system appears to use this information in judgments of glossiness and, to a limited extent, shape [11], [20]–[24].
Note that this cue to surface convexity depends upon the binocular disparity of reflections. For distant surfaces, this disparity signal will be weak, and thus no bias is predicted. Intriguingly, however, Blake & Bülthoff [20] noted informally that under monocular viewing, the addition of a specular highlight seemed to bias perception of their stimuli toward convexity, though this effect was not tested empirically. Unlike the binocular effect, such a bias does not have a straightforward geometric explanation. Yet it is important to determine whether this effect is real and quantifiable, since in the real world, disparity signals become very unreliable for distant surfaces, and other visual features (e.g., shading, texture) provide only weak cues to curvature sign. If specular highlights provide a cue to surface convexity, they may prevent observers from making large perceptual errors about distant surfaces in the environment.
Here we ask how specular highlights combine with Lambertian shading cues to determine perceived surface curvature, and to what degree this is based upon a coherent account of the scene geometry. To avoid covariation with other features found in natural scenes (stereoscopic disparity, motion, texture, etc.) we employ relatively simple stimuli (shaded ellipsoidal surfaces with constant albedo), and manipulate the location of highlights relative to the Lambertian shading gradient to vary the consistency of the two cues. In three experiments we ask:
The problem of judging surface shape for these stimuli is ill-posed: there are many possible scene configurations that could give rise to each observed image, and in particular both signs of surface curvature, convex and concave, are possible. Here we hypothesize that the human visual system attempts to determine the most probable curvature sign given the image data. In order to assess whether our psychophysical results are consistent with this principle, we construct a quantitative Bayesian model that attempts to explain the shading and highlights observed in the image in terms of the illumination field, object shape and surface material (glossy or matte). So as not to obscure the empirical results, we defer detailed presentation of the model to the Materials and Methods section, however we will discuss its qualitative properties and show the fit of the model to the psychophysical data alongside our empirical results.
In our first experiment, observers viewed a pair of shaded objects, with or without specular highlights (Figure 1a) and reported perceived sign of surface curvature (convex or concave) of one of the objects. The shading gradients on the two objects were always in opposition and were systematically varied over all angular directions, in 15 deg increments. There were four conditions:
MM: Neither object has a highlight.
SM: The target object has a highlight, the distractor object does not.
MS: The target object does not have a highlight, the distractor object does.
SS: Both objects have highlights.
Our first experiment shows that the appearance of a specular highlight biases observers toward a convex interpretation of the stimulus. For these stimuli, the geometry of reflection dictates that the highlight appears on the lighter side of the shape, aligned with the shading gradient. In Experiment 2 we ask how the effect on perceived convexity varies as a function of this alignment (Figure 4a).
Figure 4(b) shows data averaged across 10 observers. Each subplot shows perceived convexity as a function of shading orientation for a single specular highlight position (indicated by a yellow star). As in Experiment 1, objects with a highlight were judged to be convex more often than objects without. Furthermore, this effect seems to persist even when the highlight is rotated out of alignment with the shading gradient, although the magnitude of the effect is reduced.
To better understand this variation, we calculated the mutual information between the presence of a highlight and perceived shape using data from the SM and MS conditions. Figure 5a shows the results as a function of Lambertian shading orientation, averaged across highlight location. As in Experiment 1, a highlight is ineffectual when objects are bright at the top; these objects are perceived as convex (and lit from above) with or without a highlight. When a highlight appears near the top of the object, therefore, it is not possible to assess whether highlight-shading alignment (and thus highlight interpretation) modulates the effect of highlights on shape perception. However, we can examine the effects of highlight misalignment on shape by considering the mutual information between perceived and convexity and highlights appearing on the lower half of the object (Figure 5(b–d)). We see that the effect on shape is largest when the highlight is aligned, or nearly aligned, with the diffuse shading gradient.
The specular occlusion account is qualitatively consistent with the observed bias to convex surfaces induced by the appearance of a highlight, but without quantitative measurement of the prior over object shape and illuminant slant it cannot be verified quantitatively. Here we present an additional psychophysical experiment that provides an additional test of the model.
The specular occlusion hypothesis is rooted in uncertainty over the exact shape of the surface and the location of the illuminant. As a result, visual cues that shift the posterior distribution over these scene variables should alter the probability of highlight occlusion and therefore the induced convexity bias. In particular, the bias should get stronger when these cues suggest either (i) an increase in surface depth or (ii) an increase in illuminant slant (deviation from the view vector), since both variations increase the probability of specular occlusion for a concave surface.
Our third experiment directly tests this prediction of the model (we thank one of the anonymous reviewers for suggesting such an experiment). As in Experiments 1 and 2, observers viewed pairs of shaded stimuli, and reported the perceived shape (convex or concave) of one object. The shading and highlight cues to absolute depth are subtle and confounded with illuminant slant; by adding texture to the objects we provided an independent cue to depth that should allow observers to better dissociate these two scene variables (Figure 6a). The shading gradients of the two objects were always in opposition and either one or neither of the objects had a specular highlight. The two objects always had the same depth magnitude, however, this depth and the slant of the illuminant varied across trials.
To focus the experiment, we determined the shading gradient direction for each observer that produced balanced (50%) reports of ‘convex’ and ‘concave’ for the two oppositely shaded matte objects, and then examined the effect of the highlight on perceived convexity while varying object depth and illuminant slant.
Figure 6 shows example stimuli and the data from this experiment. The highlight effect is quantified by the proportion of ‘convex’ responses in the presence of a highlight (in contrast to 50% when absent). Figure 6b shows the effect as a function of illuminant slant, collapsed across stimulus depth. As the direction of illumination approaches the image plane (increasing slant), the effect of the highlight on perceived shape increases (F5 = 7.3; p<0.01). Figure 6c shows the effect as a function of stimulus depth, collapsed across illuminant slant. As object depth increases, the effect of the highlight on perceived shape again increases (F3 = 6.2; p<0.05). In summary, as predicted by the geometry of specular occlusion, increases in illuminant slant or object depth both increase the probability of convex report.
Interestingly, while increasing illuminant slant or object depth both increase the convexity bias, they have opposite effects on the position of the highlight (dashed lines in Figures 6b and c). In particular, while increasing the slant of the illuminant shifts the highlight toward the rim of the object, increasing the depth of the object shifts the highlight in the opposite direction, toward the centre of the object. Our results therefore indicate that the observer is not simply relying on the position of the highlight when judging curvature sign. Instead, our data suggest that the observer's perception is modulated by estimates of quantitative depth and illumination direction, becoming increasingly biased toward a convex interpretation as the probability of highlight occlusion increases. These results are thus a strong confirmation of the specular occlusion account of the convexity bias induced by the appearance of a highlight.
We have conducted three experiments to explore the effects of highlights on perceived convexity:
The results from all three experiments are consistent with a Bayesian model that takes into account potential light source occlusion. Does this mean that observers are constructing a complete and detailed 3D solution for the entire scene? Some have argued against this kind of ‘inverse optics’ model [14], suggesting that the underlying variables of shape, reflectance and illumination may not be estimated concurrently, so that probing the percept of each will not necessarily yield consistent results. Furthermore, while shape and material may be important for manipulating and recognizing objects, we might question whether observers require an explicit estimate of the illumination field.
On the other hand, there is evidence that observers make judgments of shape and/or reflectance consistent with a particular estimate of the illumination field without necessarily making this estimate explicit. Observers can manipulate the shading pattern of one object to appear consistent with a second object, such that the implicit illumination environments match [43], although like our observers, they relied on priors for overhead illumination and object convexity when image cues were ambiguous. Similarly, reflectance judgements for ambiguous images are consistent with a single overhead illuminant [25]. In contrast, observers are poor at making explicit judgements of illumination consistency across multiple objects [44].
In our experiments, observers are asked only to judge the convexity of objects, and not the glossiness of the surfaces or the number or direction of light sources. As a consequence, the predictions of the Bayesian model (Materials and Methods) are not based upon explicit joint estimation of these scene variables, but do depend critically on at least approximate marginalization over the unknown ‘nuisance’ variables (object depth, illumination) when judging convexity. This process of marginalizing over or ‘integrating out’ nuisance variables when judging other scene variables of interest is widely believed to explain a number of visual phenomena (e.g., [27], [45]), and the consistency of our Bayesian model with the psychophysical data suggests that it may also explain the effect of highlights on the perception of surface convexity.
The interplay between the light field, surface reflectance and surface shape is complex and many issues remain to be resolved. Our experiments reveal the effect of specular highlights on perceived convexity for ellipsoidal surfaces and point light sources. It remains to be seen whether this effect generalises to more complex surfaces and light fields (see Figure S1 for examples of ellipsoidal stimuli rendered with ray-tracing under a complex illumination field). In addition, further studies may resolve the existing inconsistencies in the literature regarding the effect of highlights on perceived curvature magnitude [8]–[14].
Overall, our results shed new light on how the human brain uses highlights to disambiguate 3D surface shape. Our Bayesian model suggests that this is more than a ‘bag of tricks’[46]. Rather, inference can be accounted for as a rational computation that selects the most probable shape interpretation, given the observed data and prior information about the relative probability of alternative scene configurations.
For all experiments, participants gave informed consent and the local ethics committee approved the study.
Stimuli consisted of two axis-aligned half-ellipsoids, compressed in depth by a factor of two relative to a hemisphere, illuminated by a single, distant light-source. The orientations of the smooth (Lambertian) shading gradients on the two objects were always in opposite directions.
When a single object (either with or without a highlight) is presented in isolation it is perceived as convex for all illumination tilts due to the widely documented prior for object convexity [47]–[50]. This convex bias is represented in our model by the prior over curvature sign : over observers. When two objects are presented with opposing shading gradients, the prior for a single illuminant counteracts the convexity prior, causing the observer to perceive the objects as having opposing curvature sign, on most trials. The two-object scene thus allows us to explore the effects of specular highlights on shape perception.
There were four stimulus configurations: (1) Highlight on neither object, (2) Highlight on the left object, (3) Highlight on the right object, (4) Highlight on both objects (Figure 1a). Stimuli were generated as grey objects under white light using the Phong lighting model implemented in OpenGL, without inter-reflections or cast shadows, under orthographic projection. Shiny objects were rendered with ambient (7% of maximum), diffuse (36% of maximum) and specular components (48% of maximum, with Phong exponent of 80). Matte objects had only diffuse and ambient components.
Under this Phong lighting model and orthographic projection, convex and concave objects generate identical images, thus rendering the estimation of the sign of surface curvature completely ill-posed, allowing us to isolate the role of highlights in the perception of surface convexity. In a real scene, however, subtly different patterns of interreflection could in theory serve to discriminate convex from concave surfaces. In practice however, these differences are relatively minor for our scenes, as confirmed by comparing ray-traced renderings, under a complex light field, with and without inter-reflections (compare Figures S1a and b).
We define a coordinate frame with origin at the centre of the display, X- and Y-axes in the horizontal and vertical directions in the plane of the screen, respectively, and Z-axis positive toward the observer. The slant of the single directional light (the angle between the lighting vector and the Z-axis) was held constant at 68°. The tilt of the lighting direction (the angle between the projection of the lighting vector and the Y-axis) varied across trials. The orientation of the shading gradient for each object was thus a function of its curvature sign and light source tilt. The room was unlit aside from the light emitted by the monitor. To eliminate binocular and motion-based depth cues, stimuli were viewed monocularly, with the observer's head fixed by a chin rest and forehead bar. At the viewing distance of 57 cm, each object subtended 5° with their centres displaced horizontally ±3.4° from the display centre. Scenes were rendered with orthographic projection, simulating an infinite viewing distance. Given the small angular subtense of our stimuli, switching to perspective projection has only a small effect on the shading gradient and position of the highlight in our images (see Figure S1c).
On each trial, the two shaded objects appeared for 1 second. Halfway through the presentation, a star appeared next to one of the objects, indicating that this ‘target’ should be judged. By a key-press, the subject reported the target curvature as either ‘convex’ or ‘concave’. The four conditions (Target Matte, Distractor Matte (MM); Target Matte, Distractor Shiny (MS); Target Shiny, Distractor Matte (SM); Target Shiny, Distractor Shiny (SS)) and the target's shading orientation were randomly interleaved. Ten observers (9 naïve and 1 author) each completed 1536 trials (24 target orientations x 4 conditions x 16 repetitions) in a single session lasting approximately 1 hour. One additional naïve observer was excluded from the analyses as the direction of the shading gradient had little effect on his/her shape judgements.
Only one of the two objects was rendered with a highlight, and the orientation of the diffuse shading component (16 equally spaced values) and the angular position of the highlight (10 equally spaced values) were varied independently, by rendering the diffuse and specular components of the image with independently positioned illuminants (Figure 4). As in Experiment 1, the two objects had opposite gradient directions and a star indicated which of the two objects should be judged (convex vs. concave). The 3840 trials (10 highlight positions x 16 shading orientations x 2 conditions (SM: only the target has a highlight, MS: only the distractor has a highlight) x 12 repetitions) were completed in 3 sessions of approximately 45 minutes. All other details were identical to Experiment 1. The 10 observers who completed Experiment 1 also participated in Experiment 2.
In our third experiment we studied the effect of object depth and illuminant slant on the convexity bias caused by a highlight. This is tricky to do in a controlled fashion using the ellipsoid objects of Experiments 1 and 2, as the variation in curvature across the shape induces changes to both the shape and size of the highlight as the slant of the illuminant is varied. To stabilize the appearance of the highlight, we replaced the ellipsoidal surfaces with sections of hemispheres that protruded from or recessed into the planar background surface. Since surface curvature is constant over the hemisphere, variations in illuminant slant induce much smaller variations in the shape and size of the highlight.
As in Experiments 1 and 2, the direction of the shading gradient on the two objects was always in opposition, such that the stimulus was consistent with one convex and one concave object, both illuminated by a single light source. The simulated depth of both objects always matched, but varied across trials (depth:radius ratio was 0.25, 0.5, 0.75 or 1), by changing the radius of the sphere from which the domes were constructed. Highlight position was yoked to the shading gradient: i.e. both were rendered with the same illuminant.
In order to determine the shading gradient that produces a roughly balanced perception of convex and concave shape for each object depth and illuminant slant, for each observer, we sampled a range of illumination tilts between 90° to 270° (7 equally spaced values). Illuminant slant varied across trials from 25° to 75° (6 equally spaced values). A green texture (see Figure 6a) was wrapped around both objects and the planar background to facilitate depth perception. We found that the sharp join between the hemisphere sections and the planar background caused the objects to appear detached from the background; to avoid this, we introduced a thin curved section to smooth this join. For specular objects, this generated an additional very thin specularity at the join (Figure 6a); this additional feature does not appear to be correlated with variations in observer reports of perceived convexity. Four observers completed 2016 trials (4 depths x 6 illuminant slants x 7 illuminant tilts x 2 specularity conditions (no highlights or highlight on the target only) x 6 repetitions in two sessions of approximately 30 minutes.
Both object depth and illuminant slant have a systematic effect on the perceived curvature sign of matte objects; shallow objects and small illuminant slants produce shading patterns that are more similar for the two objects, and perhaps for this reason the overall proportion of convex responses increases under these conditions (although this did not reach significance). To compare the effect of the highlight across these conditions without the confound of varying baseline convexity, for each condition and each subject we found the shading orientation at which the matte stimulus was perceived as convex on 50% of trials. This was found by fitting a psychometric function to the proportion of convex responses as a function of shading orientation, and obtaining the 50% threshold. We then measured the effect of the highlight by the proportion of convexity judgments relative to this consistent 50% baseline (Figures 6b and c).
Our psychophysical experiments have shown that the judgement of surface convexity is dependent upon the appearance of surface highlights and their locations relative to the shading gradient induced by surface curvature. In our view, the most important question is why a highlight has this effect. Here we put forward a specific theory: due to potential occlusion of the light source for a concave surface, highlights occur more frequently on convex surfaces in natural scenes. As a consequence, the convexity bias induced by highlights will increase the ability of the observer to correctly judge the sign of surface curvature.
While this theory is qualitatively consistent with the psychophysical data, it remains to be seen whether it is quantitatively consistent with the data. To assess this, we have constructed a Bayesian model for the discrimination of convex vs concave surface curvature given the shading gradients and highlights appearing on the two objects comprising our stimuli. Specifically, the observable variables are (Figure 7):
The model incorporates the minimal set of hidden scene variables sufficient to explain the observed shading and highlight cues. These include:
We believe this to be the minimal set of hidden variables that makes sense: removal of any one of these variables would mean that the model would not capture a basic feature of the phenomenology or relationship between observable features and observer reports (see Model complexity).
Capturing the relationship between perceived surface curvature sign and illumination requires modelling probability distributions over the angular direction (tilt) of the illuminant and corresponding observable variables. Observers have a well-documented prior for overhead illumination [1], [2], [25], [26], [34], [48], [51]–[53] that has previously been successfully modelled by a von Mises distribution [51] although the mean of this distribution varies considerably across observers [25]. We employ the von Mises distribution to model observers' prior distribution over illuminant tilt, with the general formwhere is the tilt angle, and are the mean and concentration (inverse variance), and is the modified Bessel function of order 0, required for normalization. This distribution is used to model:
For each observer, the values of the 9 model parameters (summarized in Table 2) were found (MATLAB fminsearch) that maximize the joint likelihood of the observed data for both Experiments 1 and 2. Multiple iterations of the parameter search were performed, with the initial values on each iteration determined by uniform sampling within a plausible parameter range. All equations for the model can be found in Text S1.
Our model was constructed to include only scene variables relevant to the observers' judgement of convexity for the two-object stimuli used in our experiments. Nevertheless, the model does have nine free parameters, raising the question of whether we are overfitting the data.
To address this question, we considered three models of reduced complexity and compared their ability to account for the psychophysical data (Table 1).
We find that the full model provides the best account of the data, for every observer, as indexed by the Bayesian Information Criterion (see Figure 3b). This result suggests that to account for the perception of surface convexity one must allow for a) a prior bias for convexity, b) the possibility of complex illumination fields, c) the biasing effects of highlights and d) the possibility of attributing these highlights either to specular reflection or to a local illumination effect, depending upon the consistency of the highlight with the shading gradient.
To understand the scene parameters leading to specular highlight occlusion, we can, without loss of generality, consider the viewing geometry of our scene in cross-section, in the plane defined by the viewing and illuminant vectors, with the illuminant on the right (see Figure 2a). The resulting cross-section of the surface describes a semi-ellipse. We define the depth expansion factor d to be the ratio of the length of the semi-axis in the viewing direction z to the length of the semi-axis in the horizontal direction x. Without loss of generality, we assume that the length of the semi-axis of the ellipse in the horizontal direction is 1, so that the length of the other semi-axis (in the viewing direction z) is equal to the depth expansion factor d. Centering a 2D coordinate system directly above the concave surface, at the level of the rim, the surface cross-section can be described by the equation(0.1)
Taking a first derivative yields , so that the tangent vector must be in the direction and the normal vector must be in the direction .
The specular highlight will be located at the point on the semi-ellipse where the normal bisects the angle formed by the view vector and the illuminant vector. Thus we have(0.2)
Together, Equations (0.1) and (0.2) determine the location of the highlight: solving (0.2) for and substituting in (0.1) yields(0.3)
For our stimuli, the depth expansion factor and illuminant direction were fixed at and , yielding a highlight location of .
Of course the observer does not know the exact surface depth or illuminant direction, and for a highlight appearing at this particular location there is in fact a one-dimensional family of solutions to Equation (0.3) given by(0.4)and described by the blue curve in Figure 2c.
However, not all of these solutions are physically possible: for larger illumination angles (and larger surface depths), the view of the illuminant from the required highlight location will be occluded by the rim of the surface. To quantify this constraint, we note that the angle of the vector pointing to the rim from the highlight location, relative to the view vector (Figure 2a), can be written as(0.5)
Substituting for from (0.2) yields
(0.6)and substituting for from Equation (0.4) yields(0.7)
Equation (0.7) describes the angle of the rim of the surface as seen from the potential highlight location, as a function of the estimated depth expansion factor . This function is shown by the red curve in Figure 2c. Note that for a subset of solutions with highly oblique illumination and large surface depth, the red curve lies below the blue curve. These solutions are physically infeasible because the illuminant is occluded by the rim of the surface.
For a Bayesian observer who is uncertain about the surface depth and elevation of the illuminant, a consequence is that observation of a highlight will decrease the probability of concave surface curvature relative to the probability of convex surface curvature, for which all solutions are feasible.
|
10.1371/journal.pntd.0005770 | A strategy for scaling up access to comprehensive care in adults with Chagas disease in endemic countries: The Bolivian Chagas Platform | Bolivia has the highest prevalence of Chagas disease (CD) in the world (6.1%), with more than 607,186 people with Trypanosoma cruzi infection, most of them adults. In Bolivia CD has been declared a national priority. In 2009, the Chagas National Program (ChNP) had neither a protocol nor a clear directive for diagnosis and treatment of adults. Although programs had been implemented for congenital transmission and for acute cases, adults remained uncovered. Moreover, health professionals were not aware of treatment recommendations aimed at this population, and research on CD was limited; it was difficult to increase awareness of the disease, understand the challenges it presented, and adapt strategies to cope with it. Simultaneously, migratory flows that led Bolivian patients with CD to Spain and other European countries forced medical staff to look for solutions to an emerging problem.
In this context, thanks to a Spanish international cooperation collaboration, the Bolivian platform for the comprehensive care of adults with CD was created in 2009. Based on the establishment of a vertical care system under the umbrella of ChNP general guidelines, six centres specialized in CD management were established in different epidemiological contexts. A common database, standardized clinical forms, a and a protocolized attention to adults patients, together with training activities for health professionals were essential for the model success. With the collaboration and knowledge transfer activities between endemic and non-endemic countries, the platform aims to provide care, train health professionals, and create the basis for a future expansion to the National Health System of a proven model of care for adults with CD.
From 2010 to 2015, a total of 26,227 patients were attended by the Platform, 69% (18,316) were diagnosed with T. cruzi, 8,567 initiated anti-parasitic treatment, more than 1,616 health professionals were trained, and more than ten research projects developed. The project helped to increase the number of adults with CD diagnosed and treated, produce evidence-based clinical practice guidelines, and bring about changes in policy that will increase access to comprehensive care among adults with CD. The ChNP is now studying the Platform’s health care model to adapt and implement it nationwide.
This strategy provides a solution to unmet demands in the care of patients with CD, improving access to diagnosis and treatment. Further scaling up of diagnosis and treatment will be based on the expansion of the model of care to the NHS structures. Its sustainability will be ensured as it will build on existing local resources in Bolivia. Still human trained resources are scarce and the high staff turnover in Bolivia is a limitation of the model. Nevertheless, in a preliminary two-years-experience of scaling up this model, this limitations have been locally solved together with the health local authorities.
| Bolivia has the highest prevalence of Chagas disease (CD) in the world (6.1%), with more than 607,186 people with Trypanosoma cruzi infection. In Bolivia, the management of CD has been declared a national priority. In 2009, the Chagas National Program (ChNP) had neither a protocol nor a clear directive for diagnosis and treatment of adults. The Chagas Platform has been built as a model for comprehensive care of adults with CD. From 2010 to 2015, a total of 26,227 patients were attended by the Platform, 69% (18,316) were diagnosed with T. cruzi, 8,567 initiated anti-parasitic treatment, more than 1,616 health professionals were trained. More than ten research projects were developed. The project has also produced evidence-based clinical practice guidelines, and brings about changes in policy that will increase access to comprehensive care among adults with CD. The ChNP is now studying the Platform’s health care model to adapt and implement it nationwide. It is an experience of collaboration and knowledge transfer between endemic and non-endemic countries.
| Chagas disease (CD) is caused by the parasite Trypanosoma cruzi and is one of 17 recognized neglected tropical diseases. Updated findings published in 2015 [1] estimate that between 6 and 7 million people worldwide have T. cruzi infection and that 25 million remain at risk of infection. Originally limited to Latin America, CD is now a global health problem as a result of migration flows from traditional endemic zones [2]. Bolivia has the highest prevalence in the world (6.1%), and the disease is endemic in 60% of the country [1]. The latest published data show that 607,186 persons in Bolivia are estimated to have CD and that a further 586,434 people are at risk, the number of new cases is estimated around 8,700 every year (8,087 for vector-borne transmission and 616 newborns with T.cruzi via congenital transmission).[1] Between 2 to 3% of those infected people develop cardiac and/or digestive complications every year. [3] Annually the number of estimated deaths caused by CD is around 388.[4] It is estimated that CD is responsible for 13% of deaths of people between 15 and 75 years. [3] However, the real magnitude of the problem remains unknown. Even the reduction in domiciliary infestation by the vector to below 3% in most parts of the country [5], it does not necessary correlate with advances in patient care. With rates below 3% of infestation we would have expected greater advances in the number of treated people, as treatment of positive patients is recommended by the ChNP when infestation rates are below this threshold. However the number of treated patients is very low and CD still remains an important unsolved public health problem.
The two drugs available for the aetiological treatment of CD are benznidazole and nifurtimox. While both are highly efficacious in newborns and children [6], there are concerns about their efficacy in adults [7], even if the results of recent studies indicate that efficacy is higher than previously reported [8] and that treatment decreases congenital transmission of T. cruzi.[9]
Considering that the lack of early treatment has important implications not only for individual patients, but also in terms of public health, implementation of a model to increase the number of adult patients diagnosed and treated and provide comprehensive care was prioritized [10,11]. Therefore, in 2009, a collaborative initiative was implemented by the Barcelona Centre for International Health Research (ISGlobal) and the Fundación Ciencia y Estudios Aplicados para el Desarrollo en Salud y Medio Ambiente (CEADES). This initiative was known as the Bolivian Chagas Platform.
The objective of this manuscript is to present the Chagas Platform as a model for comprehensive care of adults with CD and to show how its implementation can increase the number of people who have access to diagnosis and treatment. Healthcare coverage data are provided in order to quantify the impact of the initiative, and the limitations and lessons learned from this experience are described.
The prevention and control of CD in Bolivia was declared a national priority in Law 3374 dated March 23, 2006 [12]. However, no additional regulations were developed. The Chagas National Programme (ChNP), the responsible body for the prevention, diagnosis and treatment of CD in the country, elaborated drafted its Strategic Plan 2010–2015 [13]. Fig 1 provides a holistic and inter-sectorial overview of the Strategic Plan.
Given the epidemiological situation of Bolivia in the 1990s (infestation rates >50%), the priority area was vector control, which was financed by the Inter-American Development Bank (Banco Interamericano de Desarrollo, BID) from 1999 to 2006. This support included treatment of children but not adults with chronic CD, and the control of congenital disease (supported with funds from the Belgian Government until 2009) was prioritized.[14] Diagnosis and treatment of children under 18 years old should in theory be provided by primary healthcare centers in Bolivia, however many centers do not systematically provide this service. Even though CD affected mainly adults in Bolivia, no model for care of adults with chronic infection was defined, and therefore primary health center do not contemplate etiological treatment for this population.
In 2009, the ChNP reported that 178,012 persons had been screened. Most were pregnant women who were monitored for congenital transmission, although only 3,103 were treated (10% of those confirmed as having T. cruzi infection in the same year, representing only around 0.5% of all estimated patients with the infection) (Table 1).
Additionally, health professionals’ knowledge of CD was limited. Since their training included information on the very frequent adverse effects of aetiological treatment and the autoimmune origin of cardiac involvement [16], medical staffs were reluctant to recommend aetiological treatment of chronic CD in adults. Consequently, treatment of adults was neglected, as occurred in other endemic countries. Today, there is sufficient evidence on the role of T. cruzi in triggering and sustaining the inflammatory response [8, 16] and, therefore, on the importance of early anti-parasitic treatment.
Furthermore, the lack of information on the benefits of treatment, together with fear and an alternative understanding of CD by at-risk persons, limited patients’ active demand for treatment [17,18]. Additionally, access to healthcare was hampered by the absence of symptoms, the non-specific nature of symptoms when present, and limited access to health centres, especially in rural settings.
Following a change in international consensus, the Chagas Platform was developed as a joint initiative that arose from the need to offer diagnosis and anti-parasitic treatment to adult patients in the chronic phase of T. cruzi infection.[19,20] In 2009, ISGlobal (Barcelona, Spain) and CEADES (Cochabamba, Bolivia) pooled their expertise in the comprehensive management of adult patients by developing a care model with the Bolivian Ministry of Health and Bolivian state universities to collaborate in research and training of health professionals.
The primary objective of the Chagas Platform is to contribute to the control of CD, and the model designed to achieve this objective is based on 4 pillars:
The Chagas Platform is therefore considered a translational model in which provision of care is the initial trigger of research needs, thus initiating a circular cycle where the results of research are applied in to healthcare and are used to train staff and effect changes in health policy.
In this manuscript, we focus on comprehensive care and staff training as critical components for future scaling up of access to diagnosis and treatment.
Comprehensive healthcare based on agreed protocols was initially provided in vertical, dedicated structures for adults. Even when these structures were located in existing health structures, they were conceived as specific units for CD, instead of as units for integrating care of CD patients in the normal outpatient care circuit. This strategy made it possible to create centres of expertise and ensured sufficient capacity to increase the number of people diagnosed and treated. It simultaneously generated a critical mass of patients that allowing to pilot the use of comprehensive care strategies that could subsequently be integrated in the national health system and to advance in key research areas. Additionally, the results of the Program revealed the magnitude of the problem and the need for a national strategy for patients in the chronic phase of CD.
There are currently six centres in three highly endemic departments of Bolivia: Tarija (one), Chuquisaca (one), and Cochabamba (four: two in rural areas and two in urban or semi-urban areas). The centres were established with different organizational set-ups that varied depending on the local partners that committed to the project in each area. Although each set-up has its particularities, they all share the same protocols and database. The network of Chagas Platform centres in both rural and urban areas enables patients to be transferred from one geographical area to another. These circuits ensure better coverage and improved access to healthcare.
The Chagas Platform centres offer their services free of charge. The project was built in collaboration with the Spanish Agency for International Development Cooperation (AECID) and contributions from local partners. The ChNP covers drug costs. Of the 35 persons working in the six centres, two are covered by SEDES (the departmental health authority in Chuquisaca) and eight by the Juan Misael Saracho University (Tarija); the remaining 25 are covered by ISGlobal and CEADES with AECID funds. Local authorities are expected to progressively assume the cost of human resources in the future.
The success of the model relies on the protocols and clinical guidelines used (see Fig 2).
The main elements are as follows:
Another key element for the success of the program is the capacity of the staff to provide quality healthcare for CD patients. As health professionals were not sufficiently well trained in CD during their formal education, training of staff on current protocols became critical. After the first centres acquired expertise, the Chagas Platform started offering primary care staff a 1-week training program in the centres to learn the protocols.
Besides, the training and implementation of operational research included in the model has a relevant role giving to the National Health System personnel elements to analyze and reformulate priorities in Public Health interventions.
Finally, as the pilot project proved effective and acceptable, the Chagas Platform healthcare model has been expanded to primary healthcare centres since 2015, and a new strategy based on the network of centres managing CD was established, with the intention of expanding coverage in diagnosis and treatment in remote areas. The stages of development of the intervention have been reflected in Fig 3.
In addition to the adaptation of the protocols for managing CD in the health system care centres, this new horizontal approach was based on the training of health professionals (physicians, nurses, and biochemists) in these protocols and on referrals and counter-referral circuits between primary and specialized care centres.
Results from 2010 to 2015. Since the implementation of the Chagas Platform in 2009, a total of 26,227 adult patients have been attended in Bolivia, in the Platform centres. Around 69% had T. cruzi infection (18,316). To date, 8,567 patients have started treatment and, on average, 80% have received the complete course. Data regarding coverage of patients with T. cruzi infection are summarized in Table 2.
The number of patients was initially low because only two centres were functioning at the outset. The number of centres increased gradually until 2013, when the sixth and latest centre was opened. The increased demand has put strain on the system (organization, logistics, regulation of stocks, appointments) and the appropriate amount of medication has not always been available. Additionally, a benznidazole stock shortage in 2012 accounted for the low number of persons who started treatment during that year. Apart from the poor availability of drugs, which still limits the number of people who can be treated, the annual gap between persons with a positive diagnosis and persons treated can also be explained by the non-fulfilment of eligibility criteria and less importantly patient’s reluctance to be treated.
Strict fulfilment of inclusion criteria has improved adherence to treatment. On average, 80% of persons who initiated treatment finished the 60-day course, and conscientious follow-up ensured that data on adherence were excellent. Around 10% of patients left treatment voluntarily and around 10% of patients were advised to stop treatment owing to adverse drug reactions (ADRs) that were partially controlled with symptomatic treatment. Benznidazol and nifurtimox ADRs have been pointed out as one of the main problems for adherence to treatment, and due to the relevance of the topic, the description of them in the context of the Platform model will be described in deep in a separate manuscript.
The demand for the Platform was directly related to the implementation of community information activities. Since 2010, more than 25,000 people have received direct information about CD while in the healthcare process at Platform centres. Additionally, more than 3,500 people attended community information sessions. More than the half of these sessions was in the original language (mainly Quechua).
The percentage of people attending the centres that had T. cruzi infection was higher than expected, and most of those who were treated with benznidazole had excellent adherence to treatment.
The latest available data from the ChNP reveal the limitations of the ChNP for covering existing demand: in 2014, 29,052 adults were diagnosed with T. cruzi infection in Bolivia, and only 4,444 were treated (S1). Most of these patients (1,868, 42%) were treated in the Chagas Platform. [23]
In the areas where the Chagas Platform has its centres, yearly screening ranges from 0.7% to 2.1% of the estimated number of T.cruzi infected people. Almost two-thirds of current adult treatment is provided in the Platform centres, thus making them an important support structure for the ChNP in providing adult diagnosis and treatment.
Despite the quantitative and qualitative improvement in CD healthcare, the annual number of treated patients is less than 0.5% of the estimated total requiring treatment. There is also a considerable gap between people with T. cruzi infection and the number of people treated (only 10% among all the people with T. cruzi infection diagnosis). In this sense, the benefits of the project lie in the fact that it comprises a defined healthcare package, with concrete protocols to manage adults in the chronic phase of CD, that has proven effective and can be expanded to the National Health System.
Unfortunately, the total number of patients treated in Bolivia increased by only 14% between 2009 and 2015. While a favourable trend was observed for adults, the number of newborns and children treated decreased [15] Expanding the model to the primary health care system could reverse this trend. In collaboration with the Bolivian Ministry of Health, this is the next step proposed by CEADES and ISGlobal, with the support of AECID.
Since 2010, more than 1,600 health professionals have been trained in the specific management of patients with CD. A training fellowship program has been implemented between Universidad Mayor de San Simon (Bolivia) and Barcelona University, and five international professional exchanges have taken place since 2011. Specialized conferences and specific training activities were held during a conference in 2014, and more than 500 people participated.
Even if the Chagas Platform has proven to be highly effective, Platform centres have limited human resources to cover current demand in the management of CD in Bolivia. As recommended by the WHO, the main strategy for increasing access to healthcare requires the diagnosis and treatment of CD to be incorporated into the national health system, as part of their regular activities. In Bolivia, the recent establishment of a network including primary healthcare centres in rural areas following easy and realistic protocols is already showing positive results: it enables the management of CD in adults to be standardized and access to healthcare for people living in remote areas to be improved. To date, the integral CD healthcare model used in the Platform centres has been accepted for adaptation by the national health system before being implemented nationwide. Current Platform activities such as prevention, diagnosis, treatment, IEC, and training have been included in the implementation of the proposed model in national primary health centres. The main lessons learned from the implementation are as follows:
The increase in the number of people diagnosed and treated must be made carefully in order to avoid excessive strain on the health system, which has to adapt gradually to growing demand. Finally, as CD is a neglected disease, several external factors can hamper patient management. The poor availability of current drugs (benznidazole and nifurtimox) and the lack of new and better-tolerated medicines are key limiting factors. Furthermore, the lack of biomarkers of response to treatment hinders research on new drugs [24]. Fortunately, in the last five years, international public and private initiatives have been launched to develop new drugs and to implement clinical trials, as have studies focused on biomarkers of cure and/or response to treatment, some of which have been performed in the Chagas Platform.
Despite not being the only valid strategy, the Chagas Platform has proven to be a model of care for patients with T. cruzi infection that has been adopted by the Bolivian government. Moreover, the Chagas Platform has brought the benefits of reinforcing research capacity and training health professionals. Linking care, training, and research at the operational level is a very powerful tool that keeps health professionals updated and motivated. Additionally, the comprehensiveness of the program in the healthcare system should prove useful in other health-related issues.
The expertise that has formed the basis of guidelines and protocols is now being expanded to the national health system, thus highlighting the success of the program and enabling diagnosis and treatment to be scaled up.
|
10.1371/journal.ppat.1004528 | Experimental Cerebral Malaria Pathogenesis—Hemodynamics at the Blood Brain Barrier | Cerebral malaria claims the lives of over 600,000 African children every year. To better understand the pathogenesis of this devastating disease, we compared the cellular dynamics in the cortical microvasculature between two infection models, Plasmodium berghei ANKA (PbA) infected CBA/CaJ mice, which develop experimental cerebral malaria (ECM), and P. yoelii 17XL (PyXL) infected mice, which succumb to malarial hyperparasitemia without neurological impairment. Using a combination of intravital imaging and flow cytometry, we show that significantly more CD8+ T cells, neutrophils, and macrophages are recruited to postcapillary venules during ECM compared to hyperparasitemia. ECM correlated with ICAM-1 upregulation on macrophages, while vascular endothelia upregulated ICAM-1 during ECM and hyperparasitemia. The arrest of large numbers of leukocytes in postcapillary and larger venules caused microrheological alterations that significantly restricted the venous blood flow. Treatment with FTY720, which inhibits vascular leakage, neurological signs, and death from ECM, prevented the recruitment of a subpopulation of CD45hi CD8+ T cells, ICAM-1+ macrophages, and neutrophils to postcapillary venules. FTY720 had no effect on the ECM-associated expression of the pattern recognition receptor CD14 in postcapillary venules suggesting that endothelial activation is insufficient to cause vascular pathology. Expression of the endothelial tight junction proteins claudin-5, occludin, and ZO-1 in the cerebral cortex and cerebellum of PbA-infected mice with ECM was unaltered compared to FTY720-treated PbA-infected mice or PyXL-infected mice with hyperparasitemia. Thus, blood brain barrier opening does not involve endothelial injury and is likely reversible, consistent with the rapid recovery of many patients with CM. We conclude that the ECM-associated recruitment of large numbers of activated leukocytes, in particular CD8+ T cells and ICAM+ macrophages, causes a severe restriction in the venous blood efflux from the brain, which exacerbates the vasogenic edema and increases the intracranial pressure. Thus, death from ECM could potentially occur as a consequence of intracranial hypertension.
| Malaria remains one of the most serious health problems globally, but our understanding of the biology of the Plasmodium parasite and the pathogenesis of severe disease is still limited. Human cerebral malaria (HCM), a severe neurological complication characterized by rapid progression from headache to convulsions and unrousable coma, causes the death of hundreds of thousands of children in Africa annually. To better understand the pathogenesis of cerebral malaria, we imaged immune cells in brain microvessels of mice with experimental cerebral malaria (ECM) versus mice with malarial hyperparasitemia, which lack neurological impairment. Death from ECM closely correlated with plasma leakage, platelet marginalization, and the recruitment of significantly more leukocytes to postcapillary venules compared to hyperparasitemia. Leukocyte arrest in postcapillary venules caused a severe restriction in the venous blood flow and the immunomodulatory drug FTY720 prevents this recruitment and death from ECM. We propose a model for ECM in which leukocyte arrest, analogous to the sequestration of P. falciparum infected red blood cells in HCM, severely restricts the venous blood flow, which exacerbates edema and swelling of the brain at the agonal comatose stage of the infection, leading to intracranial hypertension and death.
| Plasmodium falciparum is responsible for an estimated 600,000 deaths annually, principally in children under the age of five [1]. Clinical symptoms range from intermittent fevers and chills to potentially fatal complications including severe anemia and cerebral malaria [2]. The mortality rate in comatose pediatric patients, most frequently due to respiratory arrest, is 15–20% despite optimal medical care [3], but the underlying pathology is unclear.
Molecular and cellular mechanisms involved in the pathogenesis of human cerebral malaria (HCM) include a predominantly pro-inflammatory cytokine profile, endothelial activation via the NF-κB pathway with upregulation of adhesion molecules, glia cell activation, and sequestration of infected red blood cells (iRBC), monocytes, and platelets within brain capillaries [3]–[6]. However, the cellular mechanisms associated with HCM cannot be directly observed in the human brain. Ophthalmological examination of the retinal pathology generally correlates with course and etiology of malarial encephalopathy [2], [7], but despite significant recent improvements [8], this technique lacks the resolution to observe the dynamic behavior of individual iRBC, leukocytes, and platelets, their exact location within the microvasculature, mechanisms of vascular leakage or possibly occlusion, and the sequence of these events. Elucidation of CM pathogenesis therefore requires the use of a robust small animal model that closely reflects clinical symptoms, histopathology, and immune mechanisms associated with the pathophysiology of HCM.
P. berghei ANKA (PbA) infected CBA, Swiss Webster, or CB57Bl/6 mice represent a well-characterized and widely used model for experimental cerebral malaria (ECM) that shares a number of similarities with P. falciparum HCM [5], [6], [9]–[12]. Both ECM and HCM are characterized by severe vasculopathy, i.e. endothelial activation and dysfunction with increased expression of adhesion molecules such as ICAM-1, VCAM-1, and E-selectin, upregulation of inflammatory cytokines, reduced blood flow, vascular leakage, acute edema of both vasogenic and cytotoxic origin, and microhemorrhages leading to neurological impairment [12]–[17]. Platelet activation, dysregulation of the coagulation cascade, thrombocytopenia, and platelet accumulation in the brain are also found in both HCM and ECM [18]–[20]. We have previously shown by intravital microscopy (IVM) that platelet marginalization and blood brain barrier (BBB) disruption are central to ECM pathophysiology [21]. Platelets are thought to impair vascular repair and increase BBB permeability by potentiating the iRBC-induced endothelial damage in the early stages of HCM development [22]–[24]. Circulating platelet-derived microparticles are increased in severe P. falciparum malaria and serve as a biomarker for neurological involvement [13], [25].
The murine PbA model has also provided ample evidence for a contribution of CD8+ and CD4+ T cells to the late stages of ECM development [26]–[30]. Both CD8+ T cells, generally considered the terminal effector cells, and CD4+ T cells must accumulate in the cerebral microvasculature for ECM to occur [11], [27], [28], [31]–[35] and may also be responsible for the ECM-associated leukocyte infiltration [36]. While ECM development was thought to involve CD8+ T cell-induced endothelial apoptosis via perforin- and granzyme B-mediated cytotoxicity resulting in BBB disruption [32], [33], [37], we recently showed by IVM that ECM closely correlates with widespread opening of the BBB and that this occurs in the absence of significant endothelial death [21]. The BBB at the level of postcapillary venules encompasses two layers, the vascular endothelium with its basement membrane and the glia limitans with associated basement membranes and astrocyte endfeet, which are separated by the perivascular space [38]. This section of the BBB is functionally distinct from other areas of the BBB, for example that at the capillary level, which consists of a single layer composed of endothelia, gliovascular membrane, and astrocyte endfeet [38]. IVM also revealed that ECM correlates with platelet deposition, leukocyte arrest, and de novo expression of the pattern recognition receptor CD14 on the endothelial surface from postcapillary venules, but not from capillaries or arterioles [21]. Strikingly, inhibition of platelet deposition and leukocyte recruitment by blockage of LFA-1 mediated cellular interactions prevented ECM and disruption of the BBB in PbA-infected mice [21]. Thus, it appears that the ultimate cause of coma and death in ECM is a universal breakdown of the BBB at the level of postcapillary venules [21].
In the PbA-infected CBA/CaJ mouse model, vascular leakage, neurological signs, and death from ECM can be prevented by treatment with the endothelial barrier-stabilizing sphingosine 1 analog FTY720 (fingolimod) [21], [39], an immunomodulatory FDA-approved drug for oral treatment of relapsing multiple sclerosis (MS) [40] that acts as an agonist for sphingosine 1-phosphate (S1P) receptors [41]. In experimental autoimmune encephalomyelitis (EAE), FTY720 prevents T cell recruitment to the brain by down-modulating the expression of S1P1 receptors on the T cell surface. This favors the CCR7-mediated retention of naïve and central memory T cells within secondary lymphatic tissues [42], leading to a reduction in the numbers of naïve and central memory T cells, but not effector memory T cells, in the blood [43]. FTY720 may also prevent stimulation of vascular endothelia or activation of CD8+ effector T cells in the spleen by decreasing CD11c+ DC migration and function and by destabilizing DC/T cell interactions thus preventing the formation of an immunological synapse [44], [45]. In addition to its involvement in T cell activation and targeting to the brain, FTY720 is also thought to have a directly stabilizing effect on endothelial junctions at the BBB [46]–[49]. However, the exact mechanism by which FTY720 prevents BBB opening remains unclear to date.
Here, we show that ECM is associated with the accumulation of numerous leukocytes within postcapillary and larger venules and that the resulting microrheological alterations severely restrict the venous blood flow. Treatment with FTY720 significantly reduced the recruitment of these leukocytes indicating their involvement in the pathogenesis of ECM [21], [39]. Leukocyte arrest likely increases the intracranial pressure, similarly to P. falciparum iRBC sequestration in pediatric HCM, which is typically associated with a poor clinical outcome [50].
Three week-old CBA/CaJ mice were infected with PbA-GFP or P. yoelii 17XL (PyXL)-RFP and monitored throughout the course of development of ECM or hyperparasitemia [21]. In this study, a total of 78 PbA-infected mice were subjected to IVM at the time of ECM (day 6–8), 18 PbA-infected mouse before ECM (day 5), 15 PbA-infected mice that failed to develop ECM (day 9), and 25 uninfected control mice (Table S1). We also examined 30 PbA-infected mice that were treated with FTY720 and failed to develop ECM (day 8–9) and 8 PbA-infected mice that developed ECM on day 8 despite treatment with FTY720. As controls for parasitemia, 62 PyXL-infected mice with >50% parasitemia (day 5) were analyzed. Throughout our IVM experiments, we examined deep microvessels, i.e. branches of penetrating arterioles and venules [51], [52], which were in direct continuation with the cortical capillary bed. Imaging of CX3CR1GFP/+ mice confirmed that the postcapillary venules, capillaries, and arterioles used for analysis are embedded in fluorescent microglia and thus clearly located in the cerebral cortex, i.e. underneath the pial microvasculature (Figure S1) [21].
IVM revealed that the venous blood flow in postcapillary venules from PbA-infected mice with neurological signs (day 6) was strikingly altered. Vascular labeling with Evans blue revealed that postcapillary venules from mice with ECM exhibited a marginal zone devoid of RBCs (Figure 1A, Video S1). Instead, this zone contained variable numbers of leukocytes that were either rolling along the endothelium, crawling, or firmly attached (Figure S2A and S2C, Video S2). Minimal projections of time sequences emphasize the boundary between the functional lumen in the center of the postcapillary venules and the RBC-free marginal zone and suggest that the functional lumen available for the blood flow is significantly restricted during ECM (Figure S2B and S2D). This phenomenon was even more pronounced in larger venules. Neither arterioles from mice with ECM (Figure 1B, Video S3) nor postcapillary venules or arterioles from mice with hyperparasitemia (Figure 1C and 1D, Video S4) showed any significant functional vascular restriction, i.e. narrowing of the passageway available for the blood flow, compared to uninfected control mice (Figure 1E and 1F, Video S5), a finding we attribute to the absence of steric hindrance generated by adherent leukocytes in these vessels. Multiple measurements of the total vascular diameter (from endothelium to endothelium) and the functional diameter (used by the blood flow) of 50 randomly chosen postcapillary venules from 4 mice with ECM revealed a mean functional diameter of 70.5±13.7% compared to data from 3 uninfected control mice, corresponding to a functional vascular cross-section of 55.8±19.1% (Figure 1G). Notably, complete vascular occlusion, whether in postcapillary venules or other microvessels, was not observed during ECM. In 50 randomly chosen postcapillary venules from 3 PyXL-infected mice with hyperparasitemia, the functional postcapillary venule diameter and cross-section was 95.5±3.4% and 92.7±5.1%, respectively. No microrheological alterations were found in postcapillary venules or arterioles from PbA-infected mice prior to ECM (day 5), in PbA-infected mice that failed to develop ECM (day 9), or in uninfected control mice. As reported previously [21], PbA-infected mice that did develop ECM despite treatment FTY720 exhibited vascular leakage suggesting that the venous blood flow restriction was similar to untreated PbA-infected mice with ECM.
Thus, ECM correlates with a significant functional constraint, but not complete blockage, of the passageway available for the venous blood flow. This is significant, because any restriction in the venous efflux from the brain likely exacerbates edema formation, a hallmark of both ECM and HCM [16], [17], [53]. A venous efflux problem would also explain the increased intracranial pressure, which is frequently observed in pediatric CM in Africa [50]. Indeed, MRI imaging has identified increased intracranial pressure as the strongest predictor of death [54], [55].
Quantitative offline IVM analysis of the various cell densities revealed that leukocytes are recruited to the cortical microvasculature not only in response to ECM, but also hyperparasitemia, albeit at a significantly lower density (Figure 2). Specifically, more CD8+ T cells, neutrophils, and macrophages were recruited in ECM compared to hyperparasitemia, while the density of all other cell types analyzed did not differ between these two infections (Figure 3A). The cortical microvasculature of uninfected control mice exhibited virtually no arrested leukocytes suggesting that in the absence of an inflammatory stimulus, innate immune cells do not monitor the BBB.
To elucidate the composition of specific cellular subtypes involved in pathology we quantified leukocytes by flow cytometry in perfused whole brains. In contrast to IVM, no significant difference in CD8+ T cell recruitment was observed between ECM and HP suggesting that flow cytometry may lack the sensitivity to distinguish important focal variations in cellular composition as arrested leukocytes were not observed in other vessels such as capillaries or arterioles.
Overall, significantly more CD45+ leukocytes were found in the brains during ECM compared to hyperparasitemia (23729.3±7573.8 vs. 4483.0±2971.6; ANOVA: F(2) = 12.42; P<0.001, Tukey's test: T = −4.49; P<0.05) and PbA/FTY720 mice (6059.7±4070.7; Tukey's test: T = −4.12; P<0.05) (Figure 3C, Table S7). Confirming the IVM data, ECM was associated with the recruitment of significantly higher numbers of Ly6G+ neutrophils than in hyperparasitemia (1470.7±325.5 vs. 518.0±317.2; ANOVA: F(2) = 9.33; P<0.05), while there was no significant difference in the number of Ly6C+ monocytes.
The largest increase in cell numbers was observed for F4/80+ macrophages during ECM compared to hyperparasitemia (9648.0±3432.1 vs. 938.0±645.9; ANOVA: F(2) = 14.35; P<0.01). Equivalent results were obtained for CD11b+ macrophages (ANOVA: F(2) = 6.50; P<0.01). Because total brain leukocytes necessarily contain a large proportion of parenchymal macrophages, we distinguished these from blood-derived macrophages by their low level of CD45 expression [36], [56]. When the CD45lo parenchymal macrophages (microglia) were excluded, the number of the remaining mostly intravascular CD45hi F4/80+ macrophages was significantly higher during ECM compared to PyXL-infected mice (1750.0±285.0 vs. 243.3±171.5; ANOVA: F(2) = 30.19; P<0.01) (Figure 3D).
FTY720 treatment of PbA-infected mice significantly reduced the number of Ly-6G+ neutrophils (Tukey's Test: T = −3.69; P<0.05), and both total and CD45hi F4/80+ macrophages (Tukey's test: T = −4.31; P<0.05) in PbA-infected mice so that no significant difference was found for any of the cell types between PbA/FTY720 mice on day 9 and PyXL-infected mice with hyperparasitemia on day 5 (Figure 3B-D, Table S7). Equivalent results were obtained for CD11b+ macrophages (Table S7). Because neither FTY720-treated PbA-infected mice nor PyXL-infected mice with hyperparasitemia exhibit vascular leakage or neurological signs, it appears that FTY720 prevents BBB opening and the associated leukocyte recruitment, although we cannot exclude that FTY720 affects the brain directly.
Flow cytometry revealed two distinct leukocyte subsets, namely CD45hi and CD45lo (Figure 4A), both of which were significantly more numerous in the brains of PbA-infected mice (ANOVA: F(2) = 10.38; P<0.05 and ANOVA: F(2) = 27.21; P<0.01, respectively) compared to PyXL-infected mice (Tukey's: T = −4.45; P<0.05 and Tukey's test: T = −7.37; P<0.01, respectively) (Table S8). FTY720 treatment significantly reduced the number of both CD45hi (Tukey's test: T = −3.08; P<0.05) and CD45lo (Tukey's test: T = −3.99; P<0.05) leukocytes. While the number of CD45hi cells after FTY720 treatment was not statistically different from PyXL infection, the number of CD45lo cells, although significantly decreased compared to PbA infection, remained significantly higher compared to PyXL-infected mice with hyperparasitemia (Tukey's test: T = −3.38; P<0.05) (Table S8). Significantly more ICAM-1+ (Table S9) and CD69+ leukocytes (Table S10) were present in the CD45hi and the CD45lo leukocyte subsets from PbA-infected mice compared to PyXL-infected mice (Figure 4B–D). FTY720 treatment significantly reduced the number of ICAM-1+, CD69+, and GrB+ leukocytes compared to PbA-infected mice with ECM (Table S9, S10, and S11). Further, the CD45hi subset from PbA-infected mouse brains contained consistently higher numbers of ICAM-1+, CD69+, and GrB+ CD8+ T cells compared to PbA/FTY720 mice, although this difference was not statistically significant (Figure S4A, Table S12). Furthermore, the median expression levels of ICAM-1, CD69 and GrB in these CD45hi CD8+ T cells were similar amongst PbA-infected, PbA/FTY720, and PyXL-infected mice (Figure S4B-D). Likely, the ECM-associated vasculopathy is caused by the high density of activated leukocytes in the postcapillary venules.
P-selectin release from platelet α-granules or endothelial Weibel-Palade bodies promotes the binding of platelets, leukocytes, and plasma proteins to the vascular wall [71]–[73]. Because both platelet marginalization and P-selectin expression have been implicated in the pathogenesis of both HCM and ECM [18], [20]–[22], [74]–[78], we determined the distribution of this adhesion molecule with respect to arrested platelets in the cortical microvasculature. Upon manifestation of neurological signs, PbA-infected CBA/CaJ mice were inoculated with a PE-conjugated mAb against P-selectin (CD62P), eFluor 450-conjugated anti-CD41 to detect platelets and Evans blue to visualize the vascular lumen [21]. IVM revealed small clusters of marginalized platelets that colocalized with patches of P-selectin on cortical postcapillary venule endothelia (Figure 7, Video S20). Occasionally, we observed strings of platelets that appeared to be attached to clusters of platelets (Video S21) as has been suggested to occur in HCM based on in vitro experiments [79]. In contrast, PyXL-infected mice with hyperparasitemia showed no evidence for P-selectin expression or platelet arrest (Figure 7, Video S22). Unlike postcapillary venules, arterioles were consistently negative for P-selectin or arrested platelets, both during ECM and hyperparasitemia. Thus, ECM, but not hyperparasitemia, is associated with marginalization of small numbers of platelets along postcapillary venule endothelia and P-selectin release, either from platelets or endothelia. However, the highly focal nature of both platelet arrest and P-selectin release contrasts with the uniform endothelial activation as evidenced by CD14 expression, ICAM-1 upregulation, and vascular leakage observed during ECM. Thus, leukocyte arrest is not limited to the P-selectin positive portions of the postcapillary venule endothelia.
FTY720 was previously shown to prevent vascular leakage, neurological signs, and death from ECM [21], [39]. To evaluate whether FTY720 protects the BBB by preserving the integrity of endothelial junctions [48], [80], [81], we determined the expression level of the tight junction (TJ) proteins claudin-5, occludin, and ZO-1 in the cerebral cortex and the cerebellum of 4 PbA-infected mice with ECM (day 6–8), 3 FTY720-treated PbA-infected mice that did not exhibit any neurological signs (day 8 or 9), and 3 PyXL-infected mice with hyperparasitemia (day 5) (Figure S6 and S7). Quantification of the fluorescence emission of specific antibodies on 3–4 immunolabeled cryostat sections per experimental condition yielded no significant reduction in protein expression under the different infection and treatment conditions compared to 3 uninfected control mice (Table S16). This finding suggests that the TJs remained morphologically intact and supports the hypothesis that the ECM-associated vascular leakage is based on a regulated, potentially reversible, mechanism of BBB opening [21], [82].
Thus, comparison of two Plasmodium infection models revealed: 1) The venous blood flow impairment during ECM is caused by the arrest of significantly higher numbers of CD8+ T cells, neutrophils, and in particular macrophages in cortical postcapillary venules compared to hyperparasitemia. While a small number of CD8+ T cells and macrophages extravasated into the perivascular space, most of the recruited leukocytes remained intravascular. 2) FTY720 treatment of PbA-infected mice reduced, but did not completely prevent leukocyte accumulation in postcapillary venules, which is consistent with the finding that low numbers of arrested leukocytes are present in PyXL-infected mice with hyperparasitemia without causing vascular leakage or neurological signs. 3) ECM closely correlates with the recruitment of large numbers of ICAM-1 expressing F4/80+ macrophages to the brain. As FTY720 treatment did not reduce the ICAM-1 expression level, the high density of these macrophages in postcapillary venules likely enhances the ECM-associated vascular pathology. 4) Leukocyte recruitment coincides with the onset of neurological signs, but follows BBB opening, as vascular leakage can be observed 1 day prior to symptomatic ECM [21].
In this study, we identify a novel key determinant of ECM pathogenesis, namely that leukocyte arrest along the wall of postcapillary venules causes microrheological alterations that severely impair the venous blood flow. Based on our findings, we hypothesize that infection with PbA opens the BBB, which leads to the recruitment of numerous activated CD8+ T cells, ICAM-1+ macrophages, and neutrophils (Figure 8). The resulting steric hindrance of the blood flow in postcapillary and larger venules impairs, but does not block, the venous efflux from the brain, which exacerbates the vasogenic edema and causes death as a consequence of intracranial hypertension.
Under physiological conditions, the luminal surface of vascular endothelia is covered with a glycocalix, a 0.5 to >1 µm layer of membrane-bound proteoglycans and glycoproteins that repels RBCs and is critically involved in inflammatory responses, blood coagulation, and blood flow regulation [83]–[86]. IVM visualizes this glycocalix as a thin red layer, covering arteriolar endothelia from infected and uninfected mice and postcapillary venule endothelia from uninfected control mice. During ECM, the thickness of the RBC-free layer in postcapillary and larger venules was drastically increased. Because the glycocalix typically degrades under inflammatory conditions, leading to exposure of adhesion molecules, leukocyte adhesion, and impairment of endothelial barrier function [85], [86], the restriction in the venous blood flow during ECM is likely not caused by components of the glycocalix, but by increasing numbers of arrested leukocytes that prevent RBC from approaching the endothelium. Although the functional cross-section of postcapillary venules was occasionally reduced by more than 80%, complete vascular obstruction was not observed. These findings argue in favor of a combined vascular sequestration and immuno-pathological etiology of ECM [87]–[89].
The reduction in the venous blood flow must be expected to have major consequences for the physiology of the brain. First, the overall hypoperfusion of the brain, enhanced by inadequate contact between RBCs and the endothelium, likely contributes to the drastically reduced O2 delivery to the cerebral parenchyma observed in ECM-susceptible C57BL/6 mice [90]. In addition, by increasing the wall shear stress, leukocyte adhesion is expected to reduce the blood volume flow in postcapillary venules dramatically [91]. Finally, a reduction of the venous efflux from the skull, caused by a generalized narrowing of the lumen of venous microvessels, necessarily increases the intracranial pressure. The finding that brains from mice with ECM, but not hyperparasitemia, are swollen and spongy and bulge out of the skull, if the Dura mater is accidentally damaged during craniotomy clearly documents the dramatically increased intracranial pressure during the agonal phase of the disease. The reduced venous efflux from the brain may exacerbate vascular leakage, brain edema, and hemorrhages - cerebral alterations that are also associated with HCM [92]. Brain swelling and edema is extremely common in adult HCM on CT scan [93], [94]. Increased intracranial pressure has long been associated with poor prognosis and neurological sequelae in severe pediatric HCM [50], [95]–[98]. In fact, recent longitudinal MRI observations in Malawian children have identified intracranial hypertension as the single most important MRI finding associated with HCM development and the most reliable predictor of death [54], [99].
PbA-infected mice also exhibit arteriolar vasospasms during the final stage of ECM [100] and it has been suggested that the reduction in the cerebral blood flow observed by MRI [101] is due to increased production of vasoconstrictive factors or inhibition of vasodilating mediators [102]. Subsequent work [103], [104] revealed that endothelin-1 (ET-1), a potent vasoactive peptide with inflammatory and platelet-activating properties [105]–[108], is upregulated during both ECM and HCM [109]–[112]. Indeed, the arteriolar vasoconstrictive effect of ET-1 could be responsible for ECM induction in the PbA-infected C57BL/6 mouse model [108], because injection of exogenous ET-1 induces neurological signs in PbNK65-infected mice, which normally do not develop ECM [113], and because blockage of the ET-1 receptor A prevents ECM development in PbA-infected mice [110]. However, ET-1 has a plasma half-life of well under one minute in rodents [114] so that vasoconstriction alone cannot explain the increased intracranial pressure observed during ECM. Because ET-1 also stimulates endothelial activation with upregulation of adhesion molecules, promotes leukocyte adhesion, and increases vascular permeability [105], [106], there is a possibility that ET-1 induces ECM by restricting the venous blood flow. Similarly, administration of nitric oxide (NO), a key messenger involved in regulation of platelet adhesion and inflammatory and immune responses [115], decreased both leukocyte accumulation and vascular resistance in larger venules of PbA-infected mice [78], [116]–[121]. We conclude that the ultimate cause of death from ECM is a combination of arteriolar vasoconstriction and severe reduction in the blood efflux from the brain due to leukocyte adhesion in the venous microvasculature.
Compared to PyXL-infected mice with hyperparasitemia, PbA infection triggered the recruitment of significantly more CD8+ T cells to postcapillary venules at the time of, but not prior to, ECM development. Together with the finding that CD8+ T cells were absent in mice that survived the critical time for ECM development, these data suggest that neurological signs and T cell recruitment are correlated and occur rapidly. CD8+ T cells from C57BL/6 mice immunized with PyXNL can confer protection against lethal PyXL infection [122], whereas CD8+ T cell accumulation in the brain of PbA-infected C57BL/6 mice was abolished and the mice were completely protected from ECM when co-infected with P. yoelii [123], [124]. However, similar numbers of CD8+ T cells accumulated in the brains of PbA-infected C57BL/6 mice with ECM and PbNK65-infected mice without neurological signs [34], further supporting the notion that ECM-eliciting parasites such as PbA induce the recruitment of a qualitatively different CD8+ T cell population to the brain.
FTY720 treatment decreased the number of CD8+ T cells to levels similar to those found in PyXL-infected mice. Despite the presence of the remaining CD8+ T cells, neither FTY720-treated PbA-infected mice nor PyXL-infected mice developed neurological signs. A small percentage of CD8+ T cells entered the perivascular space during ECM, but not hyperparasitemia. This could be explained by upregulation of the leukocyte common antigen CD45, because CD45 expression is typically enhanced in response to stress signals, leading to increased leukocyte motility [125] and brain infiltration, for example after seizure [126]. ECM coincided with larger numbers of CD8+ T cells expressing CD69, one of the earliest lymphocyte activation markers [127], [128], and FTY720 treatment reduced the number of CD69+ CD8+ T cells to levels similar to those found during hyperparasitemia. Previous work supports a correlation between ECM and CD69+ CD8+ T cells: 1) Recruitment and activation of CD8+ T cells and CD69 expression were reduced in ECM-resistant mice [129]. 2) Peripheral CD8+ T cells were predominantly CD69+ during ECM and expressed the phenotype of memory T cells [130]. 3) The ECM-associated upregulation of CD69 was reversed and disease was prevented by interference with the angiotensin I pathway [131]. 4) Ghanaian P. falciparum infected pediatric patients with clinical HCM or severe anemia showed similar T cell activation profiles with a significantly increased frequency of CD69+ cells compared to asymptomatic children [132]. The median expression levels per cell of CD69 and GrB did not differ between the experimental groups suggesting that ECM pathogenesis correlates with a high density of CD69+ and GrB+ CD45hi CD8+ T cells in postcapillary venules. FTY720 treatment of PbA-infected mice reduced the number of CD4+ T cells to levels similar to those observed for PyXL-infected mice. Together, these data suggest that FTY720 treatment prevents ECM by inhibiting the trafficking of activated CD8+ and CD4+ T cells to the brain.
FTY720 treatment of PbA-infected mice also reduced the number of macrophages and neutrophils to levels similar to those found in the brains of PyXL-infected mice, supporting the notion that these leukocytes exacerbate edema formation during ECM [26], [27]. Of particular importance, FTY720 treatment prevented the recruitment of large numbers of ICAM-1+ macrophages to the brains of PbA-infected mice. This finding sheds light on an unexpected new role of ICAM-1 in the pathogenesis of ECM. While FTY720 may preserve the integrity of the BBB primarily by preventing leukocyte recruitment to the brain, activated platelets release phosphorylated FTY720 [133], [134], which acts as a full agonist on the endothelial S1P receptor S1PR4; therefore, FTY720 may prevent the ECM-associated vascular leakage by strengthening the endothelial actin cytoskeleton [135], [136]. Further, FTY720 may regulate endothelial barrier function by directly modulating endothelial junction tightness, because the increased vascular leakage observed in mice deficient in plasma S1P can be reversed by restoring plasma S1P levels [47], [48], [80], [81]. This finding may explain why FTY720 administration must be started prior to the onset of vascular leakage [21], [39] and why attempts to rescue mice with symptomatic ECM, when leukocyte recruitment was already in progress, were unsuccessful (A. Movila and U. Frevert, unpublished observations). Further, FTY720 can cross the BBB and may thus be able to directly modulate parenchymal cells by interacting with S1P receptors in the CNS. The resulting feedback from the CNS on the activation status of the BBB may in turn alter the interaction with the immune cells. Future testing of FTY720 or related compounds in the ECM model is expected to reveal more detail on the exact mechanism of BBB opening and vascular inflammation in ECM. The ECM model may also improve our understanding of the pathogenesis of HCM. Ugandan, Malawian, and Central Indian children with HCM exhibit decreased S1P plasma levels compared to those with uncomplicated malaria and a low angiotensin-1 to angiotensin-2 plasma ratio discriminates HCM and severe non-cerebral from uncomplicated malaria and also predicts mortality from HCM [39], [137]–[140]. As these reports strongly suggest the involvement of the S1P pathway HCM, screening for novel immunomodulatory drugs and exploration of their endothelial barrier-promoting effects is warranted.
We observed significantly more neutrophils in postcapillary venules and whole brain during ECM compared to hyperparasitemia. Considering that the role of neutrophils in the pathogenesis of CM is understudied to date, this finding is of particular interest. FTY720 treatment of PbA-infected mice reduced the number of neutrophils significantly so that levels similar to those found during hyperparasitemia were reached, which supports the previously suggested involvement of neutrophils in the ECM-associated vasculopathy and edema formation [26], [141].
The number of arrested monocytes was increased similarly in PbA- and PyXL-infected mice compared to uninfected control mice suggesting that monocyte recruitment correlates with Plasmodium infection in general, not ECM in particular. This finding is in agreement with earlier reports showing that monocytes are not involved in iRBC accumulation in the brain at the time of ECM [26], [27]. Intravenous injection of fluorescent anti-CD14 revealed that monocytes were generally confined to the vascular lumen, although we occasionally detected labeled monocytes in the Virchow-Robin space. Similarly, intravenous injection of the macrophage marker anti-CD11b resulted in labeling of PVM. Because part of the PVM population derives from blood monocytes [53], these fluorescent monocytes may have been labeled before extravasating to replenish the pool of PVM in the perivascular space.
Flow cytometric measurement of leukocyte recruitment to the brain essentially corroborated our IVM findings suggesting that the cortical microvasculature reflects the ECM-associated pathological events in the entire brain. A few conceptual differences between IVM and flow cytometry are noteworthy. First, IVM confirmed the notion that the healthy murine brain has an extremely low level of immune surveillance with the almost complete absence of T cells, neutrophils, monocytes, and B cells [21], [142]. Expression of adhesion molecules is low on the healthy human postcapillary venule endothelium, and upregulation of P-selectin, E-selectin, ICAM-1, and VCAM-1 contributes to the pathogenesis of HCM [53], [58], [143]. Second, IVM provides information on the location of recruited leukocytes and their dynamics with respect to the complex and heterogeneous microvascular network of the brain. Third, intravenous labeling generally visualizes only intravascular leukocytes including those that extravasate between marker injection and imaging. However, the BBB at the level of postcapillary venules is naturally leaky, i.e. it allows passage of macromolecules including immunoglobulins into the Virchow-Robin space [38]. During ECM, this barrier becomes even more permeable, which makes distinction of recently extravasated from resident perivascular cells difficult. While CD8+ T cells do not patrol the cerebral microvasculature of uninfected mice and are therefore unlikely to extravasate into the perivascular space of naïve mice, the fluorescent CD11b+ macrophages we found in the perivascular space of mice with ECM could either have been labeled intravascularly and then extravasated or they could have been labeled after entering the perivascular space due to diffusion of the cell-specific markers across the BBB. Thus, while the majority of leukocytes remain confined to the vascular lumen during ECM, a small number of CD8+ T cells and possibly CD11b+ macrophages extravasate into the perivascular space. Fourth, flow cytometric measurement of the expression level of CD45 allowed distinction between blood-derived and parenchymal macrophages [56]. Using this approach, we found that ECM is associated not only with significantly higher numbers of ICAM-1+ CD45hi ( = blood-derived) macrophages, but also ICAM-1+ CD45lo ( = parenchymal) macrophages compared to mice with hyperparasitemia and FTY720-treated PbA-infected mice. Together with the ECM-associated activation of microglia [144] and the enhanced expression of CCR5 on non-hematopoietic cells [145], ICAM-1 upregulation on both blood-derived and parenchymal macrophages supports the notion of a mixed vasogenic and cytotoxic nature of the edema in ECM [17], [146]. Finally, IVM was critical to visualize endothelial activation markers within defined parts of the microvascular tree. Thus, IVM and flow cytometry provided highly complementary results.
BBB opening in the young CBA/CaJ mice used here begins at least one day prior to the onset of neurological signs, while leukocyte recruitment to the brain occurs only afterwards [21]. In 129,B6 mice, neither antibody-mediated neutrophil depletion nor clodronate-mediated macrophage depletion, done 1 or 2 days prior to ECM development, respectively, prevented neurological signs [26], [145]. While this was interpreted as neutrophils and macrophages not being involved in ECM development, an alternative explanation is that the BBB was already leaky at this late time point and the resulting vasogenic edema had already caused microglial activation. Clodronate liposomes do not cross the BBB [147]–[149] so that the microglia was likely unaffected and the resulting cytotoxic edema not prevented in these studies. Further, depletion of macrophages and granulocytes by administration of AP20187 to MAFIA mice on days 5, 6, and 7 after infection resulted in a >80% reduction of these cells in the blood [27], [150]. Again, as AP20187 does not cross the BBB [151], ECM development likely manifested because the activated microglia was not eliminated. Importantly, CD8+ T cells alone failed to induce any of the typical signs of ECM including convulsions and death without the direct neurotoxic effect of intravenously administered folic acid [26], [145]. Thus, it appears that both hematogenic and parenchymal macrophages play a role of in the pathogenesis of ECM. A more recent study in C57BL/6 mice emphasizes the crucial role of monocytes/macrophages in lymphocyte recruitment to the brain. Clodronate treatment 2 days prior to manifestation of neurological signs reduced the recruitment of CD8+ T cells, CD4+ T cells, and NK cells to the brain 2.8-fold, 1.8-fold, and 4.6-fold, respectively, and failed to alter the course of disease development [152]. Clodronate treatment 2 days prior to infection with PbA, on the other hand, prevented ECM development [152]. Thus, the exact mechanism by which blood-derived macrophages contribute to disease development needs further and more specific investigation. Taken together, these data suggest the following scenario: PbA infection induces BBB opening, which results in vasogenic edema and microglial activation. The ensuing cytotoxic edema then leads to endothelial activation with recruitment of leukocytes that in concert restrict the venous blood flow, which further weakens the already impaired BBB. Eventually, these events culminate in the typical irreversible damage associated with fatal malarial encephalopathy. However, the differential progression of the vasogenic versus the cytotoxic edema and their relative contribution to brainstem compression and death clearly require further experimental attention.
The critical role of ICAM-1 in the pathogenesis of severe P. falciparum malaria and the binding of P. falciparum iRBC to the endothelium is generally accepted [59], [61], [153]–[159]. However, comparison of the endothelial ICAM-1 expression levels in the brains of P. falciparum versus P. vivax infected individuals is necessary to determine whether or not HCM correlates with upregulation of ICAM-1 in postcapillary venules. ICAM-1 upregulation has also been implicated in the pathogenesis of ECM [63], [77], [78], [160]–[162]. While the cellular origin of ICAM-1 was not determined, treatment of PbA-infected mice with NO reduced ICAM-1 upregulation in the brain along with leukocyte sequestration and vascular leakage [78]. We show that upregulation of endothelial ICAM-1 correlates with Plasmodium infection in general, not with ECM in particular. Because ICAM-1 upregulation was not observed in mice deficient in RAG-1 or IFN-γ on day 6 after infection with PbA [163], the Th1 type cytokines IFN-γ and lymphotoxin, in synergy with TNF from macrophages [163], [164], may induce ICAM-1 expression throughout the brain microvasculature of ECM-susceptible mice, both during ECM and hyperparasitemia. Together with the finding that FTY720 treatment of PbA-infected mice does not prevent ICAM-1 upregulation in postcapillary venules, these data argue against a role of this endothelial adhesion molecule in the pathogenesis of ECM.
CD14 appeared on the surface of postcapillary venule endothelia selectively at the time of neurological signs (day 6–8), which is in agreement with the well-known involvement of CD14 in endothelial activation, neuroinflammation, and leukocyte recruitment to the cerebral microvasculature [165]–[168] and the finding that CD14-deficient mice are protected against ECM [169]. Further, because CD14 plays an important role in the non-phlogistic clearance of apoptotic cells [170], its expression on postcapillary venule endothelia may reflect the well-documented increase in apoptosis during ECM [171]–[173]. Interestingly, malaria-associated microparticles carry phosphatidylserine on their surface [174]. Because the plasma levels of endothelial, platelet, and erythrocytic microparticles increase at the onset of the ECM-associated neurological signs [175], CD14 may be involved in the recruitment of microparticles to postcapillary venule endothelia. Platelet-derived microparticles are known to promote macrophage differentiation [176] and may contribute to the observed upregulation of ICAM-1 on the macrophage surface.
We show that ECM occurs without apparent physical degradation of endothelial TJs, measurable loss of claudin-5, occludin, and ZO-1 and in the absence of endothelial death [21]. Together, these findings argue against widespread and irreversible endothelial injury, for example due to cytotoxic T cell-mediated apoptosis or long-lasting vascular occlusion [177], as a major pathogenetic mechanism leading to malarial coma and death [82], [88], [178], [179]. Instead, and in agreement with the observation that ECM-susceptible mice can be rescued by anti-LFA-1 treatment even minutes before imminent death [180], [181], the preserved TJ integrity suggests that BBB opening during ECM is under the control of a regulated mechanism [21], [82], [182]. In support of this notion, fast-acting anti-malarial drugs can prevent ECM one day before the expected onset of neurological signs, although CD8+ T cells still accumulate in the brain [32], [34].
Platelets play a crucial role in the early stages of the pathogenetic cascade of ECM [9], [183], [184]. Blockage of platelet activation or shedding of platelet-derived microparticles confers resistance to ECM [13], [22], [160], [184]. At the time of ECM, platelets arrest focally and in small numbers in the cortical microvasculature [21] where they closely associate with P-selectin positive areas on postcapillary venule endothelia, confirming the reported involvement of both platelets [18], [20]–[22], [74]–[76] and P-selectin [77], [78] in the pathogenesis of HCM and ECM. Although platelets produce many factors including VEGF, ROS, and thrombin that are able to directly impair endothelial barrier function [185], the highly focal nature of platelet arrest [20], [21] is difficult to reconcile with the homogenous vascular leakage throughout the entire venous microvasculature at the time of ECM [21]. In agreement with their crucial role early in the pathogenesis of ECM, platelets are more likely to accomplish their vascular permeability-augmenting effect indirectly, namely by 1) facilitating leukocyte arrest in postcapillary venules [22], [57], [186], [187], 2) boosting ROS formation in leukocytes [185], and 3) enhancing cytokine secretion and cytotoxic capacity of effector T cells [188]. Thus, platelets adhering to small foci of endothelial P-selectin may serve as early nucleation points for the gradually spreading, cytokine-induced vasculopathy observed in postcapillary venules from ECM-susceptible mice.
In conclusion, we find that steric hindrance, mediated by large numbers of arrested CD8+ T cells, macrophages, and neutrophils in postcapillary venules, causes a severe restriction in the venous blood flow. Based on this observation, we hypothesize that completely different pathogenetic mechanisms - cytoadherence of P. falciparum iRBC in HCM and sequestration of leukocytes in ECM - can result in the same pathophysiological outcome: a severe reduction in the blood efflux from the brain. If the resulting increase in intracranial pressure intensifies the cerebral edema to the point of brainstem herniation, then compression of the respiration centers in pons and medulla could cause death by respiratory arrest.
Intracranial hypertension is a known risk factor for poor outcome in pediatric HCM [189], [190], but how this leads to neural injury is unknown [191], [192]. As discussed in a recent review [193], intracranial hypertension eliminates the need to explain any selective recognition mechanism Plasmodium might use to target multiple sensitive sites in the brain [194]. Intracranial hypertension leading to general swelling and hypoperfusion of the brain can also explain all of the neurological sequelae in HCM survivors [194]. Reports associating loss of smell, deafness, and blindness with both HCM and ECM support the notion that sequestration of iRBC as well as leukocytes can cause intracranial hypertension [192], [195]–[200]. Further, leukocyte sequestration is likely also involved in intracranial hypertension during P. falciparum HCM, as artesunate treatment was more efficacious in Asian adults compared to African children with more mononuclear cell accumulation [157], [201], [202] and also failed to rescue HCM patients with a low parasite biomass in the brain [203]. Thus, HCM and ECM induce very similar neurological symptoms and sequelae.
Thus, despite fundamental differences in parasite biology, the non-cytoadherent rodent parasite PbA could be used as a model to better understand how rheological alterations might lead to the annual death from HCM of over half a million people [204]. Perhaps more importantly, venous efflux disturbances due to leukocyte sequestration could also explain the cerebral complications that are increasingly reported for severe infections with P. vivax, another knobless (essentially non-cytoadherent) Plasmodium species [205]–[214]. The marked proinflammatory responses and reversible microvascular dysfunction associated with P. vivax infections [215], [216], together with the shared propensity for reticulocyte invasion [217]–[220], may render the PbA-infected mouse model suitable for study of the pathogenesis of severe P. vivax malaria. Hence, this model promises to shed light on the ultimate cause(s) of death from cerebral malaria.
This study was conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee, NYU School of Medicine (Protocol number 101201-01). All surgery was performed under ketamine-xylazine-acepromazine anesthesia, and all efforts were made to minimize suffering.
P. berghei and P. yoelii parasites were maintained by passage through female Anopheles stephensi mosquitoes [221]. The green fluorescent P. berghei ANKA strain (PbA-GFP) was a kind gift from Dr. Andy Waters, University of Glasgow, UK [222]. The lethal P. yoelii strain 17XL, originally derived from the non-lethal 17X strain [223], was kindly provided by Dr. James Burns (Drexel University College of Medicine) [224], [225]. WT PyXL were transfected to express RedStar, an improved version of the red fluorescent protein drFP583/DsRed/RFP [226], under the control of the elongation factor 1α promoter using a novel replacement strategy [21], [227], [228] and termed PyXL-RFP [21]. Both PbA-GFP and PyXL-RFP emit fluorescence throughout the entire life cycle.
Mice were CBA/CaJ (Jackson Laboratory, Bar Harbor, ME). Animals were maintained and used in accordance with recommendations in the guide for the Care and Use of Laboratory Animals. CBA/CaJ mice were infected at the age of 3 weeks (body weight of 12–15 g) by intraperitoneal injection of 0.5–10×106 iRBC as described [21]. In our hands, 3 week-old CBA/CaJ mice responded to PbA infection with neurological signs and died from ECM comparable to adult mice used by others [3], [6], [18], [22], [23], [76], [141], [169], [229]–[244]. The parasitemia was monitored daily using Giemsa stained blood smears and mice were sacrificed upon development of ECM or hyperparasitemia.
PbA-infected mice were considered ECM positive when two or more parameters clearly indicated behavioral alteration including body position, spontaneous activity, startle response, tremor, gait, touch escape, and righting reflex [245]. Quantitative assessment of ECM-associated neurological signs was performed using the Rapid Murine Coma and Behavior Scale (RMCBS) with values of 3–7 defined here as severe, 8–16 as mild, and 17–20 as no ECM [246]. PyXL-infected mice exhibited a hunched position, pale skin color, and an increased respiration rate when parasitemia levels reached levels around 80%, but failed to present typical neurological manifestations.
For IVM, groups of 15 PbA-infected CBA/CaJ mice received one daily oral dose of 0.3 mg/kg FTY720 starting one day before infection or no treatment as described [21], [39]. At the onset of neurological signs, mice were injected with Evans blue and examined by IVM. Mice surviving the critical period of ECM development were inoculated with Evans blue and PE-conjugated species anti-species CD8a and imaged on day 9. Groups of 5 PyXL-infected or uninfected mice were inoculated with the same markers and imaged on day 5 for comparison. For flow cytometry, PbA-infected mice, FTY720-treated PbA-infected mice, and PyXL-infected mice were analyzed on day 6–8, day 8, and day 5, respectively.
Mice were anesthetized by intraperitoneal injection of a cocktail of 50 mg/kg ketamine (Ketaset, Fort Dodge Animal Health, Fort Dodge, IO), 10 mg/kg xylazine (Rompun, Bayer, Shawnee Mission, KS), and 1.7 mg/kg acepromazine (Boehringer Ingelheim Vetmedica, St. Joseph, MO) (KXA mix) and surgically prepared for intravital imaging of the brain as described [21], [247], [248]. CBA/CaJ mice were infected with PbA and subjected to brain IVM 1) on day 5 prior to the appearance of neurological signs, 2) on day 6–8 upon ECM development, or 3) on day 9 after the window of ECM development had passed. PyXL-infected mice were imaged upon the parasitemia exceeding 50% on day 5. Uninfected mice were used as controls. Prior to imaging, mice were inoculated with Evans blue and matching combinations of fluorescent markers. Despite severe illness of the animals, optimization of anesthesia, craniotomy, and injection of fluorescent markers allowed us to obtain good recordings from approximately 70% of the mice with ECM and 50% of the mice with hyperparasitemia.
iRBC were identified by fluorescent protein expression in the parasites or reflection of hemozoin [21]. The vascular lumen was visualized by intravenous injection of 100 µl of a 1% solution of Evans blue. Vascular endothelia were labeled intravenously with Alexa 488 or eFluor 450-conjugated rat anti-mouse PECAM-1 (CD31; clone MEC13.3, BioLegend, San Diego, CA) and phycoerythrin (PE)- or Alexa 647-conjugated rat anti-mouse CD14 (clone Sa2-8, eBioscience). CD8+ T cells, CD4+ T cells, monocytes, macrophages, neutrophils, or platelets were labeled by intravenous injection of 3–5 µg of the following fluorochrome-conjugated monoclonal antibodies using appropriate color-matching combinations: PE-conjugated rat anti-mouse CD8a (clone 53-6.7; eBioscience, San Diego, CA), eFluor 450 or PE-conjugated rat anti-mouse CD4+ (clone GK 1.5, eBioscience), Pacific blue-conjugated rat anti-mouse CD11b (clone M1/70, BioLegend), eFluor 450-, PE- or Alexa 647-conjugated rat anti-mouse GR-1 (Ly-6G/6c; clone RB6-8C5, eBioscience and BioLegend), and eFluor 450-conjugated rat anti-mouse CD41 (clone MWReg30, eBioscience), respectively. ICAM-1 expression was visualized with intravenously inoculated PE-conjugated rat anti-mouse CD54 (clone YN 1/1.7.4, Biolegend). P-selectin was detected with PE-conjugated rat anti-human/mouse CD62p (KO2.3, eBioscience).
Multiple time sequences and 3D stacks were recorded for quantification of the number of arrested leukocytes in the vascular lumen. Depending on the experimental conditions, 20–45 postcapillary venules and arterioles were analyzed per mouse. Arrested leukocytes were defined in each vessel segment as cells that did not detach from the endothelial lining within the observation period. CD8+ T cells, CD4+ T cells, neutrophils, monocytes, and macrophages were quantified by counting the number of cells per square millimeter of vessel surface [249]. The relative density of the various leukocyte populations was determined in multiple fields of view per experimental condition and expressed as the mean ± SEM of arrested cells as well as the percentage of the total cell number. Velocities were measured with Imaris Track as described [250].
To quantify endothelial ICAM-1 expression, confocal 3D stacks of postcapillary venules or similarly sized arterioles were acquired from 2 mice each infected with PbA, PyXL, or no parasites. The fluorescence signal intensity across 10 (PbA) or 12 (PyXL and control) vessel volumes was collected from 3D data and quantified in ImageJ. To quantify leukocyte ICAM-1 expression, the relative fluorescence emission from individual leukocytes was measured from mice infected with PbA or PyXL. As for endothelial ICAM-1 expression, confocal 3D stacks were collected and projections were created in AutoDeblur. A total of 6 stacks from 2 mice per experimental condition were analyzed.
Leukocytes were prepared from the brain (cerebrum and cerebellum) using established procedures [251]–[253]. Briefly, mice were perfused intracardially with Mg2+ and Ca2+-free PBS to dislodge nonadherent leukocytes. Next, the brain was removed and gently minced through a mesh strainer (mesh size: 100 µm; Fisher Scientific) using a syringe plunger. The homogenate was suspended in 10 ml HBSS containing 0.05% collagenase D (Roche Diagnostics, Indianapolis, IN), 0.1 µg/ml of the trypsin inhibitor TLCK (Sigma), 10 µg/ml DNase I (Sigma), and 10 mM Hepes buffer, pH 7.4. The tissue slurry was gently rocked at RT for 60 min and then allowed to settle at 1 g for 30 min. The supernatant was collected and 5 ml of the suspension was layered onto 10 ml of a density separation medium containing 7.5 ml of RPMI medium containing 10% FBS, 10 mM HEPES, and 2.5 ml Ficoll Paque (GE HealthCare) in a 50 ml conical centrifuge tube and centrifuged at 400 g for 30 min. The overlying media and tissue debris were removed, the entire gradient medium was diluted ten-fold with HBSS and centrifuged at 300 g for 10 min. Isolated leukocytes were washed twice with Mg2+ and Ca2+-free PBS before phenotyping. A total of 10 PbA-infected mice with ECM, 6 PbA-infected/FTY720-treated, and 6 PyXL-infected mice with hyperparasitemia were subjected to flow cytometric analysis. Uninfected control mice were not included in the analysis, because the cerebral microvasculature of these animals does not exhibit arrested leukocytes.
The following antibodies were used for leukocyte phenotyping: Total leukocytes were detected with PE-Cy7-conjugated rat anti-mouse CD45 (clone 30-F11; eBioscience, San Diego, CA), T cells with APC-Cy7-conjugated Armenian hamster anti-mouse CD3 (clone 145-2C11; BD Biosciences, San Jose, CA), CD8+ T cells with eFluor 450 or Alexa 700-conjugated rat anti-mouse CD8a (clone 53-6.7; eBioscience or Biolegend, San Diego, CA), CD4+ T cells with PE- Texas Red-conjugated rat anti-mouse CD4+ (clone RM4-5; Life Technologies, Carlsbad, CA, neutrophils with FITC-conjugated rat anti-mouse Ly6G (clone 1A8; Biolegend), monocytes with Alexa 700-conjugated rat anti-mouse Ly6C (clone HK 1.4; Biolegend), and macrophages with anti-mouse CD11b (clone M1/70; eBioscience) or PE- or eFluor 450-conjugated anti-mouse F4/80 (clone BM8; eBioscience). CCR5 (CD195) was detected with PE-conjugated rat anti-mouse CD195 (clone HM-CCR5; Biolegend), CD69 with Pacific Blue-conjugated rat anti-mouse CD69 (clone H1.2F3; Biolegend), ICAM-1 with APC-conjugated rat anti-mouse CD54 (clone YN1/1.7.4; eBioscience), and granzyme B with FITC-conjugated anti-mouse granzyme B (clone GB11; Biolegend). Data were acquired with a 5-laser, 17-color LSR II Analytic Flow Cytometer (BD Biosciences, San Jose, CA) and analyzed with FlowJo software (Treestar, Ashland, OR).
Time series showing the blood flow in postcapillary venules or similarly sized arterioles were acquired from mice infected with PbA, PyXL, or no parasites and converted to minimal projections to visualize the portion of the vascular lumen used for blood flow. The vascular lumen was visualized by IV inoculation of Evans blue [21]. IVM movies were converted to minimal projections to visualize the perfused center of the vessel. Multiple measurements were taken for each vessel to determine the entire vascular diameter (distance between endothelia) versus the perfused part of the vessel (dark center) (Figures 1 and S2). Vascular restriction is expressed as percent reduction in vessel diameter or cross-section.
The cortical microvasculature was imaged with an inverted Leica TCS SP2 AOBS confocal system as described [21]. Time series and 3D data sets were acquired with Leica Confocal Software [248], [250], [254]. Imaris 7.4 (Bitplane, Saint Paul, MN), Image-Pro Plus (Media Cybernetics, Bethesda, MD), AutoDeBlur (Media Cybernetics, Bethesda, MD), and NIH ImageJ were used for further image analysis, deconvolution, and 3D reconstruction.
Blood was removed by perfusion with PBS via the left ventricle [251]. Brains were snap-frozen for preparation of cryostat sections. To determine the expression level of tight junction proteins, sections were fixed in 95% ice-cold ethanol for 30 minutes and then permeabilized in acetone for 1 min at RT. After blocking in 50% goat serum for 45 min, sections were labeled with rabbit polyclonal antibody against claudin 5 (34-1600), rabbit polyclonal anti-occludin (40-4700), mouse monoclonal anti-ZO1 (33-9100), all from Invitrogen. Rat monoclonal antibody anti-CD31 (553370) was from BD Biosciences. Sections were then incubated for 2 h at RT with secondary Alexa antibodies (Life Technologies) and mounted in Vectashield with DAPI (H1200, Vector Labs, Burlingame, CA). Depending on the experimental condition, 9–16 images from various regions of the cerebrum and cerebellum were analyzed. Data were acquired at the same magnification with Zeiss Axiovison software using an Axio Imager-D2 Zeiss fluorescent microscope and imported into Image J for quantification of vascular leakage and tight junction protein expression. For tight junction protein expression, the fluorescence threshold was set with the “triangle” setting of Image J to limit the measurement of the brightest signal only, corresponding to the tight junctions. Ratios were then calculated by comparing the experimental groups with the control group using Excel. Significance was determined by t-test.
Statistical analysis was performed using Minitab® 17 (Minitab, State College, PA). Data sets were tested for normality and equal variances and when necessary data was Log10 transformed to obtain a normal distribution. Data were analyzed either by t-test, one-way ANOVA, or General Linear Model as appropriate. Where transformation was not successful, the non-parametric Mann-Whitney U test was used.
|
10.1371/journal.pcbi.1002052 | The Binding of Learning to Action in Motor Adaptation | In motor tasks, errors between planned and actual movements generally result in adaptive changes which reduce the occurrence of similar errors in the future. It has commonly been assumed that the motor adaptation arising from an error occurring on a particular movement is specifically associated with the motion that was planned. Here we show that this is not the case. Instead, we demonstrate the binding of the adaptation arising from an error on a particular trial to the motion experienced on that same trial. The formation of this association means that future movements planned to resemble the motion experienced on a given trial benefit maximally from the adaptation arising from it. This reflects the idea that actual rather than planned motions are assigned ‘credit’ for motor errors because, in a computational sense, the maximal adaptive response would be associated with the condition credited with the error. We studied this process by examining the patterns of generalization associated with motor adaptation to novel dynamic environments during reaching arm movements in humans. We found that these patterns consistently matched those predicted by adaptation associated with the actual rather than the planned motion, with maximal generalization observed where actual motions were clustered. We followed up these findings by showing that a novel training procedure designed to leverage this newfound understanding of the binding of learning to action, can improve adaptation rates by greater than 50%. Our results provide a mechanistic framework for understanding the effects of partial assistance and error augmentation during neurologic rehabilitation, and they suggest ways to optimize their use.
| Einstein once said: “Insanity is doing the same thing over and over again and expecting different results”. However, task repetition is generally the default procedure for training a motor skill. This can work because motor learning ensures that repetition of the same motor task will lead to actions that are different, as errors are reduced and motor skill improves. However, here we show that task repetition, although not “insane”, is inefficient. The machine learning algorithms used to control motion in robotics adapt the movement that was actually made rather than the planned movement in order to assure stable learning. In contrast, it had been widely assumed that neural motor systems adapt based on the planned rather than the actual movement. If this were the case, task repetition would be an efficient training procedure. Here we studied the mechanisms for motor adaptation in humans and found that, like in robotic learning, the adaptation that we experience is associated with the actual movement. This finding led to the design of an improved training procedure that avoids task repetition. Instead, this procedure continually adjusts the movement goal in order to drive participants to experience the correct movement, even if initially by accident, leading to an over 50% improvement in the motor adaptation rate.
| When learning to swim, the proper stroke motion is usually taught on the pool deck. Although a student might seem to have mastered this motion on dry land, upon entering the water she will have difficulty in accurately reproducing it underwater. However, after many laps, the student eventually learns to produce the pattern of motor output that leads to the proper stroke motion while swimming. This learning occurs via the formation of internal models of the physical dynamics experienced which allow the programming of movement to contend with the dynamics of the environment [1]–[4]. These internal models have been shown to predict the dynamics of the environment as a function of motion rather than as a function of time [5]–[8] – a strategy that makes sense in light of the viscoelastic and inertial physics of our own limbs and the objects we interact with. Consequently, the neural plasticity which underlies this learning must establish associations between motion state (i.e., position and velocity vectors) and motor output which can counteract environmental forces. Although the existence of these associations has been well established, the mechanism by which they form is not yet understood.
How does this state-dependent learning arise during the course of motor adaptation? One possibility is that on individual trials, an internal model of the environment is updated based on a combination of the errors experienced and the motion plans that led to those errors. Another possibility is that internal models are updated based on errors experienced in combination with the actual motion states associated with those errors. It is remarkable that previous work on motor learning in neural systems has widely assumed the former [4], [9]–[16], despite the fact that direct evidence for this hypothesis is scant. The idea that learning is associated with the motion that was planned (plan-referenced learning) is especially pervasive in the learning rules of the algorithms that have been proposed to model the process of adaptation in the neuromotor learning literature [4], [9], [11]–[12], [15], however it is difficult to find work that addresses the validity of this assumption, explores its implications or provides a clear rationale for its use.
The machine learning community has developed, in parallel, a series of algorithms for updating internal models in robotic systems. Interestingly, these algorithms almost uniformly involve learning rules in which internal models are updated based on a combination of the errors experienced and the actual motion associated with those errors (motion-referenced learning) rather than the motions that were planned [17]–[21]. The choice of these learning rules is grounded in the idea that adaptive changes should be provably stable in the sense that, under a set of reasonable assumptions, updated internal models should never result in worse performance [17]–[21]. Here we ask the question: Do the associations between motor output and motion state formed during human motor learning arise from adaptation based on planned or actual motions? The answer to this question is important not only for theories of motor control, and issues of stability during learning, but also because knowledge of how associations are formed during motor learning can be leveraged to improve the efficiency of training procedures.
Motor adaptation can be described as the process of tuning motor output to reduce the errors between plan and action. Thus the associations between motion state and motor output formed during this process result from the way that responsibility for these errors is assigned. This is known as a credit assignment problem. This problem can be posited as the task of assigning blame after an error is experienced to the set of actions that would be most likely to give rise to similar errors in the future. This set of actions could then be modified in order to improve performance in subsequent trials. Viewed in this way, the distinction between plan-referenced learning (PRL) and motion-referenced learning (MRL) corresponds to whether the blame for motor errors should be assigned to the planned versus actual motion. Consequently, the amount of adaptation on a given trial will be determined by the magnitude of the error, however the location of the adaptation (which future motions will benefit from the adaptation) will be determined by the credit assignment mechanism. Here we studied the generalization of motor adaptation to untrained conditions in order to elucidate the credit assignment mechanism used by the CNS, and then used our understanding of this mechanism to design a training paradigm that takes advantage of it to improve the efficiency of motor adaptation.
The adaptations that would occur at different stages of training for reaching arm movements in a velocity-dependent force-field (FF) for the PRL and MRL credit assignment hypotheses are shown in Figure 1. The green shaded region around the planned motion – which is essentially straight toward the target for short (10 cm) movements [22] – represents the space of future motions which would benefit from the adaptation to the greatest degree under PRL (Figure 1A). Alternatively, each red shaded region represents the space of future motions which would benefit maximally under MRL. A more direct visualization of the adaptive changes predicted by each credit assignment hypothesis can be made by representing motion and the resulting adaptation in velocity-space rather than position-space, since the adaptation to the velocity-dependent dynamics studied in the current series of experiments is believed to be mediated by an internal model largely composed of velocity-dependent motor primitives [8], [10], [12]–[13], [23]. These primitives are the learning elements which contribute to the compensatory motor output (i.e., compensatory force) in a velocity-dependent manner. Figure 1B shows how individual motor primitives would adapt based on PRL versus MRL credit assignment early on in training. Here each circle represents a single motor primitive (centered at its preferred velocity) with a color intensity denoting the amount of adaptation that would arise from the illustrated trial. The left and right panels of Figure 1B show the adaptations predicted by PRL (green) and MRL (red), respectively. As in Figure 1A, adaptation is centered on the planned motion for PRL and centered on the actual motion for MRL.
As training proceeds over the course of several trials, the activation levels of the adapted primitives would continue to increase. This continued increase in activation (not illustrated) leads to increased compensatory force, resulting in greater compensation of the external dynamics and thus straighter trajectories. Note that the adapted primitives would be noticeably different for the two credit assignment hypotheses early in training, but would overlap late in training as force compensation increases and the planned and actual motions converge as illustrated in Figure 1A.
Given the different implications that the PRL and MRL credit assignment mechanisms have for motor adaptation, we can assess which one is favored by the CNS by asking a simple question: After training, which motions gain the most benefit from the induced adaptation? The motions that were planned or the motions that were experienced? Since the mechanism for credit assignment determines which motions will benefit from adaptation on a particular trial, we studied how motor adaptation to a single target direction generalizes to neighboring motion directions. If a particular motion is trained, the pattern of generalization can be viewed as a record of the history of credit-assignment for the errors experienced during a training period. Specifically, the amount of generalization in the directions neighboring the trained movement constitutes the set of actions that the motor system believes should be adapted based on the history of errors experienced. Therefore, PRL and MRL should give rise to different patterns of generalization.
In order to cleanly distinguish between these hypotheses, we designed an experiment in which the planned motion and the actual motion were maintained to be distinct from one another during the entire dataset so that the patterns of generalization predicted by PRL vs. MRL would be very different from one another. This is a challenge because, training a motor adaptation generally results in improved performance such that the actual motion converges onto the planned motion, and such a scenario could hamper the ability to clearly distinguish between the PRL and MRL hypotheses. Thus, we designed an experiment in which actual motion would not converge onto planned motion during the course of training, resulting in enduring differences between the predictions of these two hypotheses. To accomplish this, subjects were exposed to a training period consisting of short, successive blocks of movements towards a single target location with a force-field (FF) that alternated between clockwise (CW) and counterclockwise (CCW) directions from block to block (see Figure 2A). The magnitudes of the CW and CCW FFs were, respectively, 9 and −9 N/(m/s). In these FFs, the peak force perturbations were 2.7 and −2.7 N, respectively, for an average movement with a peak speed of 0.3 m/s. The FF blocks were short enough (7±2 trials) that neither the CW nor the CCW FF could be learned very well before unlearning with the opposite FF occurred. After subjects were exposed to a number of these interfering FF cycles, we measured the generalization of adaptation to untrained movement directions with error-clamp (EC) trials (see Materials and Methods for details).
The predictions of PRL and MRL are strikingly different for this experiment. For the PRL hypothesis, since the adaptation is associated with motor primitives centered at the same target direction for both FFs (Figure 2B top panel, blue and orange traces), the balanced exposure to these opposite FFs would lead to cancellation of the CW and CCW FF learning resulting in near zero adaptation at the trained target direction and the adjacent directions (Figure 2B, dashed green trace). Note that although target locations are identical between CW and CCW FF trials, the actual movement directions differ. The CW FF perturbs motion towards smaller movement angles whereas the CCW FF does the opposite. Therefore, MRL predicts that smaller movement angles would be preferentially associated with adaptation appropriate for the CW FF (blue trace in the bottom panel of Figure 2B), whereas higher movement angles would be preferentially associated with adaptation appropriate for the CCW FF (orange trace in the bottom panel of Figure 2B). This would lead to the bimodal pattern of generalization illustrated in Figure 2B (red dashed trace).
We trained one group of subjects in this FF interference paradigm at a target location of 270°. We found that target directions smaller than the training direction consistently display generalization appropriate for the CW FF (negative) whereas target directions greater than the training direction display generalization appropriate for the CCW FF (positive). This is consistent with the bimodal generalization pattern predicted by MRL (compare the blue and red traces in Figure 2C: r = 0.92, F(1,7) = 36.87, p<0.001) and quite different from the essentially flat pattern predicted by PRL (green trace). Correspondingly, we found the adaptation levels at the target directions corresponding to the peaks of the predicted generalization pattern (−30° and +30°, see Figure 2C) to be significantly different from one another (t11 = 7.26, p<9×10−6) and from zero (t11 = 5.95, p<5×10−5 for −30°, and t11 = 3.89, p<0.002 for +30°). These results provide direct evidence for MRL by matching the complex pattern of generalization predicted by it.
In our experiment we balanced the direction of the FF that was presented before testing generalization, nevertheless, we noticed a small bias in the generalization function at the training direction consistent with a bias in adaptation level that we observed during the training period (see Figure S1 and Text S1). This bias is compatible with other results showing somewhat faster learning for a CW FF [8]. In order to eliminate the possibility that this bias or the target location we chose for training (270°) might have somehow contributed to the generalization pattern we observed in the data, we trained a second group of subjects in a version of this experiment that was designed to eliminate the bias and provide training at another target location (60°). We eliminated the bias by unbalancing the number of CW versus CCW FF trials in each cycle in this second group of subjects (see Text S1). We found that the close match between the pattern of generalization that these subjects displayed (Figure 2C, grey trace) and the pattern predicted by MRL persisted under these conditions (r = 0.93, F(1,7) = 42.61, p<0.001). Correspondingly, the adaptation levels at −30° and +30° were significantly different from each other (t9 = 5.37, p<3×10−4), and significantly different from zero (t9 = 3.72, p<0.003 for −30°, and t9 = 4.38, p<9×10−4 for +30°). Together, these results provide compelling evidence for MRL as the mechanism for credit assignment in motor adaptation.
We note that Equations 3 and 4 used for our simulations incorporate local motor primitives that are functions of the initial movement direction (θ) rather than of the full time series of the velocity vectors encountered during each trial. This might seem an inappropriate choice since, as we discussed above, velocity-dependent motor primitives are thought to underlie the learning of velocity-dependent dynamics [8], [10], [12]–[13], [23]. However this approximation is a good one when movements are approximately straight, which is essentially the case for the first 400 ms of the movements considered in our study. This approximation, of course, breaks down at the end of the movement when the initial movement direction no longer describes the velocities experienced. However, the amplitudes of the velocity vectors during the end-movement correction are quite low and so the unmodeled spread of learning to the actual motion experienced in this correction phase should have relatively little effect since at low velocities, viscous dynamics have small consequences. This effect can be visualized in the left panel of Figure 1B which shows that the end-movement correction which has a velocity vector that points to the second quadrant would only excite velocity-dependent primitives near the origin under MRL.
Note that the separation of the peaks in the bimodal generalization pattern predicted by MRL (red dashed line in Figure 2C) results from the size of the errors experienced during training. Consequently, larger force-field perturbations which induce larger errors would result in greater separation between the peaks. However, the separation between the peaks (about 60°) is predicted to be greater than the separation between the average errors experienced in the two force-fields (about 25°). There are two reasons for this. The first is that more adaptation occurs on trials with larger errors than those with smaller errors, skewing the center of adaptation for each force-field outwardly from the mean experienced error. The second reason is illustrated in the lower panel of Figure 2B: When the patterns of generalization for the positive and negative force-fields are summed, resulting in a bimodal generalization pattern for MRL, the peaks of this bimodal generalization pattern (red) are separated by an even greater distance than the peaks of the positive (orange) and negative (blue) components because the amount of cancellation between these components is greater at movement directions corresponding to smaller rather than larger errors resulting in further outward skew.
Previous work has attempted to measure the generalization functions (GFs) associated with learning a single FF. MRL predicts that these GFs will be shifted toward the motion directions experienced during training. Many of these studies have estimated GFs from complex datasets using a system identification framework [10], [12]–[13]. However the implementation of this framework assumed PRL in these studies, thus preventing a straightforward interpretation of their results. In one study [24] a simpler generalization experiment was conducted, in which subjects were trained with a single FF to a single target location, after which the resulting GF was measured. Because the actual motions approached the planned motions late in training, the shifts predicted by MRL would be subtle. Furthermore, the ability to detect shifts in the generalization function was hampered by a coarse sampling of the generalization function (45°). Nevertheless, careful inspection of these GFs consistently reveals subtle shifts towards the motions experienced during training as predicted by MRL. However, it is difficult to be certain whether if the shifts observed in this study result from MRL rather than innate biases in generalization functions because only a single FF direction was studied. Innate biases might stem from biomechanical asymmetries or direction-related biases in adaptation. We therefore performed a pair of single-target, single-FF experiments in order to compare the shifts in generalization associated with opposite FFs. The results of these experiments confirm the existence of subtle but significant shifts in generalization [25]. The magnitudes and the directions of these shifts are consistent with the MRL hypothesis [25].
Insights into the mechanisms for learning in the CNS can provide a platform for creating training procedures that leverage these insights to improve the rate of learning – an important goal for both motor skill training and neurologic rehabilitation. With our new understanding of how the CNS solves the credit assignment problem, we looked into the possibility of designing a novel training paradigm to take advantage of this knowledge. A key consequence of plan-referenced learning is that this mechanism for credit assignment would result in a match between what is learned and what is commanded on the next trial if the same motion plan is repeated from one trial to the next during training – like when aiming a dart at the bull's eye repeatedly. In contrast, motion-referenced learning would result in a mismatch. Motion-referenced learning, therefore, predicts that the process of training an accurate movement to a given target location in a novel dynamic environment would be inefficient if that target were repeatedly presented at the same location during training (single-target training, STT) as illustrated in Figure 3. This inefficiency arises because the motion experienced during training does not coincide with the motion that is to be learned, resulting in limited overlap between the motion-referenced learning that occurs and the learning that is desired.
The aforementioned inefficiency can be ameliorated by a paradigm which continually changes the locations of the targets presented during the training period as shown in Figure 3, second column. In this training paradigm, target directions would be shifted from one trial to the next so that the actual motion experienced repeatedly lines up with the motion to be learned. For the CW FF depicted in Figure 3, this corresponds to left-shifted training (LST). Initial target locations are placed with large leftward shifts with respect to the desired learning direction – in anticipation of the large rightward initial errors with respect to the target location. These leftward target shifts are then gradually reduced as learning proceeds and errors become smaller, in order to maintain alignment between the actual motion experienced and the movement to be learned.
The MRL hypothesis predicts that the LST training paradigm should produce faster learning than the standard STT paradigm used in previous motor adaptation studies in which a single target direction was trained [24], [26]. We tested this idea by comparing the learning curves associated with these training paradigms for adaptation to a clockwise viscous curl force-field. A different group of subjects was studied on each paradigm to avoid the effects of savings [27]–[29]. As a control for a possible increase in attention associated with changing target locations in the LST paradigm, we tested a third group of subjects with a right-shifted training (RST) paradigm. Here targets were shifted to the right, mirroring the target positions in the LST paradigm. The MRL hypothesis would predict slower learning for RST than STT or LST because right-shifted targets in a rightward pushing force-field would result in reaching movements even farther away from the desired learning direction than those expected in STT (see Figure 3, third column). In contrast the PRL hypothesis would predict fastest learning for the STT paradigm and identical learning rates for the LST and RST paradigms because the STT paradigm creates perfect alignment between the desired learning and the planned motion whereas the LST and RST paradigms create misalignments between the desired learning direction and planned motion that are opposite in direction but equal in magnitude. We used a FF magnitude of 22.5 N/(m/s) for these experiments – 2.5 times the magnitude used in Experiment 1 – in order to magnify the various misalignments discussed above. In all three paradigms, we measured learning at the desired learning direction (90°) by pseudo-randomly interspersing 90° error-clamp trials among the training trials with an average frequency of 20%.
We first collected data from a subset of subjects in the STT paradigm in order to estimate the evolution of directional errors across trials. We used this pattern of directional errors to determine the target shifts that would produce good alignment between experienced motion and desired learning direction for the LST paradigm (see Materials and Methods). As shown in Figure 4A, we obtained a good match between motion direction and the desired learning direction (90°) throughout the training period for the LST paradigm, so that misalignment between these directions was dramatically reduced compared to the STT paradigm. Correspondingly, the misalignment between motion direction and the desired learning direction was about twice as great for RST than for STT.
The plots shown in Figure 4B illustrate how the adaptation patterns predicted by MRL and PRL would evolve as training proceeds for the training paradigms discussed above. Note that adaptation spreads across a limited range of movement directions consistent with local generalization [24]–[26], but the alignment between adaptation and the desired learning direction (90°) varies from one paradigm to another (STT vs. LST vs. RST), and from one credit assignment hypothesis to another (PRL vs. MRL). The darkened dots which highlight a slice through these plots at 90° illustrate the amount of adaptation associated with the desired learning direction.
These simulations show that the PRL hypothesis predicts that in the STT paradigm, credit assignment will be perfectly aligned with the desired learning direction (90°) throughout training. PRL also predicts an equal but opposite pattern of misalignments between credit assignment and desired learning for the LST and RST paradigms (Figure 4B). These misalignments are initially large but become attenuated during the course of the training because planned and actual motions converge. This results in simulated learning rates that are highest for the STT paradigm and lower, but identical, for the LST and RST paradigms under PRL (Figure 4B–C). In contrast, the simulations for the MRL hypothesis show perfect alignment between the credit assignment and the desired learning direction for the LST paradigm. For STT, the MRL-based simulations show a gross misalignment between the credit assignment and the training direction. For RST, the misalignment is even greater (Figure 4B). This results in learning rates that are predicted to be greatest for the LST paradigm, followed by the STT and RST paradigms, respectively (Figure 4B, D). As with the PRL simulations, the misalignments become attenuated as training proceeds.
Our experimental data show a clear difference between the learning curves obtained for the three training paradigms in the early stages of training (first three EC trials; one-way ANOVA, F(2,87) = 14.57 , p<4×10−6). The LST group displays the highest adaptation levels and the RST group displays the lowest adaptation levels as shown in Figure 4E. In particular, the LST group displayed an 86% increase in adaptation levels on the first EC trial and a 52% increase over the first three EC trials, whereas the RST group displayed a 59% decrease in adaptation levels compared to STT over the first three EC trials in the training period. Post-hoc comparisons between groups over the first three EC trials indicate that the LST group showed significantly greater learning than the RST group (t58 = −5.05, p<3×10−6). This result is in keeping with the MRL prediction, but defies the PRL prediction of equal learning rates for these groups. Our data also shows that the LST group displays significantly greater learning than the STT group (t58 = −2.17, p<0.02), in keeping with the MRL prediction, but opposing the PRL prediction of a greater learning rate for STT. We also find that the STT group displays significantly greater learning than the RST group (t58 = −3.90, p<2×10−4), corroborating the group order predicted by the MRL hypothesis. These findings provide additional support for motion-referenced learning and demonstrate that a training paradigm that is designed to leverage knowledge about the mechanism for credit assignment can improve learning rates compared to standard training procedures.
Inspection of the learning curve for the RST group reveals that the adaptation for the first EC trial after exposure to the FF actually dips a bit below zero. MRL predicts reduced learning for this group but would not predict opposite learning, consistent with the finding that the adaption level at this point, although nominally less than zero, is not significantly so (t27 = −2.01, p>0.05). Additionally, we note that the third-to-last error-clamp trial in the baseline (which is illustrated along with the full learning curve in Figure S2) displays an adaptation coefficient which dips below the average baseline and falls within the error bars of the first point in the RST learning curve, suggesting that the latter is not entirely outside the range of the data. Despite the differences in learning rate predicted by MRL-based credit assignment, angular errors should decrease as the training period proceeds. This results in reduced misalignment between prescribed and actual motion directions for the STT and RST groups, leading to a predicted convergence of the adaptation levels for all three groups (see Figure S2 and Text S1). Our data bears out this prediction: despite significant differences between groups early in the training period, we find no significant difference between groups late in the training period (last three EC trials; one-way ANOVA, F(2,87) = 0.23, p>0.05). In addition, although we have shown that the MRL-based training paradigm (LST) increases the rate of adaptation, our results do not provide any information on the long-term retention for this adaptation. Further studies would be required to assess if the retention of the motor memories acquired using an MRL-based training paradigm is greater than that of memories acquired using single-target training paradigms.
Elucidating how associations are modified during the process of learning is a key step towards understanding the mechanisms underlying behavioral plasticity. Our findings demonstrate that the effect of the adaptation arising from an error sensed on a previous movement is greatest when the plan for the current motion matches the motion experienced on the previous trial. This indicates that, in the motor adaptation task we studied, the learned association binds the adaptive change in motor output to the actual motion experienced. We first showed that this motion-referenced learning hypothesis is able to explain the complex pattern of generalization that emerges when subjects are exposed to multiple blocks of interfering force-fields. We then followed up this result by showing that a manipulation of the pattern of target locations that aligned the actual motion experienced during training resulted in significantly improved learning rates, whereas a manipulation which increased misalignment resulted in significantly reduced learning rates. Together, these findings provide compelling evidence that credit assignment during motor adaptation is referenced to actual motions experienced rather than planned motions, and that this knowledge can be leveraged to improve the efficiency of motor skill training. The most general view of credit assignment would be that error-dependent motor adaptation might be composed of both motion-referenced and plan-referenced components. Although previous work overwhelmingly assumed pure plan-referenced learning [4], [9]–[16], our results indicate that motor adaptation is primarily composed of motion-referenced learning - in fact, our results are consistent with motor adaptation being fully motion-referenced. However, we cannot rule out a small contribution from plan-referenced learning. Consequently, further work will be needed to more precisely determine the relative contributions of each mechanism and to determine whether situations exist in which the levels of plan-referenced learning are substantial.
Despite the lack of direct evidence in support of it, plan-referenced learning has been widely assumed in the motor adaptation literature, particularly in modeling work in which a credit assignment scheme must be chosen, even if implicitly so, in order for a learning rule to be defined [4], [9]–[16]. Interestingly, Wolpert and Kawato (1998) assumed a hybrid credit assignment scheme: PRL for inverse-model learning and MRL for forward-model learning [4]. In principle, PRL is attractive because adaptation referenced to the previously planned motion would have the greatest effect on the same movement if it were repeated. In fact, Donchin et al. (2003), which models motor adaptation with a PRL learning rule, contains what the authors maintain is a proof that PRL-based learning is optimal in their supplementary materials [10]. However inspection of this proof reveals that its derivation is based on the assumption that motor adaptation acts to maximize the benefit that would be accrued if the same movement were repeated. In other words, this proof investigated what the optimal credit assignment procedure should be for STT and found that PRL maximizes the benefit of motor adaptation for STT. Since PRL is optimal for STT, MRL must be suboptimal for STT (as our simulations predict; see Figure 4B–D). This suggests that some training procedure other than STT would be optimal for MRL, and our data show that, for a clockwise FF, left-shifted training (LST) is indeed more effective than STT. Effectively, Donchin et al. (2003) assumed that a credit assignment procedure optimized for performance on STT would be used by the nervous system. Here we show that this is not the case. Instead, the error-dependent learning that occurs on a particular trial is referenced to the actual motion experienced on that trial rather than the planned motion, and as a result, STT produces slower learning than another training procedure (LST). Thus the human motor system does not adapt with the mechanism that would have the greatest effect on the same movement if it were repeated. Why would this be?
The problem with PRL is that the dynamics experienced are generally functions of actual rather than planned motion. For example, the dynamics experienced from moving a small dense mass would be proportional to the actual rather than the planned acceleration of that mass. Note that the dynamics that subjects experienced in our experiments were also dependent on the actual motion state, i.e., the force was based on the velocity of the actual rather than the planned motion. The key consequence of this state dependence is that since the force pattern experienced during a particular motion does not reflect the planned motion (because it reflects the actual motion), the force pattern that would have been experienced if the planned motion were achieved is unknown. This means that, in principle, the error between the current motor output and the environmental dynamics acting on the planned motion adaptation is also unknown. Because this error is unknown, no learning rule for adaptation referenced to the planned motion can be guaranteed to reduce it. If, however, errors are small enough so that the dynamics experienced in actual and desired trajectories would be very similar to each other, plan-referenced learning schemes could converge because these schemes essentially assume equality between these dynamics. On the other hand, if errors are sufficiently large, using such a credit assignment scheme might result in unstable learning which does not converge on the desired motor output. Clearly, a credit assignment scheme that could lead to instability would be a liability for the CNS.
The state dependence of physical dynamics insures that the force pattern experienced corresponds to the actual motion. Thus the error between the motor system's current estimate of the dynamics associated with the actual motion and the environmental dynamics associated with this motion can be determined. Because the motor output error corresponding to the actual motion can be determined, the motor output associated with it can be modified to reduce this error reliably, allowing for stable convergence of the motor output on the true environmental dynamics. This corresponds to motion-referenced learning. Interestingly, this reasoning is reflected in learning rules with mathematically provable stability that are widely used for the estimation of environmental dynamics in robotics and machine learning [17]–[21], [30]. These learning rules must be motion-referenced in order for stability to be assured.
One unfortunate consequence of motion-referenced learning is the suboptimal rate of motor adaptation observed if an individual were to repeatedly invoke the same motor plan when attempting to learn a novel task [30]. We demonstrate this suboptimality in the single-target training (STT) paradigm in Experiment 2 (Figure 4). Since adaptation proceeds according to the actual motion (rather than the planned motion), the STT paradigm leads to adaptation that is not aligned with the desired learning direction so that adaptation proceeds at a slower rate than if the actual motion is aligned across trials as in the LST paradigm. Our finding of motion-referenced credit assignment during motor adaptation is, therefore, compatible with the idea that the CNS favors a stable learning algorithm (MRL) over one that maximizes the effect of learning if the same motion plan is repeated at the expense of stability (PRL).
Recent studies have provided evidence for reduced learning rates for large errors [31]–[33]. One of these studies proposed the rationale that this occurs because the motor system sees large errors as less relevant than small errors [31]. However, note that in these studies the adaptation was measured not along the motion direction experienced during the training trials, but along the direction of the previously planned movement – equivalent to STT. Therefore the decreased learning rates associated with large errors observed in these studies may be, at least in part, due to misalignment in motion-referenced credit assignment, because larger errors lead to increased misalignment between desired and actual motion during adaptation. This results in a corresponding misalignment between credit assignment and the desired learning, as illustrated in Figures 3 and 4. Further work will be required to determine the extent to which the apparent reduction in learning rates that has been observed with large errors reflects this misalignment versus a true reduction in the ratio between the amount of adaptation and the size of the error.
A recent study by Diedrichsen et al. [34], provides evidence for the occurrence of use-dependent learning alongside error-based learning in reaching arm movements. This use-dependent learning describes a mechanism by which the trajectory of motion in task-irrelevant dimensions is gradually adapted to resemble the motion experienced on preceding trials. Therefore, use-dependent learning resembles motion-referenced learning in the respect that they both depend on the actual motion experienced. However, as noted by Diedrichsen et al. [34], use-dependent learning is oppositely directed from motion-referenced error-dependent learning when a perturbing force is experienced. This is because use-dependent learning would act to increase the extent to which future motions resemble the perturbed movement whereas (motion-referenced) error-dependent learning acts to oppose the effect of this force in order to reduce the extent to which future motions resemble the perturbed movement. A second key difference is that use-dependent learning is readily observed along task-irrelevant dimensions, but is either greatly reduced or entirely absent along task-relevant dimensions [34], whereas the motion-referenced learning that we demonstrate in the current study acts primarily along task-relevant dimensions in which error can be readily defined.
Taken together, the identification of motion-referenced learning and use-dependent learning expand what we know about the role of sensory information in motor adaptation, in particular sensory information about motion. In addition to the role that this sensory input plays in computing motor errors, the motion-referenced learning and use-dependent learning mechanisms respectively explain how sensed motion is specifically associated with error-dependent changes in motor output to reduce the difference between plan and action, and how sensed motion can be used to adapt which motions are planned to begin with.
Information about actual motion states is required for motion-referenced learning. This information can be acquired from delayed sensory feedback or estimated in real time through the use of a forward model, relying on an efference copy of the motor command and past sensory information [4], [35]–[39]. However, since sensory feedback signals and efference copy are noisy, actual motion must be estimated from imperfect information. Several studies have shown that the motor system integrates prior expectations about motion with noisy sensory feedback in order to estimate actual motion in accordance with Bayes Law [40]–[42]. The influence of prior expectations should increase with the level of sensory feedback noise, and so Bayesian estimation should have greater effects on motion estimation and thus on motion-referenced adaptation when noise levels are high.
What is the role of motion-referenced learning in the adaptation to a visuomotor transformation, where there is a dissociation between the actual motion of the hand and the actual motion of the cursor? A definitive answer to this question will require further experimental work since the present study looks at the adaptation to new physical dynamics rather than visuomotor transformations. A priori, it would seem that for visuomotor transformations, learning should be associated with the actual motion of the controlled object (cursor) rather than with the actual motion of the body part that is exerting this control. If motor learning were associated with the actual body motion, it would be difficult to see how large visuomotor rotations could be learned at all, because even late in adaptation, an arbitrarily large mismatch would exist between the planned motion (e.g., the motion of the cursor to its target position) and the actual hand motion. However, previous studies have shown that visuomotor rotations that are wider than the half-width of the generalization function for visuomotor rotation learning (about 30°) are readily learned [43]–[44]. A second point is that since (a) the motor planning during visuomotor transformation learning corresponds to the planned motion of the cursor (rather than the hand), and (b) the relevant motor errors involve the relationship between actual and planned or actual and predicted cursor movements (rather than hand movements) [43], [45], it would seem logical that the learning resulting from errors in this task would be associated with the cursor as well.
Linear state-space models with multiple time courses of adaptation [28], [46] have been invoked as an explanation of savings – the phenomenon that describes the increase in learning rate when an adaption is relearned compared to the initial learning. However, even when complete behavioral washout of the learning is achieved, there appears to be some capacity for savings [29]. This effect cannot be captured by the aforementioned linear models, leading to the suggestion that significant nonlinearities arise even in simple motor adaptation experiments [29]. However, motion-referenced learning provides another possible explanation: Savings after washout may be due to a mismatch between the actual movement directions experienced in the initial learning and the washout trials rather than nonlinearities in the learning process. Such a mismatch would result in incomplete washout in the actual movement directions experienced during initial learning – similar to the residual direction-dependent adaptation that we demonstrate in Experiment 1. Further work will be necessary to determine the extent to which this is the case, but if savings after washout resulted in part from a directional mismatch during washout, then the prediction would be that the amount of savings would be reduced if the washout trials spanned the movement directions experienced early in training, rather than being confined to a single target direction as in [29].
Studies with healthy subjects [47]–[48] and subjects with congenital and acquired cerebellar deficits [49]–[51] have provided evidence that the cerebellum participates in motor adaptation. It has been proposed that the simple spike firing of Purkinje cells in cerebellar cortex contributes to motor output and that error signals carried by climbing fibers modify the strength of the parallel fiber synapses onto Purkinje cells [11], [48], [52]–[53]. This plasticity alters the effect that the information carried in parallel fibers has on the output of Purkinje cells, and thus on motor output [52]–[53]. Since parallel fibers carry sensory feedback (amongst other) signals [38], [54]–[55], this plasticity alters the association between sensory feedback about the actual motion and future motor output and may represent a neural mechanism for motion-referenced learning.
A common technique in neurorehabilitation is the use of partial assistance, where a therapist or device supplements movement in order to allow patients to better approximate a desired motion [56]–[58]. Since partial assistance reduces the difference between the actual and desired motions, our findings would suggest that it improves the alignment between the adaptation that is learned and the desired motion that is being trained. This would improve the efficiency of the training procedure. However, partial assistance would also reduce the magnitude of the motor errors that drive learning. These opposing effects may decrease the overall benefit of this procedure.
Interestingly, a method known as error augmentation that can be thought of as essentially the opposite of partial assistance has recently been proposed as a means to improve the rate of motor learning during rehabilitation. In error augmentation, motor errors are increased beyond normal levels by transiently exposing patients to perturbations that are stronger than those that are to be learned [59]–[61]. The rationale behind this technique is that since error signals drive motor learning, increasing the size of this signal may improve the rate of learning. Our results indicate that, like partial assistance, error augmentation will result in two opposing effects. Whereas, partial assistance increases the alignment between the motion-referenced learning which will occur and the desired learning but reduces the magnitude of the error signal driving adaptation, error augmentation decreases the alignment between the motion-referenced learning which will occur and the desired learning but increases the magnitude of the error signal driving adaptation. Thus, unlike partial assistance, error augmentation may provide a robust error signal for learning, but could in fact lead to decreased learning rates by magnifying the misalignment between the desired motion to be learned and the learned motion in the experienced trials.
The problem of opposing effects resulting from both of these training procedures could potentially be solved by the implementation of a training procedure analogous to the LST training we studied which aligned actual and desired movements, but with stronger-than-normal perturbations. Note that the design and implementation of a training procedure that aligns actual and desired motions is somewhat challenging. Even for the simple planar point-to-point movements we studied in Experiment 2, we first ran another group of subjects to determine the magnitude of the target shifts employed in each trial of our LST paradigm. For training more complex natural motions the challenge will be even greater. With higher-dimensional complex movements, simple manipulations like the altered target position we used in our LST paradigm might not be nearly as effective as a more complex manipulation like the imitation of the entire time course of an altered motion in providing good alignment between actual and desired movement. However, if a training procedure can be created that improves the alignment of the actual motions experienced with the desired motion, even when motor errors are large, such a paradigm may be capable of simultaneously benefitting from increased error-dependent learning and improved transfer of adaptation to the desired motion – the best of both worlds from error augmentation and partial assistance.
The improvement afforded by the LST paradigm or derivatives of it might even be more substantial if used in patients undergoing neurorehabilitation. For example, chronic stroke patients are able to adapt to dynamic environments, but display slower learning rates and higher residual errors than healthy controls [62]–[63]. Interestingly, our modeling efforts suggest that MRL-based training would have an even greater effect on subjects with these types of impairments, with the advantage of LST over STT predicted to be greater in magnitude and longer lasting as shown in Figure S2, because the higher motor errors these subjects normally experience lead to greater-than-normal misalignment under STT (see Text S1). Further studies would be required to determine whether an MRL-based training paradigm could lead to clinically significant improvements in neurologically impaired subjects.
All experimental participants were naïve to the experimental purpose, provided informed consent and were compensated for their participation. All the experimental protocols were reviewed and approved by the Harvard University Committee on the Use of Human Subjects in Research (CUHS).
Subjects performed 10 cm reaching movements in the horizontal plane with their dominant hands while grasping the handle of a 2-link robotic manipulandum. Subjects were seated with their forearm leveled with the robotic manipulandum and supported by a sling. The subjects were presented with 1 cm-diameter circular targets displayed on a vertically oriented LCD monitor. The position of the subject's hand was represented on the LCD monitor by a 3 mm cursor. Position, velocity and force at the handle were measured with sensors installed in the manipulandum at a sampling rate of 200 Hz. The subjects were instructed to produce fast, continuous movements, and were provided visual feedback throughout the movement. Feedback about the movement time achieved was presented at the end of each movement. Ideal completion times (500±50 ms) were signaled by an animation of the target while a chirp sound was played. For movement completion times that were below or above the ideal range the targets were colored blue and red, respectively. The mean peak speed for the movements in all experiments was 0.302±0.017 m/s. In certain movements, the subjects' trajectories were perturbed by velocity-dependent dynamics. This was implemented by a viscous curl force-field at the handle produced by the motors of the manipulandum, Equation 1.(1)
In this equation the constant B represents the viscosity associated with this force-field and has units of N/(m/s). Note that the direction of the force is always orthogonal to the direction of the velocity vector. We assessed the level of adaptation using methods described elsewhere [28]. Briefly, we measured the force pattern that subjects produced when their lateral errors were held to near zero values in an error-clamp [28], [64]–[65]. We then regressed the measured force pattern onto the ideal force required to fully compensate for the force-field. The slope of this regression was used as the adaptation coefficient that characterized the level of learning. For a force profile that is driven by adaptation to a velocity-dependent force-field, our adaptation coefficient represents the size of the bell-shaped velocity-dependent component of the measured force profile. This velocity-dependent component of the measured force profile specifically corresponds to the force component targeted to counteract the velocity-dependent force-field perturbation.
Twenty-eight individuals with no known neurologic impairment (mean age = 19.9±1.8 years; 15 male) were recruited for this experiment. The first twelve subjects practiced the reaching task in 9 different directions (θ = 180°, 210°, 240°, 245°, 270°, 285°, 300°, 330°, 360°) for 254 movements (baseline), and were then trained to compensate velocity-dependent force-fields in a particular movement direction (270°) for 672 movements (training) with the direction of the FFs alternating every 7±2 movements between CW (B = 9 N/(m/s)) and CCW (B = −9 N/(m/s)). Thus the ratio of CW to CCW FF trials was 7∶7. After blocks of 168 training (FF) trials, the pattern of generalization was measured in each direction during a testing block of 40 consecutive EC trials spread across these directions. The direction (CW or CCW) of the last FF presented before generalization testing was balanced across the four training blocks.
A second group of subjects performed the same experiment but with different baseline/testing directions (θ = −30°, 0°, 30°, 45°, 60°, 75°, 90°, 120°, 150°) and training direction (60°). In this experiment the ratio of CW to CCW FF trials was 6∶8 for the first six subjects and 5∶9 for the subsequent ten subjects. The data from the subjects trained at 270° and that from the last ten subjects trained at 60° (CW to CCW FF trial ratio of 5∶9) are shown in Figure 2C. The CW to CCW FF ratio was adjusted to eliminate the bias towards learning the CW FF we observed in the first 12 subjects – details are provided in Text S1. The data for the subjects with the 6∶8 CW to CCW FF trial ratio are compared to the other datasets in Figure S1.
Ninety individuals with no known neurologic impairment (mean age = 22.0±5.9 years; 44 male) were recruited for this experiment. One group of subjects (N = 30) were assigned to the single-target training (STT) paradigm. Here the subjects performed 75 movements in a single direction (90°) to practice the reaching task (baseline) and then were exposed to a CW velocity-dependent force-field (CW; B = 22.5 N/(m/s)) for 125 reaching movements to the same direction (training). The learning level during baseline and training was assessed with randomly interspersed EC movements (p(EC) = 0.2). The mean trial history of angular errors 300 ms into the movement during force-field trials was obtained for this group of subjects and used to design the left-shifted (LST) and right-shifted (RST) training paradigms.
In the LST paradigm, the directions of the reaching targets were adjusted by adding a smoothed fit of the mean trial history of angular errors from the first seventeen subjects of the STT experiment to the desired learning direction on the corresponding trial (90°). We did this so that when subjects reached to these shifted targets their actual motion would be expected to line up with the desired learning direction if the directional error on that trial was similar to that observed in the STT group as illustrated in Figure 3. On the other hand, in the RST training paradigm we subtracted this trial history of angular errors from the STT experiment to the desired learning direction (90°). Therefore these target locations mirrored the LST target locations across 90°. We did this so that when subjects reached to these shifted targets their actual direction of motion would be deviated twice as much from the desired learning direction (90°) as in the STT experiment. In the LST and RST paradigms subjects (30 on each group) also performed 75 baseline movements and then performed 125 training movements using the same CW velocity-dependent FF that was learned by the STT paradigm group. The learning level during baseline and training was assessed by measuring the lateral force profiles produced during randomly interspersed EC trials (p(EC) = 0.2) directed toward the desired learning direction (90°).
We simulated the adaptation process for the STT, LST, and RST training paradigms for the PRL and MRL credit assignment schemes using the model equations and parameters described below and in Text S1. However, in this case, since the experiments and simulations were not aimed at assessing generalization, error in the simulations was defined as the difference between the desired adaptation in the target direction and the actual adaptation in that direction.
We simulated the adaptation process predicted by PRL and MRL for both experiments. We used linear state-space models [28] with local motor primitives to model the adaptation and its generalization (see Text S1 for details). These are discrete (trial-dependent) error driven models, where the error is calculated as the angular difference between the planned movement direction and the actual movement direction, Equation 2.(2)
In the learning rules presented in Equations 3 and 4, the adaptation, x, for given movement direction, θ (θ can take on values encompassing the entire movement space), in a given trial, n + 1, is the sum of the previous adaptation level for the same movement direction weighted by a retention coefficient, A, and the learning occurring in the current trial which is given by the product of the error in the current trial and a local motor primitive, B. For the PRL model (Equation 3), this local motor primitive, B, is centered at the planned movement direction, planned, implying that after a given trial, the maximum adaptation in the entire movement space occurs at the planned movement direction.(3)
Alternatively for the MRL model (Equation 4), the local motor primitive is centered at the actual movement direction, actual, which implies that after a given trial, the maximum adaptation occurs along the actual movement direction.(4)
In our data analysis a few grossly irregular trials were excluded. This included movements that were extremely fast (peak velocity >0.55 m/s) or extremely slow (peak velocity <0.2 m/s), as well as trials with extremely fast (<75 ms) or extremely slow (>2.5 sec) reaction times. This insured that subjects did not initiate movements too quickly, without correctly identifying the location of the target, or too late, indicating that they might have not been attending to the task. For Experiment 1, application of these two criteria resulted in the inclusion of 98.2% of the trials in the 270° group, 96.8% of the trials in the first 60° group (6∶8 CW to CCW FF trial ratio), and 94.9% of the trials in the second 60°s group (5∶9 CW to CCW FF trial ratio). For Experiment 2, 94.7% of the trials in the STT group, 95.2% of the trials in the STT group, and 93.4% of the trials in the RST group were included.
In order to compare the predicted and experimentally observed generalization patterns in Experiment 1, we computed the correlation coefficient between them as well as the p value and F-statistic associated with the slope of the corresponding linear regression. We assessed the significance of the difference in the adaptation between the peaks of the generalization patterns using one-sided paired t-tests. In Experiment 2, differences between learning rates for the three training paradigms (STT, LST, and RST) were assessed with one-way ANOVAs both early (first 3 EC trials) and late (last 3 EC trials) in training. When significant differences arose, post-hoc comparisons were performed using one-sided t-tests.
|
10.1371/journal.ppat.1000084 | The Malaria Secretome: From Algorithms to Essential Function in Blood Stage Infection | The malaria agent Plasmodium falciparum is predicted to export a “secretome” of several hundred proteins to remodel the host erythrocyte. Prediction of protein export is based on the presence of an ER-type signal sequence and a downstream Host-Targeting (HT) motif (which is similar to, but distinct from, the closely related Plasmodium Export Element [PEXEL]). Previous attempts to determine the entire secretome, using either the HT-motif or the PEXEL, have yielded large sets of proteins, which have not been comprehensively tested. We present here an expanded secretome that is optimized for both P. falciparum signal sequences and the HT-motif. From the most conservative of these three secretome predictions, we identify 11 proteins that are preserved across human- and rodent-infecting Plasmodium species. The conservation of these proteins likely indicates that they perform important functions in the interaction with and remodeling of the host erythrocyte important for all Plasmodium parasites. Using the piggyBac transposition system, we validate their export and find a positive prediction rate of ∼70%. Even for proteins identified by all secretomes, the positive prediction rate is not likely to exceed ∼75%. Attempted deletions of the genes encoding the conserved exported proteins were not successful, but additional functional analyses revealed the first conserved secretome function. This gave new insight into mechanisms for the assembly of the parasite-induced tubovesicular network needed for import of nutrients into the infected erythrocyte. Thus, genomic screens combined with functional assays provide unexpected and fundamental insights into host remodeling by this major human pathogen.
| The parasite Plasmodium falciparum causes malaria by replicating inside red blood cells of infected individuals. By exporting many different proteins into the host cell, the parasite changes many of its properties. Knowledge of the identity and function of all the exported proteins will both increase our understanding of the modifications required for parasite survival and provide us with targets that can be inhibited to block the growth of the parasites. Several years ago, a motif within the exported proteins was discovered that allowed them to be exported, which was used to predict the total set of proteins exported to the host cell (the secretome). We show here that the earlier studies have either under- or overestimated the total number of proteins exported into the host cell, and derive a more accurate prediction of proteins exported to the host cell. We validate the predictions by making parasites that express a fusion of predicted exported proteins to the Green Fluorescent Protein (which allows the localization of the protein to be determined visually). This revealed a positive prediction rate of ∼70%. In addition, several proteins were identified that are very likely to play an essential role in infection, with at least one involved in the formation of a structure required for nutrient import.
| Plasmodium falciparum is the protozoan parasite responsible for the most deadly forms of malaria. The symptoms of malaria, which include fevers and chills and can include coma and death, are caused by infection of human erythrocytes by the parasite. After invasion of the erythrocyte, the parasite is contained within a membrane-bound compartment, the parasitophorous vacuole (PV; Fig. 1A). Intracellular parasites induce major changes in several properties of the erythrocyte, including its deformability [1]–[3], permeability of its plasma membrane [4],[5] and its adhesiveness to the endothelium [6]. Underlying these changes are proteins produced by the parasite and exported past the PV membrane (PVM) into the host cell. Examples include ring-infected erythrocyte surface antigen (RESA), which increases the heat-resistance of the erythrocyte [7],[8], and P. falciparum erythrocyte membrane protein 1 (PfEMP1), a cell surface adhesin that, together with another parasite protein, Knob-Associated Histidine-Rich Protein (KAHRP), forms knobs on the surface of the erythrocyte that increase the adhesiveness of the infected erythrocyte [9]–[11]. Another important change in the erythrocyte is the appearance of a large membranous network, the tubovesicular network (TVN; Fig. 1A) [12], which plays a role in nutrient import into the parasite. The formation of this import organelle is entirely dependent on the parasite, and previous studies have shown that development of the TVN is linked to nutrient import into infected erythrocytes [12],[13].
For an exported protein to reach the cytosol or membrane of the erythrocyte, it needs to cross two membranes: the parasite plasma membrane and the PVM (Fig. 1A). The first step requires a canonical ER-type signal sequence, while export of many (but not all) proteins from the PV into the erythrocyte depends on a sequence motif referred to as the Host Targeting (HT)-motif [14] [alternatively known as Plasmodium Export Element (PEXEL)] [15] positioned downstream and proximal to the signal sequence. The HT-motif and PEXEL are identified by different algorithms and have slightly different specificities, but recognize the same core sequence (RxLxE/Q/D) [14],[15]. Identification of this export motif allowed prediction of the set of proteins exported into the erythrocyte (the HT-based Hiller secretome and PEXEL-based Marti secretome), and recently an expanded version of the PEXEL-based secretome (Sargeant secretome), identified with the PEXEL-based prediction program ExportPred, was published. All three secretomes are unexpectedly large, containing over 250 proteins each. Excluding the large protein families of RIFINs (165 proteins) and STEVORs (22 proteins, of which only one is synthesized in an individual parasite [16]), the Hiller secretome contains 113 proteins, the Marti secretome 158, and the Sargeant secretome 267 [17]. The overlap of the Hiller and Marti secretomes is only 59 proteins (53% of Hiller set, 37% of Marti set), underscoring the differences in the prediction algorithms.
A large majority of the identified proteins cannot be annotated, and unfortunately, export of nearly all ‘hypothetical’ proteins has not been experimentally verified. In just two cases have full-length fusions of a hypothetical protein with the Green Fluorescent Protein (GFP) been shown to be exported [14],[18]. In six additional cases only the N-terminal region of the proteins was tested [14],[17]. There was no subsequent verification with full-length proteins. Hence the contribution of additional sequences to the export signal was not established. Interestingly, all P. falciparum proteins known to be exported into the host cell are species-specific. Hence little is known about the functions of exported proteins shared by all Plasmodium species. It is these proteins that are very likely to be involved in the processes that allow the parasites to survive within the erythrocyte, and would make excellent targets for prophylaxis.
Thus identification of the exported proteins and their function will give better insight into the parasite-erythrocyte interaction. We have therefore investigated several parameters of protein export in P. falciparum to refine the export prediction. We show that the HT motif-based algorithm PlasmoHT is limited by the identification of ER-type signal sequences in P. falciparum and that this represents an important difference with ExportPred. We furthermore identify the putatively exported proteins conserved in other Plasmodium species and use this as a high value candidate set to validate the export prediction. Importantly, we find that testing N-terminal regions alone can lead to critical oversight on location of the protein and full-length fusions should be evaluated in order to validate positive predictions. Finally, this study shows that despite the fact that essential genes cannot be knocked out in blood stage Plasmodium berghei, insight into biological processes can be attained by utilizing rapid transgene expression. Even so, limitations in the genetic system of P. falciparum have not allowed any large scale validation of export predictions, and the testing here improves significantly on previous attempts and enables functional analyses.
Secreted P. falciparum proteins can have recessed signal sequences and thus be difficult to identify with standard signal sequence identification programs such as Signal P [19]. In prior attempts to identify exported proteins Hiller et al. modified SignalP to examine the first 100 residues instead of the default 70 residues [14], while Sargeant et al., abandoned SignalP but instead required the presence of a stretch of 10–25 hydrophobic residues near the N-terminus, separated by a spacer region from the start methionine [17]. However, neither prediction was based on multiple validated sequences. We therefore set out to determine how well SignalP predicts the secretion of known secreted P. falciparum proteins in order to obtain a more accurate prediction of secretion. SignalP by default examines the N-terminal 70 amino acids of a protein and determines the most likely cleavage site (represented by the C and Y scores), the maximal signal sequence residues (MaxS), and the average signal sequence residues (MeanS) as measured from the N-terminus to the most likely cleavage site [19]. We plotted the MaxS and MeanS values of 82 proteins known to be transported through the secretory pathway (with destinations as diverse as the apicoplast, parasite plasma membrane, rhoptry and erythrocyte cytosol and membrane, see Table S1) (Fig. 1B). Seventy-six proteins had a score above the MaxS threshold for a signal sequence (0.82), and in many cases the values were close to maximal. However, six (7.3%) had a MaxS score below the threshold, while 21 (25.6%) had a MeanS score below the threshold (0.47); four of these proteins (4.9%) had both a MeanS and MaxS score below threshold. Figure 1B shows that a protein (RESA) with a MaxS score of 0.571 and MeanS score of 0.113 can be secreted. Apparent from Fig. 1B is that as the MeanS scores decrease, the MaxS values become more scattered. Since a lower MeanS score often is indicative of a recessed signal sequence, it is possible that the increased length of the N-terminus before the hydrophobic core of the signal sequence allows for lower MaxS scores. The large percentage of proteins with a low MeanS score is likely a reflection of the prevalence of recessed signal sequences in plasmodial proteins. Thus clearly, while many secreted P. falciparum proteins adhere to the same principals of secretion as higher eukaryotes, there may be a degree of flexibility in the signal that allows proteins with lower MaxS scores to be secreted, which in turn leads to under-prediction of secreted proteins by SignalP.
To evaluate whether proteins predicted to be exported to the erythrocyte contained any distinguishing features in their signal sequences, we plotted the MaxS and MeanS scores of the proteins in the previously published Hiller and Sargeant secretomes [14],[17]. Since Hiller et al. used the default 0.82 MaxS cut-off to identify signal sequences, all MaxS scores predictably were above 0.82, but the MeanS scores were scattered within the same range of scores as seen in the group of known secreted proteins (Fig. 1C). The Sargeant secretome is based on the SignalP-independent algorithm ExportPred. As shown in Fig. 1D, like the Hiller secretome, the proteins in the Sargeant secretome fall on the same general curve as the known secreted set, i.e., as the MeanS decreases, the MaxS values become more scattered. However, 50 out of 267 proteins (18%) had signal sequence scores below the lowest experimentally proven functional MaxS score of 0.571 for Plasmodium; the fitness of these values for secretion remains to be validated experimentally. Thus one major reason for the different predictions in the Sargeant secretome relative to the Hiller secretome lies in the extremely low SignalP scores the former allows.
To obtain a maximal secretome whose signal sequence predictions were within the range of experimentally verified signals for Plasmodium, we re-evaluated the HT-based secretome based on a MaxS cutoff of 0.571. After manual curation this identified an additional 22 putatively exported proteins. We also accommodated recent updates in gene calling in the P. falciparum genome to revise secretome predictions obtained by MEME/MAST, the programs used to identify the initial Hiller secretome. Furthermore we utilized a second algorithm, HMMER, to expand the MAST output, which yielded an additional 171 proteins. Combining these proteins with the Hiller secretome (251 proteins) yielded an expanded secretome (van Ooij secretome) of 422 proteins (Table S2). The algorithmic predictions of this secretome were further curated for proteins with MaxS scores between 0.58 and 0.82. No further curation was undertaken based on expression, functional or structural data. Removing the RIFIN and STEVOR family members left 224 proteins (no PfEMP1 family members are present in this secretome since they contain an internal signal sequence and are thus not recognized by Signal P). The van Ooij secretome approaches the Sargeant secretome (267 proteins) in size, but utilizes signal sequence parameters that closely mimic experimental predictions and an HT-motif that is closely linked to the Hiller secretome. The MaxS and MeanS values of the proteins in the van Ooij secretome (omitting the RIFINs and STEVORs) were plotted against each other and were found to follow a similar pattern to the Hiller secretome (Fig. 1D). A comparison of the secretomes and the distribution of MaxS scores is listed in Table S3.
It has been noted previously that exported proteins are frequently encoded by genes consisting of two exons [15],[17]. This effect is exaggerated somewhat by the conservation of this exon-intron structure in the rif and stevor genes, which make up 38.5 and 7.2%, respectively, of the expanded secretome. Even after removal of the rif and stevor genes, many proteins in the van Ooij secretome are encoded by genes that have this two-exon structure, comprising 69.4% percent of the secretome, and 65% in the Sargeant secretome, with one-exon structure the second-most prevalent in both sets. Distribution of exon structures is shown in Table S4.
Previous investigations of exported proteins have focused primarily on the proteins unique to P. falciparum or members of large antigenic families unique in other Plasmodium species. We were interested in identifying those proteins that are conserved among all Plasmodium species because they likely perform functions necessary for the interaction of every Plasmodium species with the host erythrocyte. Hence they define the core of the interactions of the parasite with the host cell but remain completely unknown. Our initial analysis of the secretome of rodent malaria parasites P. berghei, Plasmodium chabaudi, and Plasmodium yoelii (Rodent Malaria Parasites (RMP); [14]) and studies by Sargeant et al. [17] on the secretomes of several Plasmodium species describe smaller secretomes than those of P. falciparum. This could reflect a less complex interaction of those species with the host erythrocyte but is at least in part due to the less complete annotation of the genomes (see Table S5 for examples of changes in annotation of RMP and Plasmodium vivax proteins that uncover signal sequence and HT-motifs). Therefore, conserved genes were detected by searching the RMP genomes for orthologues of P. falciparum secretome proteins. Since synteny breakpoints often contain species-specific genes [20], we narrowed our search to proteins encoded by genes that had maintained their genomic localization, which additionally aids in identifying those orthologues in which parts of the protein have diverged and thus have a lower score in a BLAST analysis, but are nonetheless bona fide orthologues. On the basis of these criteria, 11 proteins were identified in the original Hiller secretome (Table 1), while the van Ooij secretome contains an additional 18. All syntenic genes were also conserved in Plasmodium vivax, indicating the widespread conservation of the genes.
A similar search by Sargeant et al. identified nine putatively exported P. falciparum proteins (identified using the PEXEL) that were conserved in P. vivax and P. yoelii [17]. The overlap of core set predictions by Sargeant et al. and those listed in Table 1 consists of only four proteins (indicated by an asterisk). Two conserved proteins identified by Sargeant et al. but not listed in Table 1 were not recognized as having an HT-motif (underscoring the difference between the predictive algorithms PlasmoHT and ExportPred), while two others did not have a clearly recognizable RMP orthologue in the PlasmoDB database. The remaining protein had a MaxS score of 0.72 and was identified when the MaxS threshold was set at 0.58.
Only two of the eleven syntenic genes, pf13_0090 and pfd0495c, have the classic two-exon structure. In pf14_0607 the signal sequence is encoded by a small exon, with the HT-motif encoded close to the 5′ end of the second exon, but the entire gene contains twelve additional exons. pfl1660c is encoded by five exons, with the signal sequence and the HT-motif, as well as the majority of the protein, encoded by the first exon. The other seven genes consist of a single exon. It is possible that the classic two-exon structure may reflect a mechanism by which P. falciparum has been able to convert non-exported proteins to exported proteins through the addition of a small exon encoding a signal sequence and HT-motif. None of the syntenic genes are located near the telomeres. This is not surprising as the telomeres contain many species-specific genes.
These proteins provide a high value set to test the predictive value of each of these secretomes. But since we were interested in testing the prediction of the export motif independent of the SS, we restricted our subsequent analyses to conserved proteins identified in the Hiller secretome (Table 1), which by virtue of using a MaxS cut-off of 0.82 dramatically reduces false positive predictions for recruitment into the secretory pathway, likely to be more prevalent in the more expansive Sargeant and van Ooij secretomes. This assumption is justified by the data in Fig. 1B showing that the vast majority of known secreted P. falciparum proteins have MaxS values higher than 0.82.
Sargeant et al. showed that the N-terminal residues of one of the conserved proteins of the Hiller secretome listed in Table 1, PF14_0607, targeted a GFP-fusion to the lumen of the PV, but were not able to promote export to the erythrocyte [17]. This raised the possibility that the leader sequence did not faithfully reflect the export properties of the complete protein. We therefore tested fusions of the full-length gene and the first 89 codons of pf14_0607 to gfp. As shown in Fig. 2A, green fluorescence associated with the full-length fusion was indeed detected in bright punctate spots in the erythrocyte. However, the fusion of the first 89 codons (containing the signal sequence and the HT-motif, but missing the transmembrane domains) was not (or extremely poorly) exported to the erythrocyte (Fig. 2B). Sargeant et al. postulated that the lack of export of the fusion of the N-terminal region of PF14_0607 was due to the presence of a phenylalanine in position 4 of the HT-motif. When this residue was changed to an alanine in the 89 codon-GFP chimera, the protein was indeed robustly exported (Fig. 2D) into the erythrocyte cytosol in an HT-dependent manner (compare Fig. 2C to 2D). These data suggest that a phenylalanine residue at position 4 can indeed hinder export when acting in a short fusion protein that contains just the signal sequence and the HT motif, but in the full-length protein it can be sustained, likely due to structural constraints on the leader imposed by the rest of the protein. Moreover the fact that the mutated 89 codon-GFP fusion is delivered to the erythrocyte cytosol (Fig. 2D), while the full length gene of interest localizes to punctate intraerythrocytic structures (Fig. 2A), suggests that transmembrane and other regions of the protein influence its final destination in the host cell. These data are consistent with our prior studies that have systematically established that while the HT motif is essential for export to the erythrocyte, sequences upstream and downstream provide information [14], [21]–[23] for export resides in overall domain structure that can be conserved across evolutionary distance. The data in Fig. 2 highlight a case where sequences significantly downstream of the HT-motif in the functional protein can likely influence overall structural information needed for HT-motif dependent export. Therefore all subsequent studies were performed with full-length (or nearly full-length) protein fusions (see Table 1).
Genetic manipulation in Plasmodium, while possible, remains a slow process. Successful generation of stable transgenic parasites and initial expansion in amounts required for standard characterization can take 4–6 weeks, while integration into the chromosome via homologous recombination can take months. Therefore most studies have been limited to plasmid-based analysis, which requires continuous drug pressure in culture. Together these limitations have severely handicapped systematic analysis of transgenes in P. falciparum.
To expedite the production of the stable cell lines expressing a fusion of the syntenic genes with gfp, we utilized the piggyBac transposition system (see Fig. 3) [24],[25]. This system is based on the integration of specific DNA sequences by the lepidopteran transposase into the sequence TTAA. In the P. falciparum strain 3D7, this sequence is present 311,155 times (124,733 times in coding regions), presumably allowing for nearly random integration [25]. To make stably transfected parasites, P. falciparum strain 3D7 was transfected with two plasmids, pHTH [25], which encodes the transposase, and a second plasmid that contains a drug marker (human dihydrofolate reductase in this case) and the gfp-fusion gene, flanked by the inverted repeats recognized by the transposase. Expression of the transposase then promotes the integration of the inverted repeats into the genome. The plasmid encoding the transposase does not contain a resistance marker and is presumed to be lost during propagation of the parasites. All the transgenic parasites obtained in this study were detected within fourteen days after initiation of drug selection and could be maintained in long-term culture over several months without addition of drug while retaining the gfp transgene (Fig. 3).
Integration into genes required for export, resulting in parasites that are no longer capable of protein export, is extremely unlikely, as a subset of exported proteins is likely to fulfill an essential role in the erythrocyte (see below). In principle, integration into essential genes can also occur, but since transposition is likely to occur at many different sites, the selection process after transfection will allow the growth of only those parasites that have no growth defect relative to other transfectants.
While bioinformatics predictions are powerful in identifying candidates for export, they need to be validated in functional assays. We were interested in understanding the HT prediction, independent of the signal sequence prediction, and validated the export of the conserved proteins by making fusions with GFP at the C-terminus. Using the piggyBac system, we obtained stably resistant parasites for 10 out of the 11 genes listed in Table 1 within 14 days after transfection; only in the case of PF13_0218-GFP were we unable to obtain stable lines (Fig. 3). Each drug-resistant transformed culture displayed a uniform population of fluorescent parasites such that the export of GFP to the erythrocyte could be ascertained without subsequent cloning of the population. We determined the integration sites for three different clones, and found that transposition had occurred in intergenic sequences as well as open reading frames (Table S6).
Eight of the ten transgenic parasite lines synthesized a fusion protein of the expected size, as judged by anti-GFP immunoblot (Fig. S1). In the two other cases, PF13_0317-GFP and PFC0555c-GFP, a large majority or all of the anti-GFP signal was detected in a band approximately the size of free GFP. Hence PF13_0317 and PFC0555c could not be analysed for export and were excluded from further analysis. The entire analytical procedure is outlined Fig. 3.
Five GFP-fusions, PFA0210c, PFL0600w, PF14_0607, PFD0495c and PFC0435w, were found to be exported to the erythrocyte (Fig. 4A). PFA0210c-GFP and PFL0600w-GFP were distributed evenly throughout the erythrocyte and were also detected at high levels in the parasitophorous vacuole, which may reflect the high level of expression in the transgene system that uses the strong calmodulin promoter for better visualization of the fusion protein. PFD0495c-GFP, a transmembrane protein was detected at the periphery of the erythrocyte, in intraerythrocytic membranes and the vacuolar parasite. PF14_0607-GFP and PFC0435w-GFP, which also contain transmembrane spanning regions, displayed punctate spots in the erythrocyte as well as closely associated with the PV (Fig. 4A). PFC0435w ends with the sequence DEL but we do not think this C terminal sequence impacts its localization. This is because although xDEL at the C terminus can function as a retention signal for soluble proteins in the lumen of the ER, the ER retention of transmembrane proteins does not involve the xDEL sequence. It should be noted that PFC0435w is not a soluble protein. It is a single pass transmembrane protein with its C-terminus predicted to be on the cytoplasmic face of the ER. Hence PFC0435w is not likely to be an ER retained protein. Three fusion proteins, PFL1660c, PF10_0177 and PF13_0090, were not detected within the erythrocyte (Fig. 4B). The lack of export of PFL1660c-GFP is particularly surprising since it is also predicted to be exported by ExportPred [17]. The protein is annotated to be an aspartyl protease, was detected in small structures inside the parasite (possibly the apicoplast; Fig. 4B) and its N-terminus was recognized as an apicoplast targeting sequence by the PATS program [26]. Residue 5 of its HT-motif is a serine, which is an unusual amino acid for this position, and may be part of the reason the protein is not exported. The perinuclear staining of PF10_0177-GFP indicated the protein was not exported. It should be pointed out that for technical reasons, this fusion contained only the N-terminal 1015 residues (out of 3162, encompassing the first exon only). Although we think it unlikely, it is formally possible that lack of downstream sequences may have compromised export mediated by its HT motif. However, the two exons of the orthologues of PF10_0177 in RMP and P. vivax are annotated as two separate genes, in which case the entire gene was fused to gfp. PF13_0090, which is annotated as a possible ARF family member, also appeared to associate with the parasite, with no detectable export to the erythrocyte (Fig. 4). When the ability of the N-terminal region of PF13_0090 to direct secretion was tested by fusing the first 59 codons to gfp, the resulting fusion protein accumulated inside the parasite rather than in the PV (data not shown), suggesting that despite its high MaxS score, the predicted SS may not support recruitment into the secretory pathway.
In summary, the data in this section suggest that six of the eight tested conserved proteins contain bona fide signal sequences that enable their recruitment into the secretory pathway. Further, 5 of the 8 could be detected exported to the erythrocyte (either diffuse or in punctuate structures) when fused to GFP. Thus we could confirm export with a prediction rate of ∼70% (Fig. 3B). When the overlap of the Hiller, van Ooij and Sargeant secretomes is considered (Fig. 3B), only three of the four conserved proteins predicted to be exported by all three secretomes are exported, while one (PFL1660c) may be transported to an internal secretory destination such as the apicoplast. So while the rate of successful prediction increases when the intersection of two predictions is considered, the success rate is still below 100% (and likely not to exceed to 75%), indicating that there may be additional determinants that are currently not recognized by in silico prediction programs. Our data also show that two of Sargeant's predictions of conserved exported proteins [17] were false negatives. One of these proteins (PF14_0607) is excluded from ExportPred because the N-terminal region is not able to promote export, even though the full-length protein is exported. In the other protein (PFC0435w) the HT-motif is separated from the signal peptide cleavage site by 68 residues, which likely exceeds the length probability of the spacer region between the N-terminal hydrophobic region and the HT motif (PEXEL) allowed by ExportPred.
In order to learn more about the role of the conserved proteins during the intraerythrocytic cycle, we attempted to delete the genes encoding nine of the eleven of these proteins in the P. berghei system [27],[28]. The deletions were attempted twice for each gene, and in no case were mutant parasites obtained, while transfections performed concurrently with unrelated deletion plasmids did produce mutant parasites. The inability to delete these genes strengthens the belief that these genes encode proteins essential for the growth of the parasite within the erythrocyte, which is not unexpected considering the high degree of conservation of these genes. Results of the deletions are listed in Table 1.
Most exported proteins have no in silico annotatable features. However transgenic parasites expressing GFP fusions potentially provide powerful reagents to enable functional characterization. This is especially so when the transcriptional profile of the gene of interest mimics the largely constitutive activity of the cam promoter with peak expression from 24–40 h of infection (Fig. S2). Many secretome genes are highly stage specific and show peak expression times at segmenters and early ring stages. However the expression profile of PFC0435w peaks at 24–40 h of infection and closely parallels that of cam in the second half of the asexual life cycle (Fig. S2). We confirmed expression of the fusion of the transgene at 24, 36 and 42 h of infection demonstrating that a single fusion product is detected in rings and becomes prominent in the trophozoite and schizont stages (Fig. S3). These data confirmed that expression of PFC0435w-GFP closely mimicked that predicted for endogenous PFC0435w.
We next examined transgenic lines expressing PFC0435w-GFP at the trophozoite stage by fluorescence microscopy. We confirmed that the fusion was detected in the periphery of the parasite as well as intraerythrocytic structures. However exported PFC0435w-GFP (Fig. 5Aiii, arrow) did not colocalize with Maurer's clefts (Fig. 5Aiii, arrow head), major intraerythrocytic structures implicated in protein export to the erythrocyte membrane. Clefts are flat lamellar membranes exported from the parasite to the erythrocyte and our recent data suggest that they are targeted by the HT motif as conduits for protein export to the cytoplasm and membrane of infected erythrocyte [29]. It is possible that colocalization between one cleft (of ∼10) and PFC0435w-GFP in the trophozoite stage in Fig. 5Aiii (see asterisk) may reflect transport of the fusion through a cleft, en route to a distinct intraerythrocytic destination. A chimeric gene containing just the HT motif and a transmembrane domain expressed by the cam promoter, drives efficient export to the clefts (Fig 5Avi, arrow heads), confirming that expression of a transgene via the cam promoter does not preclude its quantitative localization to clefts. Together these data suggested PFC0435w-GFP was not a major resident protein of clefts.
We were next interested in determining the relative distribution of PFC0435w-GFP in a tubovesicular import pathway that appears to be distinct from clefts [12]. To do this we determined the distribution of PFC0435w-GFP in relation to Rhodamine B, a fluorescent dye that does not freely diffuse into erythrocytes but is actively internalized into infected erythrocytes by the TVN (Bhattacharjee and Haldar, unpublished data). As shown in Fig. 5B–D, Rhodamine B fluorescence accumulated to a high level within the parasite, as well as in tubovesicular structures that extend from the erythrocyte membrane to the parasite. Remarkably, PFC0435w-GFP could be detected in discrete localized regions of the TVN. The GFP-tagged protein apparently connects the vacuolar parasite to a large membrane loop of the TVN. Analysis of different optical sections from an infected cell confirmed that the protein appeared to form a bridge between the vacuolar parasite and the intraerythrocytic loops; curiously this bridge itself was relatively poorly stained with Rhodamine B (Fig. 5B, D), suggesting it did not retain significant amounts of the internalized probe. PFC0435w-GFP was also detected in junctions between TVN structures closer to distal regions of the TVN near the erythrocyte membrane (Fig. 5B, D and schematized in 5C). These data provide the first direct evidence of a parasite protein quantitatively located at the junction of the TVN and the PVM, and we therefore renamed the protein TVN-junction protein 1 (TVN-JP1). TVN-JP1-GFP was also seen at points where the PVM formed small buds (see arrows in Fig. 5D), suggesting it may define junction sites of loop formation at the PVM.
The TVN develops during ring to trophozoite development (which occurs approximately 24 hrs after invasion of the erythrocyte, midway in the 48 hr developmental cycle). To further investigate the function of TVN-JP1 in TVN assembly, we examined its distribution in rings. At this stage TVN-JP1 was present in small vesicles that moved rapidly (in seconds) in the host cell cytosol (see Video S1, Fig. 6). This was in sharp contrast with immobile GFP junctions connecting large membrane loops in association with the mature TVN in trophozoites. The movement of the vesicles appeared random, moving away from or towards the parasite equally, and no build-up of the protein at the periphery of the infected cell was detected, indicating that the vesicles did not fuse with compartments at the erythrocyte periphery. However movement of one vesicle (above left-hand cell in Video S1) was highly restricted, suggesting that it may be attached in the erythrocyte. Remarkably, we also detect a membrane connection between this vesicular structure and the parasite, as indicated by the presence of GFP fluorescence spanning from this vesicle to the parasite. This type of vesicle movement has been detected previously in the form of acridine orange-labeled vesicles, which were also detected primarily during the ring stage [30] but their function was unknown. Our data suggest that they are precursors to TVN assembly. The acridine orange-labeled vesicles were detected in wild-type P. falciparum, making it unlikely that the appearance of PFC0435w-containing vesicles is an artifact resulting from overexpression of the protein. Together, these data suggest that export of highly mobile vesicles containing TVN-JP1 is an early step in the formation of the TVN. Since TVN-JP1 domains in the erythrocyte are immobile in trophozoites, anchoring of these vesicles and membrane connections between them and the vacuolar parasite are likely to precede sphingolipid-dependent budding of large vesicles and loops originally described as the first step of TVN biogenesis [31]. Thus although TVN-JP1 does not stain large domains of the TVN, its export is nonetheless expected to be important in erythrocyte remodeling and for proper development of the TVN. Our studies establish that rapid genetic methodologies enable identification of genes and mechanisms linked to formation and function of the P. falciparum TVN, suggesting they provide new targets for prophylaxis.
In this study we have provided empirical testing of the prediction of protein export from the malaria parasite P. falciparum and validation of the export prediction for a set of 11 proteins conserved throughout the genus Plasmodium. This is the first demonstration of export of proteins conserved in several species of Plasmodium. Relative to the size of the entire secretome this is a relatively small set of proteins to test, but their conservation across species confirms they are high-value determinants. Due to the limitations of genetic manipulation in P. falciparum, this kind of genome-wide analysis required utilization of a robust transgene expression system such as piggyBac, which shortened the time required for establishing stably transfected parasites and had a very high success rate (10 out of 11 genes could be investigated). Since we selected from a set that had the highest MaxS scores, we maximized chances that these proteins are recruited to the secretory pathway (although the putative ARF is likely not). This increases the ability to evaluate the predictive value of the HT motif in mediating translocation beyond the PV to the erythrocyte Extrapolating the positive prediction rate of ∼70% of the tested proteins to the entire secretome predicts that well over 200 proteins will be exported, confirming original projections that host remodeling is a highly complex process. Nonetheless, due to the difficulty in predicting signal sequences in Plasmodium, we expect that in the more expansive secretomes (such as the van Ooij and Sargeant secretomes), the positive prediction rate will be lower. As pointed out previously, two of the proteins shown to be exported were not recognized by the ExportPred algorithm, likely due to a relatively large number of residues separating the HT motif and the signal sequence (PFC0435w) and additional information in the remainder of the proteins (PF14_0607), indicating that parts of the algorithm limit the prediction of the secretome. It is difficult to explain fully why a fusion containing only the 90 N-terminal residues of PF14_0607 remained in the PV while the full-length proteins is exported, other than that the remainder of the protein must aid in exporting the protein across the PV. It is interesting to note that the N-terminal sequence of PF14_0607, and that of 66 other secretome proteins, was recognized by the apicoplast-targeting prediction program PATS [26] (Table S2). This underscores the difficulty of predicting transport when multiple transport signals are present. How overlapping targeting sequences are resolved remains unclear.
Only 11 proteins of the original Hiller secretome were conserved in the RMP and P. vivax. This constitutes 9.7% of the total secretome (without the RIFINs and STEVORs), a surprisingly low number. In the Sargeant secretome the percentage of conserved proteins is even lower, 3.3%, while in the van Ooij secretome it is 12.9%. Since we found that the annotation of the orthologous genes of exported proteins often did not include the 5′ region (which encodes the signal sequence; Table S5), it is not yet possible to determine a complete secretome for the RMP or P. vivax to provide a direct comparison of the secretomes and identify all RMP or P. vivax-specific exported proteins. Even so, a large number of proteins in the P. falciparum secretome are species-specific, making it likely that they are important for P. falciparum-specific symptoms but possibly not survival of the parasite within the erythrocyte. Indeed, as shown in Table 2, none of the published deletions of genes encoding exported proteins, all P. falciparum-specific, are lethal to the parasite. Thus P. falciparum contains species-specific mechanisms for sequestration involving PfEMP1 and KAHRP [10], stabilization of the cytoskeleton through RESA [7],[8] and MESA [32], among other proteins, as well as formation of intraerythrocytic structures the Maurer's Clefts, by SBP1 [33] and MAHRP [34], but these processes are not required for the growth of the parasites in culture. Since all the exported proteins studied to date are P. falciparum-specific, little is known about the functions shared by all Plasmodium species. It is likely that the conservation of the exported proteins, along with the likely function in host cell remodeling, will make these proteins essential factors for parasite survival with the host cell. Consistent with this is the finding that none of the genes encoding the conserved exported proteins could be deleted in P. berghei, indicating that they are indeed required for parasite growth (Tables 1, 2).
The results presented here suggest that at least one conserved exported protein is involved in formation of the TVN, an organelle of nutrient uptake. At the time of its original identification [12],[31], only a parasite sphingomyelin synthase activity was known to be important for TVN synthesis, but no report has described the localization of this sphingomyelin synthase activity. Members of the P. falciparum protein family ETRAMPS as well as the P. berghei protein UIS4 have been detected on vesicular elements budding off the PVM of blood stage and liver stage parasites, respectively [35]–[37], but since they are detected primarily on the vacuole, they are not thought to be TVN resident proteins or have TVN-specific functions. This study reveals TVN-JP1 as the first TVN-specific protein marker. The distribution of the protein, in a bridge structure between large intraerythrocytic loops and the PVM, could be indicative of a structural role, which would be congruent with the finding that the protein is initially found in rapidly moving vesicles in the erythrocyte cytosol. The evidence that the TVN begins as small, mobile structures in the erythrocyte cytosol is highly unexpected since the mature TVN organelle is a relatively immobile structure in the trophozoite-infected erythrocyte as are the TVN-JP1 junction and large loops between these junctions (Fig. 6,7). The localization of TVN-JP1 at sites of PVM budding in trophozoites may reflect sites of budding that contribute to TVN development even at these later stages of growth (summarized in Fig. 7). Considering the size of the TVN within the erythrocyte cytosol and the functions in protein import it plays, there are undoubtedly more proteins involved in the formation of this organelle. The total number of conserved proteins that could be involved in TVN function and/or formation based on Hiller, Sargeant and van Ooij secretomes are expected to range from ∼10–29. The total number of P. falciparum-specific proteins associated with TVN function is presently difficult to predict.
The approach of a genome-wide screen for exported proteins combined with application of several criteria consecutively (conservation in other Plasmodium species, validation of export, timing of expression and localization data) has identified possible functions for a protein that was beyond annotation by in silico methods. By altering the criteria for selection, it should be possible to uncover the role of other pathogenic exported proteins of hypothetical function and their contribution to intracellular survival and pathogenesis.
Plasmodium falciparum parasites used in this study were of the 3D7 lineage and maintained in human erythrocytes, blood type A+, in RPMI-1640 medium supplemented with 91.9 µM hypoxanthine, 11 mM glucose, 0.18% sodium bicarbonate and 10% human serum (cRPMI). Plasmodium berghei, strain ANKA, parasites were maintained in Swiss mice.
P. falciparum transfections were performed as described by Wu et al. [38] with modifications described by Deitsch et al. [39]. Briefly, erythrocytes were loaded with 100 µg of plasmid DNA containing the transgene, and where necessary, 100 µg of plasmid containing the piggyBac transposase (pHTH), by electroporation using a BioRad GenePulser set at 0.310 kV and 950 µF. Transfected erythrocytes were washed three times with cRPMI and immediately mixed with Percoll-purified infected erythrocytes to a parasitemia of 1–5%. Drug selection was initiated by addition of WR99120 (Jacobus Pharmaceuticals, Princeton, NJ) to 2.5 nM 72–96 hours after infection. The medium was changed every other day until parasites were detected by Giemsa staining. Most transfected parasites could be detected within two weeks after drug selection. The P. falciparum strain expressing the HTTM-GFP fusion is described in Bhattacharjee et al. [29].
P. berghei transfection was performed according to published methods [27],[28]. Briefly, 5–8 ml blood was extracted from an infected Wistar rat at a parasitemia of <3%. Blood was mixed with an approximately equal volume of RPMI with L-glutamine and HEPES, supplemented with 0.085% NaHCO3 and 25% fetal calf serum (culture medium) with 0.3 ml heparin stock solution (200 I.U./ml) and spun down. Cells were resuspended in 150 ml culture medium and allowed to mature to the schizont phase overnight by incubation at 37°C in an atmosphere of 5% O2, 5% CO2, 90% N2. Mature parasites were harvested by density centrifugation on Nycodenz and subsequently transfected with 5–10 µg linearized plasmid DNA using the AMAXA device.
To search for the presence of an ER-type signal sequence, proteins were analyzed using the SignalP-NN algorithm from the SignalP V2.0.b2 program (http://www.cbs.dtu.dk/services/SignalP-2.0/) [19]. This algorithm was trained on eukaryotic sequences and the input protein sequences were truncated at amino acid 100.
MeanS scores and MaxS scores for each sequence were plotted against each other. In our previously described secretome, all sequences with a maximum S-score above the default cutoff of 0.82 determined by the SignalP-NN software, were considered to have a signal peptide. For the revised secretome set defined in this paper, a MaxS-score of 0.58 was used as the discriminating factor for secretion signals.
The HMMER software suite (http://www.psc.edu/general/software/packages/hmmer/manual/main.html) was used for making a hidden Markov model (HMM). The initial model was built using a training set that consisted of five proteins known to be exported to the erythrocyte (GBP130, PfEMP2, PfEMP3, HRP1, HRP2), and this model was used to search a database of 3D7 proteins with a signal sequence predicted using SignalP 2.0 default maximum S-score as a discriminating factor [14]. The proteins identified from this search were used as a new training set to create the final HMM. This model was used to screen all proteins predicted to have a signal sequence according to the SignalP maximum S-score cutoff of 0.58 (see Signal Sequence Prediction section below) for sequences that contain the HMM.
The chromosomal position of all 3D7 sequences in the original PlasmoHT set [14] were compared with the chromosomal position of their orthologues in the rodent malaria P. berghei and P. yoelii described in Kooij et al. [25]. e-values were used as a guide to predict functional class and synteny was used to select the functional orthologues.
The syntenic genes were amplified with the primers listed in Table S7 (which also list the amount of the gene amplified) using P. falciparum 3D7 genomic DNA as template. The resulting DNA fragments containing pf13_0317, pf13_0090, pf13_0218, pfa0210c, pfc0555c, pfl0600w, and pfd0495c were digested with the indicated restriction enzymes and cloned into pBLD70 to make a fusion with gfp. pBLD70 was made by cloning GFP mut2 followed by an XhoI site into pBluescript using the NheI and SacI sites. This fusion was liberated with XhoI and cloned into XhoI-digested pBLD194. pfc0435w, pf10_0177, pf14_0607, and pfl1660c were amplified further using AttB1 and AttB2 primers to attach full-length Att-sites on the DNA fragments, and cloned into pDONR and subsequently transferred to the GFP-fusion expression vector pBLD194 using the GATEWAY technology (Invitrogen, Inc, Carlsbad, CA).
To make a truncated version of pf14_0607, the 5′ 89 codons were amplified using the primers listed is Table S7 using plasmid pBLD201, the pf14_0607 full-length expression vector described above, as template. The resulting DNA fragment was cloned into pDONR and subsequently moved to the GFP-expression vector pBLD194. To make the F-A mutation or the replacement of the HT-motif, the 89 codon 5′ region was amplified in two parts using the primers listed in Table S7 (which include the intended mutation). The resulting fragments overlapped by 47 bp (including the introduced mutation), and were mixed with the 5′ upstream and 3′ downstream primers in a PCR to produce the entire 89-codon fragment that contains the intended mutation. This fragment was again amplified with the AttB primers to introduce full AttB sites and then cloned into pDONR and subsequently pBLD194 using the GATEWAY technology.
Vectors for the targeted deletion of the syntenic genes in P. berghei were made as follows. A region of DNA upstream and downstream of the target gene was amplified, with each fragment in the range of 650–1100 bp (with exception of the region upstream of PB000767.00.0, which was approximately 3 kb), using 100 ng P. berghei genomic DNA as template. Primers used for amplification are listed in Table S7. The amplified DNA fragments were inserted into the pDONR vector using the GATEWAY technology. The upstream fragments were released from the resulting vectors using SacII and SpeI and cloned into corresponding sites in pL0001. Subsequently the corresponding downstream regions were cloned into the resulting vector with HindIII and KpnI to form the final deletion vector with the upstream and downstream regions flanking the hDHFR resistance cassette. For transfection the plasmids were linearized by digestion with KpnI.
To obtain stable parasite lines with integrated copies of the transgene, parasites were transfected as described above with 100 µg of vector containing the transgene and 100 µg of vector containing the transposase, pHTH [24],[25]. To identify the sites of integration, genomic DNA was extracted from ∼2–9×108 cloned parasites. Parasites were extracted from infected erythrocytes by lysis with 0.05% saponin in PBS. Parasites were pelleted and lysis was repeated. Parasites were washed with PBS twice and frozen at −80°C. Genomic DNA was isolated by standard procedure. Briefly, parasites were resuspended in lysis buffer (10 mM Tris, 20 mM EDTA, 0.5% SDS, 25 µg/ml Proteinase K) and incubated at 37°C for 3 hours. The lysate was extracted once with phenol and once with chloroform and treated with RNase A (20 µg/ml) for 15 minutes at 37°C. The DNA was subsequently extracted twice with phenol-chloroform and precipitated with ethanol. One µg of DNA was digested with Sau3AI overnight, ethanol precipitated, and self-ligated in a 100 µl-reaction overnight. The ligated DNA was precipitated with ethanol and resuspended in 20 µl. Of this, 1 µl was used for PCR using previously described primers [25]. The resulting DNA fragments were cloned into pGEM-T and sequenced using M13 forward and M13 reverse primers.
Parasites were cloned by limiting dilution. Parasitemia of cultures was determined by Giemsa staining and parasites diluted to 3 parasites/ml in cRPMI. Of this dilution, 100 µl was added to one well of a 96-well plate, and 10 µl of 50% human erythrocytes and 90 µl of cRPMI were added. Medium was changed and erythrocytes added every four days. Parasites were detected approximately 18 days after dilution.
Parasites were prepared for microscopy as described [21]. Briefly, ∼3×108 erythrocytes infected at a parasitemia of 1–10% were spun down and resuspended in 1 ml PBS with 11 mM glucose containing 10 µg/ml Hoechst 33342 and left at room temperature for five minutes. The cells were washed twice with 3 ml PBS-glucose and resuspended to a parasitemia of ∼30%. A 3 µl aliquot was placed on a slide and covered with a coverslip, which was sealed with nail polish. The sample was viewed on an Olympus IX inverted fluorescence microscope and images were collected with a Photometrix cooled CCD camera (CH350/LCCD) driven by DeltaVision software from Applied Precision Inc. (Seattle, WA).
For immunofluorescence staining, we followed the protocol of Apodaca et al. [40]. Briefly, cells were deposited on poly-L-lysine coated coverslips for thirty minutes at 37°C in PBS containing 11 mM glucose. Cells were washed twice with PBS and fixed with 1% formaldehyde in PBS for ten minutes, then quenched with 50 mM ammonium chloride in PBS for ten minutes. Then the cells were washed, permeabilized with 0.05% saponin in PBS and blocked with 0.7% fish skin gelatin, 0.01% saponin in PBS (FSP) for thirty minutes at 37°C. Primary antibody against SBP1 (LWL) and GFP diluted in FSP were added and the cells were incubated for one hour at 37°C. Cells were washed three times with FSP before addition of secondary antibody (goat anti-mouse TRITC and goat anti-rabbit FITC) and incubation at 37°C for one hour. Cells were then washed three times with FSP, twice with PBS, once with PBS containing 0.01% saponin, once with PBS, once with PBS containing 0.1% Triton X-100 and once with PBS. Cells were then fixed again with 4% formaldehyde in 0.1 M sodium cacodylate for thirty minutes. The cells were washed once with PBS, then stained with 10 µg/ml Hoechst 33342 in PBS for five minutes. The cells were washed two more times with PBS before being mounted on glass slides on a drop of DABCO and sealed with nail polish. Slides were stored at −20°C.
Parasites were harvested from erythrocytes with 0.05% saponin as described above and frozen at −80°C. In some cases the parasites were resuspended directly in SDS-PAGE loading dye, boiled, and an equivalent of 1–5×107 parasites was loaded on an SDS-PAGE gel after boiling. In other cases the parasites were resuspended in 100 µl RIPA buffer, incubated on ice in the presence of protease inhibitors. An equivalent volume of SDS-PAGE loading dye was added, the samples were boiled and the samples run as above. Proteins were transferred to nitrocellulose. The resulting blot was blocked with 5% milk in PBS with 0.05% Tween-20 (PBS-T) for one hour at room temperature or at 4°C overnight, and incubated with anti-GFP antibody (Santa Cruz, Molecular Probes) at room temperature for 1 hour or 4°C overnight, followed by incubation at room temperature in the presence of HRP-linked secondary antibody. Protein was visualized using the ECL system (SuperSignal West Pico chemiluminescence, Pierce).
For labeling with Rhodamine B, ∼2×107 cells were pelleted and resuspended in PBS containing 11 mM glucose and 1 mM Rhodamine B and 10 µg/ml Hoechst 33342. Subsequently cells were incubated at 37°C. Cells were washed five times with PBS-glucose and viewed. |
10.1371/journal.pbio.0060219 | LTP Promotes a Selective Long-Term Stabilization and Clustering of Dendritic Spines | Dendritic spines are the main postsynaptic site of excitatory contacts between neurons in the central nervous system. On cortical neurons, spines undergo a continuous turnover regulated by development and sensory activity. However, the functional implications of this synaptic remodeling for network properties remain currently unknown. Using repetitive confocal imaging on hippocampal organotypic cultures, we find that learning-related patterns of activity that induce long-term potentiation act as a selection mechanism for the stabilization and localization of spines. Through a lasting N-methyl-D-aspartate receptor and protein synthesis–dependent increase in protrusion growth and turnover, induction of plasticity promotes a pruning and replacement of nonactivated spines by new ones together with a selective stabilization of activated synapses. Furthermore, most newly formed spines preferentially grow in close proximity to activated synapses and become functional within 24 h, leading to a clustering of functional synapses. Our results indicate that synaptic remodeling associated with induction of long-term potentiation favors the selection of inputs showing spatiotemporal interactions on a given neuron.
| In the central nervous system, excitatory contacts between neurons occur mainly on postsynaptic protrusions called dendritic spines. For decades, these structures have been considered static, and the adaptive properties of neuronal networks were thought to be only due to changes in the strength of neuronal connections. But recently, new imaging techniques used on living neurons revealed that spines and synapses are dynamic structures that undergo continuous turnover and can be formed or eliminated as a function of activity. The functional consequences of this structural remodeling, however, were still unknown. This work shows that application of learning related paradigms (such as induction of long-term potentiation or rhythmic activity) to hippocampal neurons allows them to operate a selection of synaptic inputs that show coincident activity. This is done through a competitive mechanism that promotes a selective stabilization of synapses activated by the learning paradigm and a replacement of non-activated inputs by new spines. Furthermore these new dendritic spines preferentially grow in close proximity to activated synapses and become functional. These findings provide evidence that learning related paradigms play a major role in shaping the structural organization of synaptic networks by promoting their specificity.
| Integration of synaptic signals during learning processes is critical to the function of cortical networks. This processing is achieved through various mechanisms that involve generation of coincident rhythmic activity, induction of properties of plasticity such as long-term potentiation (LTP), but also growth of new protrusions and remodeling of synaptic networks [1–5]. The precise functional contribution of this structural remodeling to network properties remains unclear. In vitro experiments have demonstrated that LTP induction results during the next few hours in the growth of new filopodia and spines [6–9] which then rapidly become functional [10] and show all characteristics of morphologically mature synapses over the course of 24 h [11]. Also, work by several laboratories has shown that under in vivo conditions, spines and varicosities undergo a continuous turnover and replacement that vary in intensity as a function of development [12–16]. This process is further regulated by sensory activity, because under conditions of deprivation such as whisker trimming [17] or unbalanced activity such as chessboard whisker trimming [12,18], spine turnover increases, new spines form synapses and become stabilized, and others are eliminated. These experiments therefore clearly demonstrated that stable synaptic contacts can be removed or created de novo through experience, raising the possibility that synapse remodeling, together with Hebbian forms of plasticity, could contribute to information processing and learning [3,19]. It remains unclear, however, whether and how sensory activity regulates this synaptic remodeling and whether it could actually affect signal integration by the neuron and/or the network. Also, the rules and mechanisms determining which synapse should be removed or restructured and where new synapses should be created are unknown. These are important issues because both the number and localization of spines may greatly affect the properties of integration of synaptic responses by a neuron. Recent studies have shown that spatiotemporal clustering of synaptic currents on small or remote dendrites represents a critical aspect for the expression of plasticity and the contribution to neuronal firing [20–22]. Identification of the mechanisms that underlie spine and synapse remodeling is therefore critical to a better understanding of the processing properties of synaptic networks. We investigated these issues, using a repetitive imaging approach applied to hippocampal slice cultures, and analyzed how precisely learning-related activity patterns affected the long-term behavior of identified spines.
Hippocampal slice cultures were transfected to express enhanced green fluorescent protein (EGFP) using a biolistic approach; we then monitored the behavior of identified protrusions (spines and filopodia) over several days following induction of learning-related activity patterns (Figure 1). For this, we used two different conditions that trigger LTP, a property believed to underlie learning mechanisms: first, we applied theta burst stimulation (TBS) to Schaffer collaterals, which triggers robust LTP, and second, we treated slice cultures for 20–60 min with carbachol (Cch, 10 μM), a cholinergic agonist, which, in the hippocampus and in slice cultures, triggers rhythmic activity in the theta and gamma range and induces a lasting synaptic enhancement (Figure 2C, inserts) [1,23]. In humans, these theta activities have been directly implicated in memory processes [24]. Fluorescent cells and dendritic segments were then imaged repetitively and the changes in protrusion number and long-term spine stability monitored (Figure 1A–1C) through analysis of single z-stack images (Figure 1D and 1E; see criteria in Materials and Methods). Control experiments with propidium iodide staining showed that transfection and repetitive confocal imaging of slice cultures did not alter cell viability over periods of weeks.
Analysis of protrusion turnover over periods of 3–8 d showed that the dynamics of synaptic networks is high at this developmental stage (11 d in vitro) with an average of 20.3% ± 1.1% new protrusions formed per 24 h and 20.8 ± 0.9% disappearing within the same period of time (Figure 2A and 2B). The other protrusions either remained stable without changes or underwent some sort of morphological transformations (16.2% ± 0.3% [25]). These values are in the range of those reported in vivo in the cortex of very young mice [12,14,15].
Following theta burst activity we found that this basal turnover rate markedly increased. The effect was not short-lived [7,9], but the increase lasted for several days following a brief stimulation episode. This lasting increase in turnover rate was observed both following LTP induction by TBS (Figure 2C) and by Cch-induced rhythmic activity (10 μM; Figure 2D). The insert in Figure 2C shows the potentiation of the slope of evoked excitatory postsynaptic potentials (EPSPs) recorded in slice cultures following TBS. In Figure 2D, we illustrate the spontaneous baseline activity of 1–3 Hz observed under control conditions, the increased 5–10 Hz field activity recorded in the stratum pyramidale of CA3 during application of 10 μM Cch and the synaptic enhancement observed in the cornus ammonis 1 (CA1) region. The proportion of new and lost protrusions, which includes spines and filopodia, increased markedly under both conditions to values of 34% ± 6% and 43% ± 6% of new protrusions and 33% ± 2% and 44% ± 3% of lost protrusions over the first 24 h for TBS and Cch, respectively (see also Figure 3A). These changes reflected a similar increase in the formation of thin spines and filopodia, filopodia representing only a very small fraction of the new protrusions both under control conditions and after stimulation (4.3% ± 0.9%, n = 30 cells, control; 4.7% ± 1.4%, n = 17, LTP and 3.4% ± 1.6%, n = 17, Cch). Together, these experiments indicate a 70% and 115% increase in protrusion turnover rate following TBS or Cch treatment, respectively. To allow comparisons, the data obtained at the different observation times are expressed in Figure 2C and 2D as percentage of the basal rate of protrusion formation or loss observed under control condition. To test for the specificity of the effect, we then carried out the same experiments, but applied the N-methyl-D-aspartate (NMDA) receptor antagonist D(−)-2-amino-5-phosphonopentanoic acid (D-AP5; 100 μM) during the stimulation protocol or during the application of Cch. As shown in Figure 2C and 2D, D-AP5 specifically prevented the lasting increase in protrusion turnover under both conditions. As an additional control, we also analyzed hippocampal slice cultures stimulated in the same way at low frequency (0.3 Hz), but without induction of rhythmic activity. These controls showed no significant changes in turnover rate over time. Finally, we also tested whether this increase in protrusion turnover was dependent upon protein synthesis. For this, slice cultures were incubated in the presence of 25 μM anisomycin (Ani) and stimulated with either TBS or Cch. Under these conditions, both forms of potentiation were prevented (ratio of potentiation at 60 min: 1.13 ± 0.2, n = 6 and 1.08 ± 0.11, n = 3 for TBS and Cch, respectively) and, as shown in Figure 3A, no significant increase in the rate of protrusion formation or loss could be observed over the next 24 h. Note also that Ani treatment of cultures for 5 h without TBS or Cch stimulation did not affect the rate of formation and loss of protrusions over 24 h. These results thus indicated that the changes in protrusion turnover associated with induction of LTP lasted several days and included formation and elimination of spines and filopodia.
To assess these results further and test for possible changes in spine stability and/or occurrence of populations of transient spines or filopodia, we next analyzed protrusion growth each day over a period of 5 h, a period during which most new events can be detected [25]. Following LTP induction by TBS, the rate of protrusion formation expressed per 5 h and per 100 μm of dendritic segment increased by a factor of 2, and this for several days, an effect fully prevented by D-AP5 applied during the stimulation protocol (Figure 3B and 3C). We then also assessed spine stability, restricting the analysis to spines, since filopodia are essentially transient [25] and mostly disappeared within 24 h. The stability of pre-existing spines, calculated as the proportion of spines still present on consecutive days, significantly decreased following LTP induction (Figure 4A), a change also dependent upon NMDA receptor activation. The stability of the new spines formed within the first 5 h following LTP induction was however not affected (Figure 4B) and remained particularly low as under control conditions. Thus, LTP induction promoted protrusion growth, but also destabilization of pre-existing spines. Altogether, these different effects approximately cancelled each other, so that the protrusion density did not greatly vary; actually, a significant increase was only observed transiently 2 d following LTP induction (Figure 4C). A similar situation was observed following Cch treatment. Protrusion growth increased in association with a decrease in stability of pre-existing spines and no effect on the process of new spine stabilization or on protrusion density (Figure 4D–4F). With both types of experiments, therefore, the net effect on several days of this increased turnover was to promote the replacement of existing spines by new ones.
We then wondered how this increased spine remodeling could contribute to the specificity of the synaptic network and thus investigated whether it affected similarly activated and naive synapses. For this, we transfected pyramidal neurons with the red fluorescent dye monomeric red fluorescent protein (mRFP) [26], to visualize the structural changes in spine morphology, and costained them 3 d later with Fluo-4 AM, a calcium indicator, to identify spines activated by single pulse and TBS stimulation protocols (Figure 5A–5C, see also Material and Methods). Figure 5 illustrates the example of a dendritic segment with one spine that showed a clear increase in calcium fluorescence upon stimulation, while another one on the same segment remained silent. In all experiments carried out, we verified that spines activated by stimulation were always surrounded by other silent, nonactivated spines in order to exclude global activation effects. Also, we checked that analyses were done on spines of similar size (see Figure 6) and that the maximum calcium signal perfectly coincided with the center of the spine head. We then assessed the stability of activated and nonactivated spines for the next 3 d. Overall, with the stimulation pulses used under these conditions, on average, 36% of all spines tested on analyzed dendritic segments were found to be activated (n = 349 spines, 18 cells or segments). TBS was then applied to the same synapses using the same stimulation pulses in ten cells (62 activated and 130 nonactivated spines analyzed), which resulted in a differential effect on spine stability: activated spines showed a striking increase in stability in comparison to nonactivated spines present on the same dendritic portions (Figure 5D; p < 0.001). Nonactivated spines actually underwent pruning with regard to spines in nonstimulated slice cultures (Figure 5E; p < 0.05). Interestingly, this differential stabilization was prevented by D-AP5 applied during TBS (Figure 5E). We also verified that simple activation of spines without TBS did not affect the long-term stability of spines (Figure 5E, squares).
Although for technical reasons we could not directly assess LTP in these stimulated spines, we found that most of them exhibited an enlargement of their head over the next 5 h. Several previous studies have indeed reported an enlargement of the spine head as a consequence of LTP induction [27–29] or used this criteria for identifying potentiated synapses [30]. In the group of 272 activated and nonactivated spines analyzed before TBS, there was no difference in mean head width (Figure 6A). However, when analyzed 5 h after TBS, most activated spines now exhibited an enlargement of their head, an effect not observed with nonactivated spines (Figure 6B). Interestingly, we also found that this differential enlargement was transient, as most activated spines reversed their size after 24 h and the differences with nonactivated spines then became nonsignificant (Figure 6C). Note, in addition, that the head width of nonactivated spines tended to become smaller after TBS and that the size of spine heads, when analyzed individually, showed regular fluctuations over consecutive days for both activated and nonactivated spines. A robust effect, however, was the close correlation observed between activated spines, spines that showed an enlargement 5 h after stimulation, and spines that became stabilized by activity. When using spine enlargement as a criteria to analyze spine stability, we found, as for activity, that enlarging spines exhibited the same differential stabilization (Figure 6D). Thus LTP induction is very likely to promote a long-term stabilization of potentiated synapses.
To verify whether Cch-induced rhythmic activity also produced the same selective stabilization process, we then analyzed how Cch treatment affected spine size. Analysis of 218 spines taken from nine dendritic segments showed that 34% of them exhibited enlargement of their head 5 h after Cch treatment. We then tested the stability of these spines over the next 2 d. As shown in Figure 6E, spines that enlarged as a result of Cch-induced rhythmic activity also became significantly more stable, while nonenlarging spines tended to be eliminated, showing the same differential behavior as after TBS-induced potentiation.
We then asked how these mechanisms could affect spine organization and distribution and analyzed whether newly formed spines could appear at specific hot spots. As shown in Figure 7A and 7B, we found that, indeed, newly formed spines tended to appear in close proximity to activated spines. In Figure 7C, we analyzed the proportion of activated versus nonactivated spines that had a new protrusion formed within a distance of 1.5 μm in the next 48 h (defined as hot spot). As indicated, almost half of activated spines had a new spine growing close by, something that did not occur with nonactivated ones. As shown by Figure 7D, we then examined all newly formed spines and asked how many actually grew close to an activated or a nonactivated spine. The results show that, again, about half of newly formed spines grew less than 1.5 μm from an activated spine, while only a small number of them grew close to a nonactivated spine, the others growing close to spines that could not be determined. The overall stability of newly formed spines was, however, not dependent on their localization (Figure 7E), because new spines generated close to or far from an activated spines showed the same probability of being present on subsequent days.
We then tested whether these newly formed spines became functional. For this, TBS was applied to an mRFP-transfected neuron, and the new spines formed within the next 24 h monitored by repetitive imaging and their functionality tested through loading with Fluo-4 AM and stimulation trials of Schaffer collaterals. Figure 8A shows an example of such a newly formed spine. Line scan analysis performed 24 h after TBS shows that this newly formed spine did indeed respond to stimulation through a calcium signal (Figure 8B and 8C), indicating that it was functional. Similar results were obtained in 30 spines out of 47 analyzed (n = 5 cells), indicating that a majority of them were functional. The mean ΔF/F0 ratio (i.e., [fluorescence − basal fluorescence]/basal fluorescence) at the peak of the calcium signal recorded in these experiments was 4.3 ± 0.8 (n = 30). For the other spines, it remains unclear whether they were silent or whether we simply could not activate them. We then asked whether the new functional synapses were also likely to be more stable than those that did not exhibit any calcium signal in response to stimulation. Of the 47 newly formed spines analyzed here, we found that the probability to persist for 48 h was 82% ± 12% for the 30 functional spines (n = 5), but only 30% ± 10% for the 17 nonactivated spines (Figure 8D), indicating that activity is a major criteria for long-term stability. Together these results indicate that LTP induction favored a clustering of new functional spines around activated spines, promoting in this way possibilities of spatiotemporal interactions between them.
Together, these experiments provide evidence for an important new functional role of LTP-inducing activity in promoting a refinement of synaptic networks. Previous work in hippocampal slice cultures has shown that LTP induction is associated with two major types of structural remodeling. First, within minutes, potentiated synapses become larger and express larger and more complex postsynaptic densities [27–30], a change possibly associated with receptor expression and/or spine stabilization [31]. Second, within minutes to hours, LTP induction also results in the growth of new filopodia and spines [7,9,32], which then eventually become functional synapses [8,10,11,25]. These in vitro data are consistent with other in vivo experiments indicating that sensory deprivation or unbalanced activity does indeed affect cortical spine turnover and promote formation of new synapses [17,18,33]. Here we add three new pieces of information providing a novel, important function for structural plasticity: namely, to operate as a selection process for the long-term stability of synaptic contacts and the promotion of spatiotemporal interactions between spines.
First, we provide the first (to our knowledge) direct evidence that spines stimulated with LTP-inducing protocols are selectively stabilized over periods of several days. Although LTP could not be directly assessed together with repetitive imaging, we find that stabilization occurred specifically at spines stimulated with TBS and not at nonstimulated spines. Also, stabilized spines did exhibit an enlargement of the head at 5 h, a characteristic now demonstrated to be directly associated to LTP by several recent studies [27–30]. Finally, spine enlargement and spine stabilization were both D-AP5 sensitive and protein synthesis dependent. It seems therefore likely that the stabilization of stimulated synapses revealed here represents a central mechanism for the persistence of potentiated synapses.
The second new feature uncovered by these experiments is that LTP is not only associated with a short-term increase in protrusion growth, but a lasting, enhanced turnover that affects pre-existing spine stability, probably through competition mechanisms. Consistent with previous data [7,9], protrusion growth initially tended to predominate over spine loss, leading to a transient increase in spine or protrusion density. However, all together, LTP mainly affected turnover, resulting not only in protrusion growth, but also in an increased loss and destabilization of spines, which, importantly, specifically affected nonstimulated spines. The net effect of LTP over several days was therefore to promote the replacement of nonactivated spines by new ones. This selective destabilization of nonactivated spines was quantitatively significant, because in these experiments more than 10% of the spines of the neurons were actually replaced. Accordingly, regular occurrence of activity susceptible to induce LTP works as a selection mechanism leading to a progressive stabilization of inputs showing coincident activity, increasing in this way the coherence of the synaptic information provided to the neuron and reducing background noise.
The last important finding of these experiments is that newly formed protrusions do not appear just anywhere, but tend to cluster around activated spines. These new spines also become functional, and when functional, tend to remain stable. Together with the evidence that LTP induction is facilitated between spines located close to each other [30], this result indicates that LTP will actually promote the creation of hot spots of functional synapses. This provides therefore a means to promote spatiotemporal clustering of synaptic signals, a property recently shown to be critical for determining the characteristics of plasticity and processing at synapses on small or remote dendrites [20–22].
At the molecular level, an interesting implication of these results is that LTP mechanisms are likely to involve specific changes that could directly affect spine stability. Spine enlargement has been previously proposed to reflect this process [31] and, consistent with this idea, we indeed found that activated spines did enlarge 5 h after stimulation. Curiously, however, this effect did not seem to remain stable over 24 h, and analyses of spine head width suggest that most spines regularly exhibit significant variations of their size [30]. It could be, therefore, that stability is not only reflected in the size of the spine, but is linked to the expression of specific molecules. The current evidence indicating a contribution of protein synthesis to the long-term changes in synaptic strength and to the regulation of spine turnover as reported here could actually suggest such a mechanism [34]. In order to become stable, activated spines would need to accumulate the machinery required for protein synthesis [35] and/or express specific molecules conferring stability to the synaptic contact.
Taken together, the mechanisms reported here provide a new framework for understanding how the specificity of cortical networks may progressively develop. These results might be particularly important during critical periods when refinement of connections represents a major process shaped by rhythmic activity and dynamic regulations between excitatory and inhibitory transmission [36]. This network plasticity might, however, also contribute in the adult and provide the functional rules underlying the spine dynamics described in association with sensory activity [18] or following brain damage [37]. Together the synaptic mechanisms described here certainly point to the important role played by structural plasticity in association to Hebbian changes in synaptic strength for the refinement and specificity of cortical networks.
Transverse hippocampal organotypic slice cultures (400 μm thick) from 6- to 7-d-old rats were prepared as described [38] using a protocol approved by the Geneva Veterinarian Office (authorization 31.1.1007/3129/0) and maintained for 11–18 d in a CO2 incubator at 33 °C. Transfection was done either with a pc-DNA3.1-EGFP or a pCX-mRFP1 [39] plasmid using a biolistic method (Helios Gene Gun, Bio-Rad) 2–3 d before the first observation. Fluorescence usually started to be expressed after 24–48 h and then remained stable for at least 15 d.
For electrophysiological recordings, slice cultures were maintained at 32 °C in an interface chamber under continuous perfusion as described [40]. EPSPs were evoked by stimulation of a group of Schaffer collaterals and recorded in the stratum radiatum of the CA1 region with pipettes filled with medium. Potentiation was analyzed by measuring EPSP slopes expressed as percent of baseline values using an acquisition program written with Labview. LTP was induced by TBS (five trains at 5 Hz composed each of four pulses at 100 Hz, repeated twice at 10-s intervals). As controls, we used slice cultures stimulated at low frequency (0.3 Hz) and recorded in the same manner as well as slice cultures stimulated with TBS but in the presence of 100 μM D-AP5. In these experiments, D-AP5 was only applied for 30 min during application of TBS. Cch treatment was applied for 20–60 min at a concentration of 10 μM with or without concomitant application of D-AP5. The protein synthesis inhibitor was Ani applied 1 h before TBS or Cch treatment at a concentration of 25 μM and then maintained for 3 h.
Short imaging sessions (10–15 min) of transfected slices were carried out with an Olympus Fluoview 300 system coupled to a single (Olympus) and a two-photon laser (Chameleon; Coherent) as described [25]. Laser intensity in all these experiments was kept at the minimum and acquisition conditions maintained mostly unchanged over the different days of observation. Control experiments showed that transfection and repetitive confocal imaging of slice cultures did not alter cell viability over periods of weeks.
We focused on dendritic segments of about 35 μm in length and located between 100 and 300 μm from the soma on secondary or tertiary dendrites using a 40× objective and a 10× additional zoom (final resolution: 25 pixels per micron; steps between scans: 0.4 μm; Figure 1). We did not find differences in protrusion turnover within the limits of these dendritic locations. For calcium imaging of spine activity, transfected cells were additionally loaded with the cell-permeable calcium indicator Fluo-4 AM (F-14201, Invitrogen). For this, 50 μg of Fluo-4 AM was dissolved in 10 μl Pluronic (F-127, Invitrogen) and then diluted in 90 μl of standard pipette solution (150 mM NaCl, 2.5 mM KCl, 10 mM Hepes) for a final dye concentration of 500 μM. A standard patch pipette was then filled with 10 μl of dye solution and placed at a distance of about 10 μm from the soma of a mRFP1-expressing CA1 pyramidal cell. Dye was ejected by short pulses of pressured air at a frequency of three per minute during one-half hour. Calcium transients in 10–26 identified spines per dendritic segment were then recorded using line scans through the spine heads obtained during application of stimulation pulses to Schaffer collaterals. These pulses were of identical intensity and duration to those used for subsequent induction of LTP. Confocal aperture was set to the minimum during line scans, and matching with the mRFP fluorescence in the red channel was systematically checked. For each spine tested, calcium transients evoked by two or three consecutive stimulation pulses were recorded, and spines were determined as activated whenever the fluorescence signal increased by more than 20% over background in any of the recordings. In average, 36% of all spines tested corresponded to these criteria with the stimulation pulses used. To avoid biases, we then also verified that the size distribution of the spine heads did not differ between spines classified as activated and nonactivated (0.56 ± 0.02 μm versus 0.58 ± 0.02 μm, respectively; Figure 6A).
In this study we refer to protrusions, whenever analyses were carried out by considering filopodia and spines. Filopodia were defined as protrusions devoid of enlargement at the tip, while we classified as spines all protrusions exhibiting an enlargement at the tip. All turnover and stability analyses were carried out on single z-stacks of raw images (Figures 1E and S1) using a plug-in specifically developed for OsiriX software (http://www.osirix-viewer.com). The measures of turnover were carried out by analyzing all protrusions, i.e., filopodia and spines. We counted as new protrusions all new structures (spines or filopodia) appearing between two observations (5 or 24 h) and characterized by a length of >0.4 μm. All filopodia were counted as separate protrusions. We also counted spines located behind each other on z-stacks whenever distinction was possible (Figures 1E and S1). For disappearances, we counted all protrusions (spines and filopodia) that could no longer be identified on the next observation. Dubious situations due to possible changes in protrusion shape, size, or orientiation were discarded, but overall accounted for only a small number of cases (less than 1%). To further ensure reliability of analyses, all measurements of spine turnover and stability were carried out blind by two experimenters. Comparisons of the analyses made in this way showed variations in the results that were less than 3%. Furthermore, we used high numbers of n for both cells and spines, and labeled all new or lost protrusions directly on the raw data (Figure 1E) to allow multiple checks.
Due to the lack of survival of filopodia on several days, stability analyses carried on 48 or 72 h periods only included pre-existing spines, i.e., spines present at the beginning of the experiment. For analyses of spine width, we measured the maximum diameter of the spine head on individual z-images, setting the fluorescence level on the levels obtained in the dendrite. Situations that did not allow a precise spine head width measurement (two spine heads overlapping each other on the same z sections) were excluded. Calcium fluorescence intensities were acquired and analyzed with Fluoview software (FV300, Olympus). Note that for illustration purposes, images presented in the figures are maximum intensity projections of z stacks, further treated with a Gaussian blur filter. All statistics are given with the standard error of the mean. Normality was tested for each distribution (D'Agostino and Pearson test), and α was set to 5% for all tests.
|
10.1371/journal.pntd.0006810 | Spatial distribution of Taenia solium exposure in humans and pigs in the Central Highlands of Vietnam | Taenia solium, a pork-borne parasitic zoonosis, is the cause of taeniasis and cysticercosis in humans. In Vietnam, poor sanitation, the practice of outdoor defecation and consumption of raw/undercooked pork have been associated with infection/exposure to T. solium in both humans and pigs. The broad-scale geographic distribution of the prevalence of T. solium varies throughout the country with infection restricted to isolated foci in the north and a more sporadic geographic distribution in the Central Highlands and the south. While cross-sectional studies have allowed the broad-scale geographic distribution of T. solium to be described, details of the geographic distribution of T. solium at finer spatial scales have not been described in detail. This study provides a descriptive spatial analysis of T. solium exposure in humans and pigs and T. solium taeniasis in humans within individual households in village communities of Dak Lak in the Central Highlands of Vietnam.
We used Ripley’s K-function to describe spatial dependence in T. solium exposure positive and negative human and pig households and T. solium taeniasis exposure positive and negative households in villages within the districts of Buon Don, Krong Nang and M’Drak of Dak Lak province in the Central Highlands of Vietnam. The prevalence of exposure to T. solium in pigs in Dak Lak province was 9 (95% CI 5 to 17) cases per 1000 pigs at risk. The prevalence of exposure to the parasite in humans was somewhat higher at 5 (95% CI 3 to 8) cases per 100 individuals at risk. Spatial aggregations of T. solium exposure-positive pig and human households occurred in some, but not all of the villages in the three study districts. Human exposure-positive households were found to be aggregated within a distance of 200 to 300 m in villages in Krong Nang district compared with distances of up to 1500 m for pig exposure-positive households in villages in M’Drak district. Although this study demonstrated the aggregation of households in which either T. solium exposure- or taeniasis-positive individuals were present, we were unable to identify an association between the two due to the very low number of T. solium taeniasis-positive households.
Spatial aggregations of T. solium exposure-positive pig and human households occurred in some, but not all of the villages in the three study districts. We were unable to definitively identify reasons for these findings but speculate that they were due to a combination of demographic, anthropological and micro-environmental factors. To more definitively identify characteristics that increase cysticercosis risk we propose that cross-sectional studies similar in design to that described in this paper should be applied in other provinces of Vietnam.
| Taenia solium is a pork-bone zoonotic parasite. Humans acquire taeniasis from consumption of raw/undercooked pork contaminated with T. solium cysticerci. Pigs and humans acquire cysticercosis following consumption of food contaminated with eggs shed from the feces of humans with T. solium taeniasis. In Vietnam, the geographic distribution of T. solium varies throughout the country with hotspots or foci of infection in communities in the North and a more sporadic distribution in the Central Highlands and the South. While information on the distribution at the regional and provincial level is available, there is no available information on the spatial distribution of T. solium at fine spatial scales and factors influencing its distribution at fine spatial scales have not been described in detail. In this cross-sectional study, we collected information on the geographic coordinates of study households and utilized spatial analytical techniques to quantify both the fine scale spatial pattern of exposure to T. solium as well as the tendency for T. solium exposure-positive households to be located close to other T. solium exposure-positive households (spatial autocorrelation) in three districts in Dak Lak province. We found that in some of the study villages T. solium exposure-positive households were more likely to be surrounded by other T. solium exposure-positive households. Human exposure-positive households were found to be aggregated within a distance of 200 to 300 m in villages in Krong Nang district; whilst spatial aggregation of pig exposure-positive households was found up to distances of 1500 m in villages in M’Drak district. Although households that had either T. solium exposure- or taeniasis-positive cases were aggregated, we were unable to quantify their spatial association due to the extremely low number of T. solium taeniasis-positive households. This study shows that in the Central Highlands of Vietnam, T. solium exposure tend to cluster within foci. This information can be used to inform community intervention programs to lower its incidence in both humans and pigs.
| Taenia solium is a pork-borne zoonosis of major public health and economic importance. The parasite causes cysticercosis/neurocysticercosis in humans and pigs in many low-income communities in Latin America, Africa and Asia [1]. Poor sanitation, allowing pigs to free roam, the practice of outdoor defecation and consumption of raw/undercooked pork are risk factors for T. solium infection. Cysticercosis infection in humans and pigs occurs due to the accidental ingestion of T. solium eggs shed through the feces of humans with T. solium taeniasis (tapeworm carriers). Taeniasis occurs when humans consume raw/undercooked pork with T. solium cysticerci.
T. solium infection results in not only an economic burden in low-income communities due to loss of productivity in affected individuals and the cost of treatment, but also losses arising from the condemnation of pig carcasses destined for human consumption. It was estimated that approximately US $185 million and 2.1 million disability-adjusted life years (DALYs) were lost in north India due to human cysticercosis in 2011 [2]. It was estimated that the number of DALYs for human cysticercosis in Mozambique in 2007 [3] was 6.0 per thousand person-years, 0.2 in 2011 in Mexico [4] and 0.7 in 2012 in Tanzania [5]. The pork industry in four provinces of Laos estimated losses of between US $55,000 to 96,000 arising from 20% of carcasses identified as cysticerci-positive over a 21 month period [6]. In Latin America, the number of neurocysticercosis infections has been estimated to be between 11 and 29 million with 1.3 million individuals suffering from neurocysticercosis-related epilepsy [7]. Globally, cysticercosis was estimated to be the cause of over 28,000 deaths in 2010 [8].
Although risk factors for T. solium taeniasis and cysticercosis include outdoor defecation, the consumption of raw/undercooked pork and allowing pigs to free roam, transmission patterns and the prevalence of the parasite can vary considerably and may prove inconsistent within and/or between regions and communities [9] thus impeding efforts to achieve parasite eradication. Because of the negative impact of T. solium on human health and the economy [10], controlling the disease is a priority. Understanding the spatial distribution of T. solium is one important step towards development of effective control strategies. Studies in communities of Latin America and Africa, where T. solium infection is hyperendemic, identified clustering of T. solium cysticercosis at both the household and community level [11–13] and a strong geographical association between T. solium carriers and cysticercosis in humans and pigs [14,15]. Spatial analyses were used to define an appropriate radius for a control area in a community where T. solium was hyperendemic in Peru. Targeting interventions within this control area was effective in reducing the number of T. solium carriers and the sero-incidence of porcine cysticercosis [16].
T. solium is endemic in Vietnam. A systematic review of cross-sectional studies carried out in Vietnam between 1999 to 2011 showed that the prevalence ranged from 0 to 130 T. solium positive individuals per 1000 individuals at risk [17]. The distribution of T. solium in Vietnam is characterized by hotspots or foci of infection in communities in the northern provinces, including Phu Tho and Bac Ninh [18,19]. In the Central Highlands and the south of the country the distribution of T. solium is sporadic [20]. While previous studies [15,16,17] have provided useful information at the regional and provincial level, we are aware of no investigations that have investigated the distribution of T. solium at finer spatial scales. With this background, the aim of this study was to describe the spatial distribution of households in which one or more individual humans or pigs were T. solium exposure positive and households in which one or more individual (humans) were T. solium taeniasis positive. Quantitative estimates of the prevalence and geographic distribution of T. solium exposure and T. solium taeniasis positive households will provide evidence to allow public health authorities to decide between treatment programs applied at the whole community level as opposed to treatment programs applied at either the individual household or small area level.
This study was reviewed and approved by the Behavioral and Social Sciences Human Ethics Sub-committee, the University of Melbourne (reference number 1443512) and the Animal Ethics and Scientific Committee, Tay Nguyen University (reference number 50.KCNTY). This study was conducted under the supervision of the local Center for Public Health and the local Center for Animal Health, Dak Lak, Vietnam. This research on pigs was based on the International Guiding Principle for Biomedical Research Involving Animals issued by the Council for the International Organization of Medical Sciences.
The cross-sectional study was carried out between May and October 2015 in Dak Lak province in the Central Highlands of Vietnam. Dak Lak is comprised of 15 districts with approximately 70% of the total population of 1.8 million people living in rural areas [21]. Within the province, three districts namely Buon Don, Krong Nang and M’Drak were chosen as the study sites based on their diverse geographic characteristics (Fig 1). The characteristics of these districts have been described in detail elsewhere [22]. A sampling frame listing the name of all villages in Dak Lak province was obtained from Sub-Department of Animal Health office within the Ministry of Agriculture and Rural Development. Villages eligible for sampling comprised those with more than 1000 pigs, as recorded by the Sub-Department of Animal Health. All eligible villages within each district were assigned a number and two numbers chosen at random to select villages from each district for inclusion in the study. A list of householder names within each selected village was obtained from each village head person, and householder names were assigned a numeric code. A sheet of paper listing numeric household codes for each village were cut into pieces and placed face-down on a table. The village head person was asked to select 50 households at random for human sampling and between 100 and 140 households for pig sampling. All households selected for human sampling and pig sampling were visited several days before the proposed sampling date to obtain consent from householders to take part in the study. Householders eligible for inclusion in the study were individuals who were healthy, not pregnant and over seven years of age. Householders requested to take part in the study signed a consent form. Those that were under 18 years of age were required to provide written consent as well as written consent from either their parents or legal guardians. At the time of consent each study participant was given a labeled stool container, with instructions that the container would be collected on the date of sampling, several days later. Sampling of households (Fig 2) was carried out in two stages. Humans from each of the 50 consenting study households were sampled between May and October 2015; pigs from each of the 100 to 150 consenting study households were sampled between June and October 2015. At the time of each household visit, a questionnaire was administered to each of the study participants soliciting details regarding demography, sanitation and hygiene status, food culture and religion, practice of pig management and the longitude and latitude coordinates of the main doorway of entry of the dwelling used for sleeping.
On the date of sampling, consenting householders were visited by staff from the Sub-Department of Health and 5 mL of venous blood collected into plain clotting tubes from consenting study participants. Stool samples were fixed in 5% potassium dichromate (w/v) for molecular analysis.
Pigs that were pregnant, ill or aged less than 2 months of age were excluded from sampling. Approximately, 10 mL of blood was obtained from the cranial vena cava of each pig into plain blood collection tubes. Blood samples were allowed to clot at ambient temperature prior to centrifugation at 3200 × g for 5 minutes to collect serum. Serum was dispensed into 1.5 mL aliquots and stored at -20°C until analysis.
The longitude and latitude of the main doorway of entry of sampled households was recorded using a handheld global positioning system (GPS) device (Garmin GPSMAP64, Taiwan).
Human stool samples were tested to determine T. solium tapeworm carriers using a real-time PCR (T3qPCR) described by Ng-Nguyen et al. (2017) [22]. The assay has been reported to have a diagnostic sensitivity of 94% and a diagnostic specificity of 98%.
Human serum samples were tested for the presence of antibody against T. solium cysticerci using a lentil-lectin purified glycoprotein-enzyme-linked immunoelectrotransfer blot (LLGP-EITB). This assay has a diagnostic sensitivity of 98% and a diagnostic specificity of 100%. The LLGP-EITB was performed using the methodology described by Tsang et al. (1989) [23].
Pig serum samples were tested for the presence antibody against T. solium cysticerci using rT24H antigen in the enzyme-linked immunoelectrotransfer blot (EITB) format. The rT24H-EITB assay showed no cross-reaction to T. hydatigena and had a diagnostic sensitivity and specificity of 100% when tested on 29 cysticercosis-negative USA pig sera, 12 necropsy-positive T. solium-positive Peruvian pig sera and four T. hydatigena necropsy-positive Vietnamese pig sera [24]. The performance of the rT24H-EITB assay was carried out using the methodology described by Noh et al. 2014 [25]. Positive samples resulting from the rT24H-EITB assay were confirmed using the LLGP-EITB assay.
Details collected during each of the household visits and laboratory test results were stored in a relational database (Microsoft Access 2007, Microsoft Corporation, Redmond, USA). Longitude and latitude coordinates of household locations (recorded in degrees, minutes and seconds) were converted to decimal degrees and re-projected to the Universal Transverse Mercator Zone 48N projection using the World Geodetic System 1984 datum.
Analyses were carried out to describe the spatial characteristics of: (1) T. solium exposure-positive and T. solium exposure-negative households for humans; (2) T. solium exposure-positive and T. solium exposure-negative households for pigs; and (3) human or pig T. solium or T. solium taeniasis-positive and negative households.
Our classification of T. solium exposure-positive households for humans and pigs was based on the LLGP-EITB and rT24H-EITB assays, respectively. Classification of Taenia-positive households was based on parallel interpretation of the test results of the LLGP-EITB, rT24H-EITB and T3qPCR assays (positive for either T. solium exposure or T. solium taeniasis).
Ripley’s K-function [26] provides a summary measure of spatial dependence among point locations as a function of their Euclidean distance. The K-function is defined as the expected number of points that are located within a distance h of an arbitrarily selected point location, divided by the overall density of points [27]. Where there is spatial dependence in a point pattern, point events are likely to be surrounded by other point events and, for small vales of distance h, K(h) will be relatively large. Conversely, if point events are regularly spaced, each point is likely to be surrounded by empty space and, for small values of distance h, K(h) will be small. To facilitate inference, we developed separate K-function plots for T. solium exposure-positive and T. solium exposure-negative households. For each value of h we then calculated the K-function difference as D(h) = K(h)positive – K(h)negative. If exposure-positive households were spatially aggregated, over and above that of the exposure-negative households, then D(h) will appear graphically as peaks (or troughs) as a function of distance.
Three sets of K-function analyses were carried out for: (1) human T. solium exposure-positive households and human T. solium exposure-negative households; (2) pig T. solium exposure-positive households and pig T. solium exposure-negative households; and (3) T. solium taeniasis exposure positive and T. solium taeniasis negative households.
Monte Carlo simulation was used to construct critical envelopes for each K-function difference plot. Here, we randomly assigned the observed number of positive households across the population of study household locations and re-computed D(h) each time. The critical envelopes are based on 1000 Monte Carlo simulations of the data. Departures of the observed value of D(h) above the limits of the upper and lower critical envelopes provided an indication of spatial aggregation of exposure-positive households beyond that which would be expected by chance, and at what spatial scale.
The total numbers of households visited for collecting samples from humans and pigs, respectively, were 190 and 408 in Buon Don, Krong Nang and M’Drak districts. Within the 190 households, a total of 342 individuals consented to participate in the study. The number of pigs sampled from the 408 households was 1281. Four of the 190 households (2.1%, 95% CI 0.6 to 5.6) housed T. solium tapeworm carriers and the percentages of households housing individuals and pigs that were T. solium exposure positive was 8.9% (17/190, 95% CI 5.5 to 14) and 2.7% (11/408, 95% CI 1.4 to 4.9), respectively. Amongst the 11 T. solium exposure-positive households for pigs, there was one household that had more than one pig antibody-positive to T. solium cysticerci; all other exposure-positive households had a single pig that was seropositive. All T. solium exposure-positive households for humans had a single individual that was antibody-positive. Of the 561 households that were visited for either human or pig sampling, 31 had either humans or pigs that were either human T. solium exposure-positive, pig T. solium exposure-positive or T. solium taeniasis positive (Table 1). There were 29 households with single infections; one household with exposure-positive pigs and one household had an individual infected with T. solium taeniasis and an individual that was T. solium exposure-positive. Human T. solium exposure- and taeniasis-positive households were present in all three districts. There was no pig T. solium exposure-positive households in Buon Don.
In three study districts, the prevalence of having a household latrine was relatively low ranging from 13% to 47%. This meant that outdoor defecation was a practice reported by between 15% and 74% of the study population. Our data showed that allowing pigs to free roam was common practice in Buon Bon, Krong Nang, and M’Drak. The percentage of pigs that consumed human feces was high in Buon Don and M’Drak (Table 2). Of 11 exposure-positive households for pigs, there were nine households in M’Drak. Pigs kept in seven of the nine households were allowed to roam freely within the village.
No exposed pigs were detected in Buon Don (Fig 3C and 3D) and two exposed pigs were detected in Krong Nang (Fig 4C and 4D). In M’Drak, the four exposed pigs that were detected were within a distance of 250 m of each other (Fig 5C) and pairs of exposure-positive pig households were less than 100 m apart (Fig 5D). The K-function difference plot for T. solium exposure in pigs in M’Drak shows K(h)positive in excess of K(h)negative up to a distance of 1500 m (Fig 6F).
There were small numbers of human T. solium exposure-positive households in close proximity in Krong Nang (Fig 4B). The K-function difference plot for exposure to T. solium in humans in Krong Nang supported this observation, where K(h)positive was in excess of K(h)negative up to a distance of 200 to 300 m (Fig 6C). The spatial distribution of human exposure-positive households in Buon Don (Fig 3A and 3B) and M’Drak (Fig 5A and 5B) were more regularly distributed; there was no evidence of significant differences between K(h)positive and K(h)negative up to a distance of 3000 m (Fig 6A and 6E).
When we considered households that were either human T. solium exposure, taeniasis or pig T. solium exposure as a single group, the K-function difference plot showed all T. solium exposure-positive and taeniasis-positive households were aggregated up to a distance of 1000 m in M’Drak (Fig 6g). Similar associations were evident in Buon Bon and Krong Nang but the observed K-function difference plot did not exceed the simulation envelope limits at any distance (Fig 6B and 6D). On inspection, however, we observed a group of taeniasis- and exposure-positive households in close proximity to each other in the village of Cu Mta in M’Drak district (Fig 7).
This study describes the fine scale spatial distribution of T. solium exposure in pigs and humans in Vietnam for the first time. The geographic distribution of T. solium exposure- and taeniasis-positive households varied markedly across the districts of Buon Don, Krong Nang and M’Drak of Dak Lak province. A prominent feature of this data is that the prevalence of T. solium exposure in both species and T. solium taeniasis was relatively low (in humans 9 exposure- and 2 taeniasis-positive households per 100 households at risk; in pigs 3 exposure-positive households per 100 households at risk) making it difficult to definitively identify characteristics of the spatial distribution of positive households that are likely to exist across all districts of Dak Lak, and indeed all districts of Vietnam.
Spatial aggregations of T. solium exposure-positive households for humans occurred in some (the village of Ea Wer in Buon Don district (Fig 3B), Dlieya in Krong Nang district (Fig 4B) and Krong Jing and Cu Mta in M’Drak (Fig 5A and 5B)), but not all, of the villages in the three study districts. Our K-function difference plots showed that T. solium exposure-positive households for humans showed the same pattern of spatial dependence as T. solium exposure-negative households in Buon Bon and M’Drak (Fig 6A and 6E). In Krong Nang, compared with human exposure-negative households, human exposure-positive households were aggregated up to a distance of 200 to 300 m (Fig 6C). We speculate that if the prevalence of exposure was higher in Buon Don and M’Drak and sufficient resources were available to allow larger sample sizes in each of the two districts to be collected, a similar pattern of spatial dependence would be evident.
Although spatial aggregation of T. solium exposure-positive households for humans in Krong Nang was beyond that expected by chance (and entirely due to a collection of five positive households in the village of Dlieya), its overall magnitude was relatively small (Fig 6C). Dlieya is small village comprised of less than 200 households in a remote area of Krong Nang. In Dlieya the number of individuals per household was considerably larger than that of the other villages in the study (median 5; minimum 2 to maximum 11) and a notable feature was that it was common for several generations of a family to live together in close proximity, and a highly prevalent custom was that food was shared with neighbours and relatives on a daily basis, often associated with community ceremonies (e.g. weddings and anniversary of deaths). In the district of Krong Nang, houses are typically surrounded by a large garden comprised of coffee or pepper trees. We hypothesize that in this district, where a high proportion of the study population were known to defaecate outdoors, individuals were more likely to defecate in their own garden (as opposed to communal areas) which means that it was more likely for T. solium eggs to be present in close spatial proximity to a given household where exposure-positive individuals were present. We speculate that the anthropological and fine-scale environmental characteristics of Dlieya were sufficient to allow spatial clustering of human exposure infection to be detected even in the presence of a modest sampling effort (n = 30 households).
Spatial aggregations of exposure to T. solium in pigs occurred but this was infrequent. In three study districts, there was a single aggregation of pig exposure-positive households in M’Drak (Fig 5C). Our K-function difference plot for M’Drak (Fig 6F) showed pig exposure-positive households were clustered within a distance of 1500 m. Presumably, this was due to the larger range over which free-roaming pigs forage. Copado et al., 2004 [26] reported that free-roaming pigs travel daily within a distance ranging from 1000 to 3000 m. In a 12 hour period, pigs traveled a distance of up to 4000 m and spent, on average, 47% of their time outside of their homestead [28]. Our findings are supported by those of Ngowi et al. (2010) [12] who conducted a cross-sectional study of cysticercosis in 784 pig-owning households in northern Tanzania. In the study of Ngowi et al. (2010) it was shown that porcine cysticercosis was clustered within the distance of 600 m and 10 km. Morales et al. (2008) [29] conducted a cross-sectional study of 562 pigs in the state of Morelos in Mexico in 2003. In this study the prevalence of porcine cysticercosis was relatively high (13%; 95% CI 11 to 17) and while free-roaming pigs had a greater risk of being cysticercosis-positive, no geographical clustering of positivity was found.
Spatial clustering of T. solium exposure in pigs in M’Drak could have been associated with the age of the resident pig population, the absence of pigsties and the regular habit of coprophagy amongst pigs. In total, there were 11 T. solium pig exposure-positive households in Dak Lak. Nine of these 11 positive households were in M’Drak, aggregated in groups of two to four households (Fig 5C and 5D). Of the nine T. solium pig exposure-positive households in M’Drak, in seven households pigs were allowed to roam freely within the village, increasing the chance of exposure to T. solium eggs. The terrain in M’Drak is generally flat. The quality of the soil is poor supporting predominantly natural grasslands. For these reasons, the practice of allowing pigs to free roam is more common compared with the two other districts. M’Drak had a high proportion of pigs that were not confined (17% [211 of 1281], 95% CI 15 to 19). Of the total number of pigs sampled in this study, 511 (40%) were from M’Drak. Of the 511 M’Drak pigs that were sampled, it was reported that 193 (38%) consumed human feces and 211 (41%) regularly scavenged for food (Table 2).
When we considered households that were human and/or pig T. solium exposure-positive or taeniasis positive as a single group, our K-function difference plot showed these positive households were aggregated up to a distance of 1000 m in M’Drak (Fig 6G), but not in Buon Bon (Fig 6B) and Krong Nang (Fig 6D). O’Nea at al. (2012) and Pray et al. (2017) showed that human and/or porcine cysticercosis cases were strongly associated with the presence of tapeworm carriers [14,15]. Individuals and pigs living in close proximity to tapeworm carriers are more likely to be infected with T. solium cysticercosis [11,15,30]. Given this unique data set, with contemporary sampling of humans and pigs, it was of interest to us to determine if there was a spatial dependence between T. solium exposure- and T. solium taeniasis-positive households. Although households that had either T. solium exposure- or taeniasis-positive cases were spatially aggregated, we were unable to identify an association between the two because of the extremely low number (n = 4) of taeniasis-positive households across the three study districts. On inspection we observed a group of taeniasis- and exposure-positive households in close proximity to each other in the village of Cu Mta in M’Drak district (Fig 7). Madinga et al. (2017) [26] and Morales et al. (2008) [29] indicated that there was no spatial correlation of T. solium exposure in pigs and T. solium taeniasis.
Spatial aggregations of human and pig exposure-positive households occurred in some, but not all, of the villages in the three study districts. We can only speculate about the reasons for this pattern, as discussed above. Since cysticercosis occurs throughout Vietnam [17] it is likely that foci of infection are present in other areas. The relatively low prevalence of exposure to T. solium indicates that massive deworming programs in the communities of Dak Lak province are, for the most part, unnecessary. Instead, we recommend that if a human is identified as T. solium positive then either: (a) individuals resident in the immediate area should be tested to rule out the presence of an exposure or infection cluster; or (b) anthelmintic treatment is offered to individuals resident within a 2000 m radius of the identified case. With respect to the second approach, privacy issues would need to be handled appropriately, particularly in small communities.
A limitation of this study was that for logistic reasons sampling of humans and pigs were carried out independently resulting in a lack of overlap of the locations of households where humans and pigs were sampled (see, for example, Ea Nuol in the district of Buon Don, Fig 3A and 3C). While this limited our ability to identify an association (if any) between human and pig T. solium exposure-positive households, assessment of the spatial dependence of exposure status by species (human, pigs) was possible. Although households that had either T. solium exposure- or taeniasis-positive cases were spatially aggregated, we were unable to quantify their spatial association due to the extremely low number of T. solium taeniasis-positive households.
|
10.1371/journal.ppat.1007860 | The 2nd sialic acid-binding site of influenza A virus neuraminidase is an important determinant of the hemagglutinin-neuraminidase-receptor balance | Influenza A virus (IAV) neuraminidase (NA) receptor-destroying activity and hemagglutinin (HA) receptor-binding affinity need to be balanced with the host receptor repertoire for optimal viral fitness. NAs of avian, but not human viruses, contain a functional 2nd sialic acid (SIA)-binding site (2SBS) adjacent to the catalytic site, which contributes to sialidase activity against multivalent substrates. The receptor-binding specificity and potentially crucial contribution of the 2SBS to the HA-NA balance of virus particles is, however, poorly characterized. Here, we elucidated the receptor-binding specificity of the 2SBS of N2 NA and established an important role for this site in the virion HA-NA-receptor balance. NAs of H2N2/1957 pandemic virus with or without a functional 2SBS and viruses containing this NA were analysed. Avian-like N2, with a restored 2SBS due to an amino acid substitution at position 367, was more active than human N2 on multivalent substrates containing α2,3-linked SIAs, corresponding with the pronounced binding-specificity of avian-like N2 for these receptors. When introduced into human viruses, avian-like N2 gave rise to altered plaque morphology and decreased replication compared to human N2. An opposite replication phenotype was observed when N2 was combined with avian-like HA. Specific bio-layer interferometry assays revealed a clear effect of the 2SBS on the dynamic interaction of virus particles with receptors. The absence or presence of a functional 2SBS affected virion-receptor binding and receptor cleavage required for particle movement on a receptor-coated surface and subsequent NA-dependent self-elution. The contribution of the 2SBS to virus-receptor interactions depended on the receptor-binding properties of HA and the identity of the receptors used. We conclude that the 2SBS is an important and underappreciated determinant of the HA-NA-receptor balance. The rapid loss of a functional 2SBS in pandemic viruses may have served to balance the novel host receptor-repertoire and altered receptor-binding properties of the corresponding HA protein.
| Influenza A viruses infect birds and mammals. They contain receptor-binding (HA) and receptor-destroying (NA) proteins, which are crucial determinants of host tropism and pathogenesis. It is generally accepted that the functional properties of HA and NA need to be well balanced to enable virion penetration of the receptor-rich mucus layer, binding to host cells, and release of newly assembled particles. This HA-NA-receptor balance is, however, poorly characterized resulting in part from a lack of suitable assays to measure this balance. In addition, NA is much less studied than HA. NA contains, besides its receptor-cleavage site, a 2nd receptor-binding site, which is functional in avian, but not in human viruses. We now show that this 2nd receptor-binding site prefers binding to avian-type receptors and promotes cleavage of substrates carrying this receptor. Furthermore, by using novel assays, we established an important role for this site in the HA-NA-receptor balance of virus particles as it contributes to receptor binding and cleavage by virions, the latter of which is required for virion movement and self-elution from receptors. The results may provide an explanation for the rapid loss of a functional 2nd receptor-binding site in human pandemic viruses.
| Influenza A virus (IAV) particles contain hemagglutinin (HA) and neuraminidase (NA) glycoproteins. HA functions as a sialic acid (SIA)-binding and fusion protein. NA has receptor-destroying activity by cleaving SIAs from sialoglycans. The HA and NA protein functionalities are critical for host tropism, and need to be well balanced in relation to the host receptor repertoire for optimal in vivo viral fitness [1–3]. However, there is no standard assay and unit for measuring a functional balance and the precise mode by which HA- and NA-receptor interactions contribute to the balance at the molecular level remains mostly unexplored. An optimal HA-NA balance is hypothesized to allow virions to penetrate the heavily sialylated mucus layer, to attach to host cells prior to virus entry, and to be released from cells after assembly [4–7].
Aquatic birds constitute the natural reservoir of IAVs. Occasionally IAVs from birds cross the host species barrier and manage to adapt to non-avian species, including humans. The human receptor repertoire differs from avians and requires adaptations in the SIA-interacting HA and NA proteins for optimal interaction. The HA protein of avian IAVs prefers binding to terminally located SIAs linked to the penultimate galactose via an α2,3-linkage. Human IAVs preferentially bind to α2,6-linked sialosides [8–11]. Internal sugars and their linkages as well as glycan branching have been shown to determine fine specificity of HA-receptor binding [12–17]. Changes in the receptor-binding properties of the HA proteins are achieved by mutations in the receptor binding site, which have been well documented for several HA subtypes [1, 10, 11, 18]. Much less is known about the adaptations in NA required to match the corresponding HA proteins.
NA is a type II transmembrane protein that forms mushroom-shaped homotetramers. Tetramerization is essential for its enzymatic activity [19, 20]. The enzyme active site is located in the globular head domain that is linked to the endodomain via a thin stalk. The active site is made up by catalytic residues that directly contact SIA and by framework residues that keep the active site in place [21, 22]. The catalytic and the framework residues are extremely conserved between avian and human IAVs [23]. Nevertheless, although both avian and human NA proteins preferentially cleave α2,3-linked SIAs, human viruses appear relatively better at cleaving α2,6-linked SIAs [24–27].
Adjacent to the catalytic site, NA contains a 2nd SIA-binding site (2SBS; also referred to as hemadsorption site) (S1 Fig)[28–31]. The 2SBS is made up by three loops, which contain residues that interact with SIA. Mutations in these loops in N1, N2 and N9 affected NA binding of erythrocytes [28, 32–35] or sialosides [26, 33] and enzymatic cleavage of multivalent substrates [28, 33] but not of monovalent substrates [26, 28, 33]. A detailed analysis of the receptor binding properties of the 2SBS of most NAs is lacking. N1 and N2 proteins bind to α2,3- as well as α2,6-linked SIAs based on binding of resialylated erythrocytes [28, 35] whereas N1 and N9 proteins mainly bind, via their 2SBS, to α2,3-linked sialosides present on glycan arrays [33] or in biolayer interferometry assays [26]. Interestingly, the high conservation of SIA-contact residues in the 2SBS of avian IAV is lost in N1 and N2 of human IAVs [1, 26, 28, 30] that, supposedly, all lack a functional 2SBS. For N2 of avian viruses, the conservation of the 2SBS is only lost in viruses of the H9N2 subtype, which mainly infect Galliformes species, in contrast to other N2-containing viruses, which mainly infect Non-Galliformes species [36] (S2 Fig). Conservation of the SIA-contact residues in the 2SBS of N2 is also lost in canine and not restored in swine viruses, the latter of which are generally derived from human viruses (S2 Fig). It is tempting to hypothesize that the loss of a functional 2SBS in pandemic viruses is part of a required adaptation of the HA-NA balance in order to deal with the altered receptor repertoire in the novel human host [1, 28]. At first, to test this hypothesis, a detailed analysis of the contribution of (mutations in) the 2SBS to receptor binding and cleavage in the context of IAV particles is necessary as the interplay with HA proteins binding to either avian- or human-type receptors needs to be taken into account.
We define the HA-NA balance as the balance between the activities of HA and NA in virus particles in relation to their functional receptors on cells and decoy receptors present e.g. in mucus. We have recently established novel kinetic assays based on biolayer interferometry (BLI) with which, in the context of virus particles, HA binding, NA cleavage and their balance can be monitored in real time using synthetic glycans and sialylated glycoproteins [37]. Multivalent IAV-receptor binding is established by multiple low affinity interactions of several HA trimers and sialosides [38, 39]. This enables a dynamic binding mode in which individual interactions are rapidly formed and broken without causing dissociation of the virus but providing access of NA to temporarily free SIAs. Cleavage by NA results in reduced SIA-receptor density, in virus movement and ultimately in virion dissociation [37]. How fast this occurs depends on the HA-NA-receptor balance governing the dynamics of virus-glycan interactions.
In the present study, we applied these novel BLI assays to study the HA-NA-receptor balance of viruses that have a single amino acid substitution in the 2SBS. We first performed a detailed analysis of the functional importance of the 2SBS in N2 for substrate binding and cleavage by comparing NA of the pandemic H2N2 virus from 1957, containing a mutated 2nd SIA-binding site, with an avian-like NA, in which the 2nd SIA binding site was restored. Preferred binding to α2,3-linked sialosides was shown to result in enhanced cleavage of substrates containing these glycans. Analysis of the HA-NA-receptor balance of viruses containing these N2 proteins in combination with H3 proteins that prefer binding to avian or human-type receptors clearly demonstrated a role for the 2SBS in the complex and dynamic interplay between HA, NA and receptor, which has been largely overlooked until now. The functional importance of the 2SBS for the HA-NA-receptor balance may explain the conservation and loss of this site in avian and human IAVs, respectively.
We first analysed the receptor binding and cleavage activity of N2 NA with and without a functional 2SBS using purified recombinant soluble NA expressed in HEK293T cells [20]. In NA of A/Singapore/1/1957 (H2N2) pandemic virus (referred to as human N2 [hN2]) one of the SIA-contact residues in the 2SBS is mutated compared to the avian consensus sequence (S367N, S3 Fig). Introduction of the reciprocal mutation (N367S) in this NA restored the 2SBS (referred to as avian-like N2 [aN2]) [28]. hN2 and aN2 displayed similar specific activities when using the monovalent MUNANA [2’-(4-Methylumbelliferyl)-α-D-N-acetylneuraminic acid] substrate (Fig 1A, S4A and S4B Fig), indicating that mutation of the 2SBS did not affect the catalytic activity of the N2 proteins per se. Similar results were obtained previously using membrane-associated proteins [28], indicating that the activity of the recombinant soluble proteins accurately reflects the activity of their membrane-bound counterparts as concluded earlier for N1 [20]. Cleavage of SIAs from fetuin and transferrin sialoglycoproteins was quantified by enzyme-linked lectin assay (ELLA), by analysing the increase or decrease in binding of lectins depending on their binding specificities (S4 Fig). ECA (Erythrina Cristagalli lectin) specifically binds glycans containing terminal Galα1,4GlcNAc corresponding to non-sialylated N-linked sugars [40], while PNA (peanut agglutinin) binds to terminal Galβ1,3GalNAc, which generally corresponds to non-sialyated O-linked sugars [41]. NA activity thus results in increased binding of these lectins. MAL I (Maackia Amurensis Lectin I) and SNA (Sambucus Nigra Lectin) specifically bind α2,3- or α2,6-linked SIAs, respectively [42, 43]. Binding of SNA and MAL I is decreased by NA activity. For all lectins analysed, aN2 was more active than hN2 using fetuin, containing α2,3- and α2,6-linked SIAs (Fig 1A) [44]. In contrast, no statistically significant difference was observed using transferrin that only contains α2,6-linked sialoglycans (Fig 1A) [45, 46]. Plotting the specific activities of the NA proteins relative to their specific activities as determined by the fetuin-ECA combination resulted in similar activity profiles (Fig 1B), which mimic those determined previously for N1 and N9 [26, 33]. Both hN2 and aN2 preferred cleavage of α2,3- (determined with fetuin-MAL I) over α2,6-(determined with fetuin-SNA) linked SIAs (Fig 1B). In agreement herewith, the specific activities were higher when determined with the fetuin-ECA than with the transferrin-ECA combination as fetuin, but not transferrin, contains α2,3-linked SIAs (Fig 1B).
These results show that an avian-like 2SBS in N2 contributes to cleavage of the sialoglycoprotein fetuin containing α2,3- and α2,6-linked SIAs. We next used BLI to study the kinetics of NA activity on a multivalent surface coated with either an avian receptor (3’SLNLN: NeuAcα2-3Galβ1-4GlcNAcβ1-3Galβ1-4GlcNAc) or a human receptor (6’SLNLN: NeuAcα2-6Galβ1-4GlcNAcβ1-3Galβ1-4GlcNAc). NA activity can be directly monitored in real-time by the specific binding of the lectin ECA to terminal Galβ1-4GlcNAc glycotopes that become available upon removal of SIA by NA (Fig 1C and 1D, red and black lines). Note that cleavage of the small SIA moiety is not detected directly by BLI (Fig 1C and 1D, dashed red and black lines). Binding of ECA to a sensor coated with LNLN (Galβ1-4GlcNAcβ1-3Galβ1-4GlcNAc, Fig 1C and 1D blue lines) rapidly reaches the maximum ECA binding signal (representing 100% de-sialylation) assuring that ECA binding during the relatively slow accumulation of de-sialylated glycans by NA activity (red and black lines) reflects the cleavage kinetics of the N2 proteins. Both hN2 and aN2 more efficiently cleaved 3’SLNLN over 6’SLNLN. Especially the aN2 protein displayed much more efficient cleavage of 3’SLNLN. We conclude that restoration of the 2SBS in hN2 to the avian consensus sequence results in enhanced cleavage of substrates containing α2,3-linked SIAs.
The increased cleavage by aN2 of substrates containing α2,3-linked SIAs is expected to result from specifically increased binding to α2,3-linked SIAs due to the presence of an avian 2SBS, although N2 proteins were reported to bind both α2,3- and α2,6-linked SIAs by using resialylated erythrocytes [28]. We observed hemagglutination for recombinant soluble aN2 but not for hN2 (S5A Fig). However, no specific binding to synthetic α2,3- and α2,6-linked sialoglycans by BLI could be observed for the recombinant soluble N2 proteins, which could be due to low affinity of the 2SBS. By embedding the N2 proteins in membrane vesicles highly multivalent receptor interactions may increase receptor-binding avidity. To this end, full length N2 proteins were expressed in 293T cells. N2 virus-like particles (VLPs) [47] were directly harvested from the culture supernatant, while cells were treated with hypotonic and hypertonic buffers, resulting in the release of N2 protein-containing vesicles [48]. Preparations containing similar amounts of N2, based on MUNANA activity (Fig 1A) were used to determine the receptor specificity of the 2SBS by BLI [26]. Negligible binding was obtained for hN2 VLPs (Fig 2A and 2B) or vesicles (S5D and S5E Fig) to α2,3- or α2,6-linked SIAs, regardless of the presence of the NA inhibitor oseltamivir carboxylate (OC), which binds the NA catalytic site. In contrast, highly 3’SLNLN-specific, binding was observed for aN2 VLPs and vesicles (Fig 2A and 2B; S5D and S5E Fig) in the presence of OC leading to the conclusion that aN2 has much higher lectin activity than hN2 due to the presence of a functional 2SBS. The observed α2,3-linked SIA specificity is in agreement with the particularly enhanced cleavage of substrates containing α2,3-linked SIAs (Fig 1). No binding of aN2 VLPs to 3’SLNLN was observed in the absence of OC, which is likely explained by immediate self-elution of VLPs carrying active NA proteins.
To examine the contribution of the 2SBS to the HA-NA balance of virus particles we examined the replication phenotype of recombinant viruses containing either aN2 or the hN2 in the background of the 1968 pandemic virus A/Hong Kong/1/68 (H3N2) (referred to as hH3aN2 and hH3hN2) [28]. The hH3hN2 virus, lacking a functional 2SBS, produced large and clear plaques on Vero cells (Fig 3A and 3B, S6A and S6B Fig) as compared to the smaller, fuzzy plaques of the hH3aN2 virus with a functional 2SBS. Staining of plaques at 48 h post infection indicated that all cells within the plaques of hH3hN2 virus were infected, whereas many non-infected cells could be observed in the hH3aN2 plaques. This could be due to the more active aN2, which may destroy receptors on cells before the virus can enter into the cells. hH3aN2 reached lower titres than the hH3hN2 virus at 24 and 48 h post infection when Vero cells were used (Fig 3C), while no significant differences were observed for replication in MDCK cells (Fig 3D). Differences in cell surface sialosides and their distribution may explain differences between replication in Vero and MDCK cells. Although the sialylation patterns of MDCK and Vero cells are poorly characterized, both cell lines can be infected with human and avian IAVs and express α2,3- and α2,6-linked SIAs [49–51]. From these results we conclude that the absence or presence of a functional 2SBS in N2 may affect virus replication kinetics in a cell type-dependent manner.
Using a recently established BLI-based kinetic binding assay [37] an enhanced initial binding rate to 3’SLNLN but not 6’SLNLN (Fig 4A and 4B), was observed for hH3aN2 virus containing a functional 2SBS in comparison to hH3hN2. As a result the hH3aN2 virus displayed a higher initial binding-rate ratio 3’SLNLN/6’SLNLN than hH3hN2 (Fig 4C, red and black bars). Next, two recombinant soluble glycoproteins containing mainly N-linked glycans (lysosomal-associated membrane glycoprotein 1 [LAMP1], ca. 18 N- and 6 O-linked glycans [52, 53]), or O-linked glycans (glycophorin A, ca. 16 O- and a single N-linked glycan [54, 55]) were used in BLI as recently described for recombinant fetuin [37]. LAMP1 and glycophorin A mimic the presumed functional and decoy receptors found on cells (LAMP1) and on mucins (glycophorin A) that are rich in N- or O-glycans, respectively. Analysis of the glycans on these glycoproteins by lectin binding using BLI confirmed the presence of sialylated N-linked glycans (both α2,3- and α2,6-linked) on both proteins, while only glycophorin A was shown to contain sialylated O-glycans (S7 Fig). Again, a functional 2SBS (present in aN2) contributed to virus binding (Fig 4D and 4E). This contribution was larger for binding to glycophorin A than to LAMP1 as judged from the initial binding rates (Fig 4F).
The HA of the 1968 pandemic H3N2 virus (referred to as hH3) prefers binding to terminal α2,6-linked SIAs [56, 57]. The results above implicate that, besides adaptations in HA, also adaptations in the 2SBS may contribute to a specificity-switch when an avian IAV adapts to humans. We therefore studied the effect of the 2SBS in NA when combined with an avian-type HA preferring binding to α2,3-linked SIAs. We generated the corresponding recombinant A/Hong Kong/1/68 (H3N2) viruses containing 7 amino acid substitutions in the HA (see S8A Fig). These substitutions reverted the HA back to the avian consensus sequence (referred to as avian-like aH3), including the crucial substitutions Q226L and G228S, which enable HA preferential binding to avian-type receptors [56, 57]. The resulting viruses are referred to as aH3hN2 and aH3aN2, depending on the absence and presence of the functional 2SBS, respectively. We confirmed the receptor-binding specificities of soluble hH3 and aH3 proteins by solid phase fetuin- and transferrin-binding assays and BLI (S8B and S8C Fig). As expected, aH3 displayed higher binding levels to fetuin, containing α2,3- and α2,6-linked SIAs, than hH3, while hH3 bound better than aH3 to transferrin, which only contains α2,6-linked sialoglycans. BLI analysis using H3-containing vesicles obtained from cells expressing full-length versions of hH3 or aH3 confirmed the different receptor-binding properties of these H3 proteins to 3’SLNLN and 6’SLNLN (S8D and S8E Fig). In contrast to viruses containing hH3, the presence of a functional 2SBS in aN2 enhanced replication of viruses with aH3 both on Vero (Fig 3E) and MDCK (Fig 3F) cells. Differences in virus replication were smaller for MDCK than for Vero cells. We next analysed receptor-binding properties of aH3hN2 and aH3aN2 viruses using BLI. As observed before for the hH3-containing viruses (Fig 4C; red and black bars), a functional 2SBS enhanced binding to 3’SLNLN but not 6’SLNLN when N2 was combined with aH3 (Fig 4C). However, viruses containing aH3 displayed similar binding kinetics in the presence of OC regardless of the presence of a functional 2SBS for both LAMP1 and glycophorin A (Fig 4G, 4H and 4I). From these results we conclude that a functional 2SBS site in NA contributes to virion-receptor binding in a HA- and receptor-dependent manner.
The NA enzymatic activity of the different recombinant viruses with and without a functional 2SBS was analysed using the monovalent soluble substrate MUNANA, by ELLA and by BLI. The different viruses displayed a similar NA activity per particle using the monovalent soluble substrate MUNANA (Fig 5A). As also the NA proteins do not differ in their MUNANA activity regardless of the presence or absence of a functional 2SBS (Fig 1A), we conclude that similar amounts of NA are incorporated into virions of the four viruses. The viruses differed, however, in their specific activities when the multivalent glycoprotein fetuin was used as substrate in an ELLA (Fig 5B). hH3hN2 virus was less active compared to viruses containing aN2 and/or aH3, indicating a contribution of receptor binding via HA and the 2SBS to NA enzymatic activity in the context of virus particles. In agreement with the results obtained with the recombinant proteins (Fig 1B), cleavage of α2,6-linked SIA found on transferrin was less efficient and did not appear to differ significantly between the different viruses (S9 Fig).
The ELLAs (Fig 5B and S9 Fig) indicate that both receptor binding via HA and the 2SBS of NA contribute to the sialidase specific activity of virus particles. These endpoint assays do not, however, elucidate the HA-NA balance of these viruses, for which kinetic BLI assays are required [37]. Preliminary experiments showed inefficient cleavage of the synthetic glycans by the recombinant viruses. Kinetic assays to determine the HA-NA balance of these viruses were therefore performed with the glycoprotein receptors (LAMP1 or glycophorin A). In the absence of OC, that is, with active NA proteins, no appreciable binding of hH3-containing viruses could be detected indicating efficient receptor cleavage by NA (Fig 5C and 5D). Limited binding could be detected, however, for the aH3-containing viruses in the absence of OC. The binding curve of the virus with a functional 2SBS (aH3aN2) bended earlier and had a smaller area under the curve than that of the virus without a functional 2SBS (aH3hN2) for both LAMP1 and glycophorin A. This bending of the curves is explained by ongoing cleavage of SIAs by viruses attached to the sensor-attached glycoproteins, resulting in release of bound virus particles [37]. The earlier bending and smaller area under the curve observed for the aH3aN2 virus is indicative of more efficient cleavage of the sensor-attached receptors by this virus than by aH3hN2, lacking a functional 2SBS.
The effect of receptor binding via NA and HA on NA activity of virions was analysed further by NA-dependent virion self-elution from a receptor-coated BLI sensor after prior binding of the virions in the presence of OC. Self-elution of IAV particles requires NA activity and self-elution is not observed when NA activity is blocked by OC [37]. After binding of the four recombinant viruses to LAMP1 and glycophorin A in the presence of OC, OC was removed by repeated short washes in Dulbecco’s phosphate buffered saline (PBS) with Calcium and Magnesium and virus self-elution was monitored. Clearly, viruses with aN2 proteins eluted faster from the sensors than the viruses with hN2 (compare hN3aN2 with hN3hN2 and aH3aN2 with aH3hN2; Fig 5E and 5F), for both glycoprotein receptors. Of note, NA-depended self-elution of virus particles is often preceded by an apparent increase in virus binding [37] represented here as negative self-elution, particularly in the case of aH3aN2 and aH3hN2 (Fig 5F). The larger negative area of self-elution for aH3hN2 reflects the reduced NA activity of this virus compared to aH3aN2. Also the identity of HA affected the virus self-elution rate. Viruses with hH3 eluted faster than corresponding viruses with aH3 (e.g. compare hH3aN2 with aH3aN2). For hH3aN2, self-elution was faster from glycophorin A than from LAMP1. For aH3-containing viruses, the opposite was observed. Differences in virion self-elution observed for different HA-receptor combinations could be due the different receptor repertoires present on the two proteins (S7 Fig). The results indicate that receptor binding via the 2SBS of NA contributes to enzymatic cleavage by NA in virions and to virion self-elution from a receptor-coated surface. Virion self-elution was also shown to depend on the identity of the HA and the glycoprotein receptor used.
The S367N mutation in the 2SBS of N2 was rapidly obtained after emergence of the H2N2 pandemic virus in 1957 and was observed in human H2N2 viruses until 1958. Most viruses isolated thereafter did not contain the S367N mutations but rather contained the S370L mutation, which also results in loss of a SIA-contact residue in the 370 loop (S3 Fig) and hemadsorption activity [28]. Both single mutations had a similar negative effect on catalytic activity of the 1957 NA [28]. These results indicate that there was not a selection against S367 per se, but rather against a functional 2SBS, which is achieved by either mutation. However, several additional mutations accumulated in time in the three loops of the 2SBS. N2 from the A/Hong Kong/68 (H3N2) (referred to as HK N2) contains five mutations (S370L, N400S, N401D, W403R, P432K) in the 2SBS compared to the avian consensus sequence (Fig 6A). To analyse the contribution of the 2SBS to the enzymatic activity of these different N2 proteins, a comparative analysis of recombinant proteins and viruses using monovalent and multivalent substrates was performed. The HK N2 protein was 4–5 fold less active than hN2 both on the monovalent substrate MUNANA and the multivalent substrate fetuin. Cleavage of sialoglycans attached to transferrin was not significantly affected (Fig 6B). We also compared the NA activity of recombinant H3N2 viruses only differing in their NA segment. HK H3N2 virus, containing the 1968 HK N2 protein, displayed 2-fold lower NA activity per virus particle than the hH3hN2 virus, containing the 1957 N2 protein, as determined by MUNANA cleavage (Fig 6C). Similarly, the time required for 50% self-elution of virions from the multivalent receptor LAMP1, was 2-fold longer for HK H3N2 than for hH3hN2 (Fig 6D). Thus, hN2 from 1957 and 1968 HK N2 differ to a similar extent in their catalytic activity both when monovalent or multivalent substrates are used. As receptor-binding via the 2SBS only increases NA activity for multivalent, but not monovalent substrates, we conclude that these differences do not result from differences in receptor-binding by their (non-functional) 2SBS. Moreover, the difference observed when comparing the two recombinant proteins is similar to the difference in activity of the two NAs in the virus context.
Since the discovery of hemadsorption activity in NA 1984 [58] and the structural evidence of the 2SBS in N9 1997 [29], only few studies have addressed 2SBS-mediated receptor binding and the functional consequences thereof for NA activity [26, 28, 33, 34, 59]. We now show that the 2SBS is an important factor in the complex interplay between HA, NA and receptors, referred to as the HA-NA-receptor balance. A functional 2SBS in N2 was shown to prefer binding to α2,3-linked sialosides similarly to N1 [26] and N9 [33]. In agreement herewith, it enhances catalytic activity against substrates carrying α2,3-linked SIAs. The contribution of the 2SBS to the HA-NA-receptor balance of virus particles was shown to be receptor- and HA protein-dependent as demonstrated by kinetic analysis of receptor-binding and -cleavage of virions using BLI. The 2SBS was shown to contribute to receptor binding also when NA was combined with a receptor-binding HA in IAV virions, as well as to cleavage of receptors by virions and to virion self-elution from a receptor-coated surface. The absence or presence of a functional 2SBS also affected virus replication in a cell type- and HA-dependent manner. Our results indicate that mutation of the 2SBS as observed in early human pandemic viruses negatively affects the catalytic activity of NA and may serve to restore the HA-NA-receptor balance of viruses carrying HA proteins with altered receptor-binding properties in relation to a novel host sialome. Conservation of the 2SBS in most avian strains, with the notable exception of H9N2 viruses, is lost in human [26, 29, 30, 34], swine and canine variants (S2 Fig). Strong conservation usually reflects a critical function. It would be very interesting to investigate in depth whether a critical function for the 2SBS in avian strains, for instance related to the HA-NA-receptor balance, is not required for efficient replication and transmission of human, canine and swine strains.
N2 prefers binding of α2,3- over α2,6-linked SIAs via its 2SBS. The specificity of the N2 2SBS correlates with the enhanced cleavage of substrates carrying α2,3-linked SIAs compared to substrates carrying only α2,6-linked sialosides. Of note, enhanced activity was also observed for α2,6-linked SIAs at least when these sialosides were linked to substrates additionally carrying α2,3-linked SIAs (Fig 1A; fetuin-SNA combination). These results indicate that the 2SBS enhances catalytic activity by bringing sialosides on multivalent substrates close to the catalytic site and that, depending on the substrate used, the enhanced cleavage of SIAs not necessarily matches the specificity of the 2SBS. Preferred binding of avian-type receptors via its 2SBS was previously also observed for N9 [33] and N1 [26], suggesting that this is a conserved feature for NAs of different subtypes. We cannot exclude, however, that the 2SBS of different NA subtypes may differ in their receptor-binding fine specificity, as structural differences were observed in the interactions between ligands and the 2SBS for different NA subtypes [30]. In N9, the conserved K432 residue in the 2SBS forms a hydrogen bond with SIA [29] and mutation K432E in N1 has a large negative effect on the cleavage of multivalent substrates [26]. In contrast, several other avian NA subtypes, including N2, contain a Q or E residue at this position, which does not form a hydrogen bond with SIA in the few available crystal structures [30]. Previously, it was shown that N1 and N2 NAs bound with similar efficiency to both avian and human type receptors SIAs [28, 35]. This discrepancy is probably explained by the different methods used to analyse the receptor specificity of the 2SBS. In the previous reports, a red blood cell binding assay was employed, in which desialylation of erythrocytes was followed by resialylation using α2,3- or α2,6-sialyltransferases. Binding to resialylated erythrocytes might be affected by prior incomplete desialylation. Alternatively, a higher receptor density on erythrocytes compared to the BLI sensor surface might allow for binding of α2,6-linked SIAs. The ability of the 2SBS to bind human-type receptors to some extent is also suggested by the modestly increased or decreased cleavage of SIAs from substrates only containing α2,6-linked SIAs upon the introduction of mutations in the 2SBS (this study and [26, 28]).
The 2SBS contributed to receptor-binding also when NA was combined with a receptor-binding HA in IAV virions. In combination with HA preferring binding to α2,6-linked SIAs (hH3), the 2SBS enhanced binding for all receptors analysed, except 6’SLNLN, to which the recombinant aN2 protein did not bind. Binding to glycophorin A, carrying many O-linked sugars also found on mucins, was more enhanced by the 2SBS than binding to LAMP1, which carries mostly sialylated N-glycans. The functional significance of this difference remains to be determined. When combined with HA that prefers binding to α2,3-sialosides (aH3), the enhancing effect of the 2SBS was not observed for the glycoprotein receptors analysed. Thus, the contribution of NA to virion-receptor binding depends on the specificity/affinity of the corresponding HA and the receptors present. Previously it was shown that the active site of NA contributes to virion-receptor binding in case of a low-activity catalytic site [37], a characteristic which is also appears to be displayed by recent H3N2 viruses [60, 61]. As we now show that a functional 2SBS in NA can also contribute to virion-receptor binding, two mechanisms exist by which NA can assist in binding of virions to host cells.
A complex interplay between HA, NA and receptor determines the attachment of virus particles to and release from a receptor-containing surface. This HA-NA-receptor balance can be experimentally determined using kinetic BLI assays by analysis of virus binding in the absence or presence of NA inhibitors and self-elution from different receptors (this paper and [37]). The HA-NA-receptor balance determines the residence time of a virus on a sialylated surface and the speed by which it moves over this surface. We assume that an optimal balance is important for virions to efficiently pass the heavily sialylated mucus layer, while still allowing virion attachment to host cells resulting in endocytic uptake. The complexity of the HA-NA-receptor balance is exemplified by the contribution of NA to receptor binding [37] and of HA to the apparent catalytic activity of NA (this paper)[37, 62]. We now show that the HA-NA-receptor balance as reflected for example in virion self-elution (Fig 5) is affected by a functional 2SBS, depending on the particular HA with which NA is combined and the receptors used. Changes in the 2SBS of NA should thus be considered in the context of mutations affecting the receptor-binding site of HA and the catalytic site of NA.
The 2SBS of N2 appears to accumulate more mutations than other surface exposed parts of the NA protein (S3 Fig). While the 1957 N2 protein has a single substitution in the 2SBS, the 1968 N2 protein contains five mutations in this site. The accumulation of several mutations in the 2SBS was found to have no further negative effects on the enzyme-enhancing function of 2SBS as compared to a single mutation of a SIA contact residues in the 2SBS of an early pandemic virus from 1957. Although we cannot exclude that the accumulation of mutations in the 2SBS of N2 indicates ongoing adaptation of NA to the human host or serves to restore subtle deviations in the HA-NA-receptor balance resulting from other mutations in HA and/or NA, it seems more likely that it rather results from continuous immune pressure on this site [22, 63] in combination with loss of functional importance of the 2SBS in human viruses.
An important role for the NA 2SBS in IAV replication in vivo is suggested by the conservation of this site among NA subtypes of most avian viruses, the rapid loss of this site in human pandemic viruses ([1, 26, 28, 30, 36] and S2 Fig), the important role of this site in HA-NA-receptor balance (this study) and observations that this site affects virus replication in vitro ([26, 34, 59] and this study). Of note, we now show that the presence or absence of a functional 2SBS affected virus replication depending on the receptor-binding properties of HA, with which NA was combined. Replication of viruses with a human or avian-like HA is enhanced by the absence or presence of a functional 2SBS, respectively, although some cell-dependent differences were observed. The absence or presence of a functional 2SBS was reported not to affect influenza viral replication in ducks [34]. However, in this latter study recombinant viruses were used containing HA from a H2N9 and NA from a H3N2 virus. This may have resulted in a mismatched HA-NA combination in which the presence of the 2SBS might be of minor influence on replication. Alternatively, the 2SBS may be important for virus transmission rather than for replication in ducks per se. Clearly, additional experiments are needed to demonstrate the importance of the 2SBS for IAV replication and transmission in vivo. Interestingly, both for H9N2 and H7N9 viruses, the well-known Q226L mutation in the receptor-binding site of HA, resulting in a shift from avian to human receptor specificity, is associated with mutations in the 2SBS that negatively affect receptor binding [33, 36]. These avian viruses thus display a striking parallel with the changes observed in the receptor-binding sites of HA and NA of avian-origin pandemic viruses. We propose that mutations in the 2SBS of avian viruses may be indicative of an as of yet underappreciated, increased potential of avian viruses to cross the host species barrier. Of note, also upon introduction of coronavirus OC43 into humans, the lectin function of the receptor-destroying hemagglutin-esterase protein was lost through progressive accumulation of mutations resulting in reduced cleavage of multivalent substrates [64]. Thus, both coronaviruses and IAVs appear to adapt to the sialoglycome of the human respiratory tract by tuning the virion receptor-binding and cleavage functions, the latter among others by mutation of the lectin domain of the receptor-destroying NA.
Human-codon optimized cDNAs (Genescript) encoding the N2 ectodomain of A/Singapore/1/57(H2N2) (GenBank accession no. AY209895.1; referred to as human N2 [hN2]) and a variant thereof containing the N367S mutation (referred to as avian-like N2 [aN2]) were cloned into a pFRT expression plasmid (Thermo Fisher Scientific) in frame with sequences encoding a signal sequence derived from Gaussia luciferase, a Strep tag and a Tetrabrachion tetramerization domain, similarly as described previously [20]. The corresponding full length (FL) NA-coding plasmids were generated by replacement of the non-NA coding sequences by sequences encoding the NA transmembrane domain and cytoplasmic tail of N2 of A/Singapore/1/57(H2N2). Human-codon optimized cDNAs encoding FL H3 or the H3 ectodomain of A/Hong Kong/1/68 (H3N2) (GenBank accession no. CY033001; referred to as human H3 [hH3]) or of an variant thereof containing 7 amino acid substitutions, which revert the HA back to the avian consensus sequence [56] (referred to as avian-like H3 [aH3]) were cloned in pCD5 expression vectors similarly as described previously [65]. Codon optimized glycoproteins LAMP1 and glycophorin A ectodomain-encoding cDNAs (Genescript) were genetically fused to Fc-tag, for Protein-A based purification, and a Bap tag [66], for binding to octet sensors, and cloned in a pCAGGs vector, similarly as described previously for fetuin [37]. NA and glycoprotein expression plasmids were transfected into HEK293T (ATCC) cells using polyethylenimine (PolyScience) [20]. An expression vector encoding BirA ligase was cotransfected with the LAMP1- and glycophorin A-coding vectors [37]. Five days post transfection, cell culture media containing soluble NA proteins and glycoproteins were harvested and purified using Strep tactin or protein A containing beads [20, 37]. Purified NA proteins were quantified by quantitative densitometry of GelCode Blue (Thermo Fisher Scientific)-stained protein gels additionally containing bovine serum albumin (BSA) standards. The signals were imaged and analysed with an Odyssey imaging system (LI-COR). HEK293T cells were transfected with full-length NA constructs to obtain membrane vesicles. To this end, cells were vesiculated as described previously [26, 48]. VLPS and membrane vesicle preparations were purified using Capto Core 700 beads (GE Healthcare Life Sciences) according to the manufacturer’s instructions and as detailed previously [67] to remove proteins smaller than 700 kDa. The amount of NA protein in the VLPs and vesicle preparations was determined using the MUNANA assay described below.
Generation of recombinant virus HK H3N2, which harbours all genes from the pandemic virus A/Hong Kong/1/68 (H3N2) has been described before [57]. Also the generation of hH3hN2 and hH3aN2 viruses, which carry the N2 gene of the pandemic A/Singapore/1/1957 (H2N2) in the background of A/Hong Kong/1/68 (H3N2) has been described before [28]. The hH3aN2 virus contains substitution N367S in the N2 protein. aH3hN2 and aH3aN2 viruses were generated as described previously [56] in the background of A/Hong Kong/1/68 (H3N2). These latter viruses carry the H3 protein of A/Hong Kong/1/68 (H3N2) containing 7 amino acid substitutions in HA which revert the HA back to the avian consensus sequence [56] combined with the N2 protein of A/Singapore/1/1957 (H2N2) with (aH3aN2) or without (aH3hN2) the N367S substitution. Virus stocks were grown in MDCK-II cells (ECACC). Viruses were inactivated by UV radiation using UV Stratalinker 1800 (Stratagene) on 50,000 μJoules prior to their use in the binding and cleavage assays. UV inactivation did not affect the enzymatic activity of NA as determined with the MUNANA assay.
The NA enzymatic activity was determined by using a fluorometric assay [68] in combination with 2’-(4-Methylumbelliferyl)-α-D-N-acetylneuraminic acid (MUNANA; Sigma-Aldrich) as described previously [20]. Enzymatic activity of the NA proteins towards multivalent glycoprotein substrates was analysed using a previously described enzyme-linked lectin assay (ELLA) [33]. In brief, fetuin- or transferrin-coated plates were incubated with serial dilutions of recombinant soluble NA proteins. After overnight incubation at 37°C, plates were washed and incubated with either biotinylated Erythrina Cristagalli Lectin (ECA, 1.25 μg/ml; Vector Laboratories), biotinylated peanut agglutinin (PNA, 2.5 μg/ml; Galab Technologied), biolinylated Sambucus Nigra Lectin (SNA, 1.25 μg/ml; Vector Laboratories) or biotinylated Maackia Amurensis Lectin I (MAL I, 2.5 μg/ml; Vector Laboratories). Cleavage of SIAs from fetuin and transferrin was quantified by analysing the increase (PNA and ECA) or decrease (MAL I and SNA) in binding of different lectins depending on their binding specificities (S4 Fig). The binding of ECA, PNA, SNA and MAL I was detected using horseradish peroxidase (HRP)-conjugated streptavidin (Thermo Fisher Scientific) and tetramethylbenzidine substrate (TMB, bioFX) in an ELISA reader EL-808 (BioTEK) by measuring the optical density (OD) at 450 nm. The data were fitted by non-linear regression using the Prism 6.05 software (GraphPad). The resulting curves were used to determine the amount of NA protein corresponding to half maximum MUNANA cleavage or lectin binding. The inverse of this amount is a measure of specific activity (activity per amount of protein) and was graphed relative to other NA proteins or substrate-lectin combinations.
Plaque assays were performed in Vero cells (ATCC) as described previously [69]. One hour after infecting the cell monolayers with 30–50 plaque forming units of the virus in 1 ml of maintenance medium, the virus inoculum was removed and cells were covered the Avicel RC-581 overlay medium and cultures were incubated at 37°C in 5% CO2 atmosphere. After three days of incubation, the overlay was removed by suction and the cells were fixed with 10% formalin and stained with 1% crystal violet solution in 20% methanol in water. For immunostaining, cells were fixed with 4% paraformaldehyde solution for 30 min at 4°C, washed with PBS and permeabilized by incubation for 10–20 min with buffer containing 0.5% Triton-X-100 and 20 mM glycine in PBS. Cell layers were incubated with monoclonal antibodies specific for the influenza A virus nucleoprotein (kindly provided by Dr. Alexander Klimov at Centers for Disease Control, USA) for 1 hour followed by another 1 hour incubation with peroxidase-labeled anti-mouse antibodies (DAKO, Denmark) and 30 min incubation with precipitate-forming peroxidase substrates True Blue. Stained plates were washed with water to stop the reaction, scanned on a flatbed scanner and the data were acquired by Adobe Photoshop 7.0 software.
To characterize replication kinetics of different recombinant viruses, two replicate cultures of Vero or MDCK cells in 12-well plates were infected with each virus at MOI 0.001 (Vero cells) or 0.0001 (MDCK cells). Inocula were removed 1 hpi, fresh medium was added, and cultures were incubated at 37°C. Samples of culture supernatant were taken 24, 48 and 72 hpi and stored frozen. They were titrated together using focus formation assay in MDCK cells as described previously [57]. Numbers of infected cells per well were counted for the virus dilution that produced from 30 to 300 infected cells per well and recalculated into numbers of focus forming units (FFU) per ml of the original undiluted virus suspensions.
For the full length protein-containing vesicles and VLPs, similar amounts of NA activity, and thus NA protein, were applied in the BLI assays using the Octet RED348 (Fortebio). Inactivated virus preparations were analysed using Nanoparticle Tracking Analysis (Nanosight NS300, Malvern) as detailed below in order to use similar number of virus particles in the BLI assays. BLI assays were performed as described previously [37]. All experiments were carried out in Dulbecco's PBS with Calcium and Magnesium (Lonza) at 30°C and with sensors shaking at 1000 rpm. Streptavidin biosensors were loaded to saturation with biotinylated synthetic glycans 2,3-sialyl-N-acetyllactosamine-N-acetyllactosamine (3’SLNLN), 2,6-sialyl-N-acetyllactosamine-N-acetyllactosamine (6’SLNLN), N-acetyllactosamine-N-acetyllactosamine (LNLN), LAMP1 or glycophorin A glycoproteins. Synthetic glycans were synthesized at the Department of Chemical Biology and Drug Discovery, Utrecht University, Utrecht, the Netherlands. For the NA kinetic cleavage assay, the sensors loaded with synthetic glycans were incubated in 100 μl buffer containing 4 μg recombinant soluble aN2 or hN2 in the absence or presence of 8 μg ECA. As controls, sensors were also incubated with ECA in the absence of N2. Association of the N2 VLPs, vesicles and virus particles was analysed for 30 minutes in the absence or presence of 10 μM OC (Roche). For viruses, the virus association phase in the presence of OC was followed by three 5 s washes and a dissociation phase in the absence of OC. Initial binding rates were determined similarly as previously described [37]. For lectin binding, the sensors loaded with recombinant glycoproteins were incubated with the different lectins (8 μg/100 μl) for 15 minutes.
NTA measurements were performed using a NanoSight NS300 instrument (Malvern) following the manufacturer’s instructions. The UV-inactivated virus preparations were diluted with PBS to reach a particle concentration suitable for analysis with NTA. All measurements were performed at 19°C. Per analysis, the NanoSight NS300 recorded five 60 second sample videos, which were then analysed with the Nanoparticle Tracking analysis 3.0 software, resulting in quantitative information on particle number and particles sizes (S10 Fig). Each virus preparation was analysed twice and mean values were used. NTA measurements were validated by analysis of virus stocks quantified earlier by silver staining of viral proteins after electrophoresis on polyacrylamide gels [37]. Results obtained via both methods correlated well (less than 25% deviation).
|
10.1371/journal.ppat.0030026 | A Specific Primed Immune Response in Drosophila Is Dependent on Phagocytes | Drosophila melanogaster, like other invertebrates, relies solely on its innate immune response to fight invading microbes; by definition, innate immunity lacks adaptive characteristics. However, we show here that priming Drosophila with a sublethal dose of Streptococcus pneumoniae protects against an otherwise-lethal second challenge of S. pneumoniae. This protective effect exhibits coarse specificity for S. pneumoniae and persists for the life of the fly. Although not all microbial challenges induced this specific primed response, we find that a similar specific protection can be elicited by Beauveria bassiana, a natural fly pathogen. To characterize this primed response, we focused on S. pneumoniae–induced protection. The mechanism underlying this protective effect requires phagocytes and the Toll pathway. However, activation of the Toll pathway is not sufficient for priming-induced protection. This work contradicts the paradigm that insect immune responses cannot adapt and will promote the search for similar responses overlooked in organisms with an adaptive immune response.
| Due to the common practice of vaccination and prominence of AIDS, people are already aware of the distinction between adaptive and innate immunity without realizing it. All organisms have an innate immune response, but only vertebrates possess T cells and the ability to produce antibodies. It has been a long-standing assumption that invertebrate immune systems are not adaptive and respond identically to multiple challenges. In this study, we demonstrate that the fly innate immune response adapts to repeated challenges; flies preinoculated with dead Streptococcus pneumoniae are protected against a second, otherwise-lethal dose. Although the underlying mechanisms are likely to be very different, this primed response is reminiscent to vaccine-induced protection in that it exhibits coarse specificity (dead S. pneumoniae only protects against itself), persists for the life of the fly and is dependent on phagocytic cells. This result prompts the obvious question of whether the innate immune system of vertebrates shares a similar biology. Such a finding is of particular interest since immunocompromised individuals only possess an innate immune system.
| Immune responses are typically characterized as being either adaptive or innate. Adaptive immunity, which requires T and B cells, is specific, has memory, and is generally considered to be restricted to vertebrates. In contrast, the innate immune response is thought to act naïvely to each encounter with a pathogen [1,2]. Innate immunity depends on the recognition of broadly conserved molecular moieties and exhibits only weak specificity, such as the ability to distinguish between different structural classes of peptidoglycan [3]. However, recent work suggests that the invertebrate innate immune response may exhibit adaptive characteristics (reviewed in [4] and [5]).
Functional immune adaptation can be defined most broadly as any case where an immune response differs between a first and second challenge. The simplest form involves the immune system remaining activated after an initial challenge. This sort of response has long been known in invertebrates as shown by Hans Boman and coworkers [6]. They found that antibacterial activity in Drosophila hemolymph persists after bacterial challenge and can provide protection against subsequent challenges. More recently, Moret et al. [7] found a similar persistence of humoral antibacterial activity in mealworms.
More complex adaptive phenomena have also been observed in arthropods; for example, flour moths [8] and Daphnia [9] possess strain-specific immunity that is passed from a mother to her offspring. The molecular mechanisms underlying the maternal transfer of strain-specific protection have not been characterized. Specific memory has also been examined in cockroaches [10] and bumblebees [11]. In both insects, the initial immune activation is nonspecific and confers protection against many types of challenges. However, both cockroaches and bees are also able to mount long-term specific protection: a priming dose of a particular species of bacteria only protects against that species (or class of species in the case of bumblebees). From this work, it seems that innate immunity possesses the adaptive characteristics of specificity and memory; unfortunately, the animals used in past studies have not been amenable to deeper analysis. We examined the well-understood Drosophila innate immune response for specificity and memory because this model organism would give us genetic and physiological assays to dissect adaptive aspects of innate immunity.
Drosophila has been proved to be a powerful model organism to study innate immunity [1,2]. The Drosophila innate immune response has three effector mechanisms: the humoral response, melanization, and the cellular response [1,2]. The humoral immune response involves the secretion of soluble factors, such as antimicrobial peptides (AMPs), into the hemolymph following immune activation. Melanization is the process whereby melanin is deposited at wound sites and parasite surfaces, resulting in the release of toxic reactive oxygen species. The cellular immune response consists of hemocytes that phagocytose, encapsulate, and kill invading microbes, much like vertebrate macrophages. These mechanisms depend in various ways on pathogen detection via the Toll or imd signaling pathways [1,2,12,13].
We found that S. pneumoniae–primed flies are protected against a subsequent lethal challenge with S. pneumoniae. This response is specific for S. pneumoniae and persists for the life of the fly. In this paper, we demonstrate that the Toll pathway, but not the imd pathway, is required for this protective effect. Notably, activation of the Toll pathway is not sufficient to elicit a primed response. We show that AMPs are not involved and that phagocytes are the critical effectors of the primed response. Taken together, we demonstrate that the Drosophila primed response is specific and persists for the life of the fly. We identified a signaling pathway required for the process and ascertained which branch of the fly immune response is responsible for the primed response.
We found that previous exposure to S. pneumoniae permanently alters the fly's response to this bacterium. S. pneumoniae is a Gram-positive encapsulated bacterium that is the causative agent of otitis media, pneumonia, and meningitis [14]. The injection of 3,000 colony-forming units (CFU) directly into Drosophila hemolymph is normally lethal, killing the fly within 2 d (Figure 1A), and death is correlated with bacterial proliferation (Figure 1B). However, flies primed with a sublethal dose of bacteria were protected against a lethal challenge of S. pneumoniae administered 1 wk later (Figure 1C). Flies primed with S. pneumoniae and challenged with a lethal dose died at the same rate or slower than wounded controls. A priming dose of dead heat-killed S. pneumoniae was also sufficient to protect flies against a subsequent lethal challenge. Priming flies with S. pneumoniae thus induces long-term changes in the fly immune response, conferring protection against an otherwise-lethal challenge.
We considered two scenarios that could explain enhanced survival in S. pneumoniae–primed flies. First, bacterial numbers may not differ between naïve versus primed flies, and primed flies might survive the stress of the infection better than naïve flies. Second, S. pneumoniae–primed flies could kill the bacteria faster, and bacterial clearance would correlate with enhanced survival. To distinguish between these two possibilities, we examined bacterial load in naïve versus primed flies. Flies were injected with a priming dose of either dead S. pneumoniae or phosphate-buffered saline (PBS) 1 wk prior to a challenge of 400 CFU S. pneumoniae. This dose was chosen to emphasize the difference between naïve and primed flies; 400 CFU is the lowest dose that is lethal to naïve flies but not S. pneumoniae–primed flies (unpublished data). Within 1 d, S. pneumoniae–primed flies had killed almost all of the S. pneumoniae, whereas naïve flies still contained bacteria (Figure 1D). This result indicates that the survival difference between naïve and S. pneumoniae–primed flies results from different rates of S. pneumoniae killing.
Functionally, immunological memory is characterized by a more effective immune response upon repeat exposure that persists for the life of the animal. Although the best-described model for immune memory involves T and B cells and recombination-derived variation of receptors, we note that the definition of memory is independent of mechanism and immune memory could arise in a variety of ways. We demonstrated that preinoculation with dead S. pneumoniae alters the fly immune response such that it is more effective against subsequent challenges with S. pneumoniae. To determine how long these immune changes persist in the fly, we next varied the length of time between the priming and challenge dose. Flies primed with dead S. pneumoniae on day 0 were challenged between 1 and 14 d later with a lethal dose of S. pneumoniae. S. pneumoniae–primed flies always died significantly more slowly than naïve PBS-injected flies (Figures 1E and S1). An interval of 2 wk between the priming dose and challenge dose was the longest time we could assay; at 3 wk postpriming, the flies are actually 5 wk old and die from the stress of wounding alone (unpublished data). In summary, protection due to a priming dose of S. pneumoniae was detectable within 24 h and persisted for the life of the fly, or as long as we could assay survival differences.
To explore the specificity of this immune response, we asked whether other pathogens can induce a protective response against themselves. We chose a broad range of microbes pathogenic to wild-type Drosophila, including a Gram-negative bacterium, a Gram-positive bacterium, a Mycobacterium, and a natural fungal pathogen [15–18]. Priming doses of heat-killed Salmonella typhimurium, Listeria monocytogenes, and Mycobacterium marinum did not elicit a protective effect against a subsequent lethal challenge of the same bacteria used in the priming dose (Figures 2A and S2). Dead bacteria were used as priming doses in these experiments because all of these bacteria have an LD50 of one bacterium. Protection by priming is thus not a general characteristic of all microbial challenges in the fly. However, a priming dose of the natural fungal pathogen, Beauveria bassiana [17], conferred a protective effect against a subsequent lethal challenge (Figures 2A and S2).
Perhaps S. pneumoniae is a uniquely powerful immune activator; priming with S. pneumoniae might protect against challenges with other pathogens. To test this hypothesis, we challenged S. pneumoniae–primed flies with lethal doses of the same panel of microbes as above. S. pneumoniae–primed flies were not protected against lethal challenges of other pathogens (Figures 2A and S2). The protective effect of the primed response thus is not due to general activation of the Drosophila immune response.
Having seen that the priming-induced protective response was specific in the sense that S. pneumoniae was incapable of protecting against other immune challenges, we next determined whether other immune activators were capable of inducing a protective response against S. pneumoniae. To test this, we primed flies with a mixture of strong immune activators (dead Escherichia coli, dead Micrococcus luteus, and dead Beauveria bassiana) [17,19–21]. Flies injected with this mixture were not protected against a lethal challenge of S. pneumoniae (Figure 2B). Furthermore, injection with this mixture does not interfere with the ability to induce a protective response because the addition of dead S. pneumoniae to the mixture protected the flies from a second lethal challenge (Figure 2B). These results also demonstrate that a priming dose of B. bassiana does not protect against a lethal dose of S. pneumoniae. We conclude that protection conferred by a priming dose of S. pneumoniae specifically protects against lethal doses of S. pneumoniae and persists for the life of the fly.
Toll and imd are the best-characterized fly immunity pathways that control the majority of genes found to be induced by bacterial and fungal infections, including AMPs [1,2,17,19–21]. The mixture of microbes used in Figure 2B was chosen to strongly activate both the Toll and imd pathways [19,20]. Although activation of Toll and imd signaling is insufficient to induce a protective response, it remained possible that the pathways are necessary for protection. We tested loss-of-function mutants in each pathway to determine whether they were necessary for a protective response. Because both Toll and imd pathway mutants are immunocompromised with respect to S. pneumoniae (Figure S3), we reduced the lethal challenge dose of S. pneumoniae (20 CFU for Toll mutants, 100 CFU for imd mutants). These doses are normally sublethal to wild-type flies, but higher doses killed the mutant flies too quickly to detect a difference in survival. Loss-of-function mutants of imd (Figure 3) and dTak1 [22] (unpublished data) were protected by a priming dose of dead S. pneumoniae. Thus, although the imd pathway normally contributes to killing S. pneumoniae, it is not necessary to elicit a protective response. In contrast, flies homozygous for partial loss-of-function mutations that disrupt the Toll pathway, PGRP-SA [23,24] (Figure 3) and Dif [25] (unpublished data), were not protected by a priming dose of dead S. pneumoniae. The Toll pathway is therefore necessary for the primed response, but the imd pathway is not required.
Next, we wanted to determine the contribution of melanization, the humoral response, and the cellular response to priming (Figure 4A) [1,2]. We do not observe melanization in response to S. pneumoniae infections and therefore do not expect that it plays a strong role in the protective response (unpublished data) [26]. Because the Toll pathway is required for the priming-induced protection, we first examined the contribution of the inducible humoral response.
We found four lines of evidence suggesting that AMPs are not responsible for protecting flies against a second lethal challenge of S. pneumoniae. First, S. pneumoniae was not a strong AMP inducer because the peak of AMP transcription in response to S. pneumoniae was not as high as the peak induced by the positive control elicitor E. coli [27]. Flies were injected with a priming dose of dead S. pneumoniae, media (wounding control), or E. coli (positive control), and quantitative real-time reverse transcription–PCR (qRT-PCR) was used to assess transcript levels of three different AMPs: defensin (Figure 4B), attacin (Figure S4A), and diptericin (Figure S4B) [27]. These AMPs were chosen because they are strongly induced by Gram-positive bacterial infections. Second, we found that not only was S. pneumoniae a poor inducer of AMPs but also AMP transcript levels did not remain elevated 1 wk later (Figure 4B). Thus, AMP transcription should be back to the ground state by the time the challenge dose is administered, 1 wk after the priming dose. It has been reported that AMPs can persist in the hemolymph [28]. However, we have shown that simultaneous activation of Toll and imd, and thus AMP induction, is not sufficient to protect the fly (Figure 2B). Finally, we asked if this ground state is “sensitized”—that is, whether AMP induction is enhanced in flies that have been primed with S. pneumoniae compared to media. Using qRT-PCR, we measured defensin (Figure 4C), attacin (Figure S5A), and diptericin (Figure S5B) transcript levels after a second challenge of S. pneumoniae, media (wounding control), or E. coli (positive control). None of the AMPs were differentially induced in S. pneumoniae–primed flies compared to naïve media-injected flies. Thus, we find no evidence to support the involvement of AMP induction in the primed immune response.
In light of these data, and the fact that hemocytes are altered in Toll pathway mutants [12,13], we examined whether the cellular immune response is the main effector of the primed response—that is, a priming dose of S. pneumoniae might specifically increase S. pneumoniae clearance by phagocytes upon a second exposure. We first assessed the contribution of phagocytosis to S. pneumoniae killing in naïve flies by injecting flies with polystyrene beads to block phagocytosis prior to infection [29]. Bead-inhibited flies were extremely sensitive to S. pneumoniae; 3,000 bacteria were required to kill a wild-type fly, whereas 20 CFU was sufficient to kill a wild-type fly that lacks phagocytosis (Figure 4D). We demonstrated above that the survival difference between S. pneumoniae–primed flies and naïve flies is linked to enhanced clearance (Figure 1D); here we show that phagocytosis is required to kill S. pneumoniae in naïve flies.
We then asked if the enhanced clearance in primed flies is due to increased killing by phagocytes and not a second, mechanistically, different method of killing. To test this, we inhibited phagocytosis in both primed and naïve flies and then challenged with a lethal dose of S. pneumoniae 1 wk later, at which time point phagocytosis remained inhibited [29]. Primed flies died at the same time as naïve flies and were therefore not protected by a priming dose of S. pneumoniae, regardless of whether they were injected with beads before (Figure 4E) or after (unpublished data) the priming dose. Fly phagocytes are therefore an essential effector of the primed response.
These data suggested that a priming dose of S. pneumoniae activates fly phagocytes to kill S. pneumoniae more efficiently. Is this enhanced killing specific to S. pneumoniae, or does a priming dose of S. pneumoniae simply cause general phagocyte activation? If phagocytes are generally more activated in S. pneumoniae–primed flies, these flies should be able to clear other bacteria more rapidly. To test this hypothesis, S. pneumoniae–primed flies were tested for their ability to kill E. coli (Figure 4F). There was no difference between naïve PBS-injected and S. pneumoniae–primed flies in their ability to clear E. coli. Combined with the fact that a priming dose of S. pneumoniae does not offer protection against any other lethal challenges of bacteria (Figure 2A) and the fact that the primed response is specific for S. pneumoniae (Figure 2), we conclude that a priming dose of S. pneumoniae alters the fly immune system in a persistent manner that specifically allows phagocytes to recognize and kill S. pneumoniae more efficiently.
We have presented evidence that the fly modulates its immune response as a result of multiple challenges: a priming dose of S. pneumoniae is sufficient to protect the fly against a subsequent lethal dose of S. pneumoniae. Using a functional immune assay, we have shown that the fly immune system exhibits the adaptive characteristics of specificity and persistence. The mechanism underlying this protective response requires the Toll pathway, although its contribution is not via activation of AMPs. We have eliminated contributions from the imd pathway and AMPs and identified phagocytes as the critical effectors of the primed response. This system is uniquely positioned to further characterize the molecular basis underlying specific phagocyte activation and other adaptive aspects of innate immunity.
Flies were maintained on standard dextrose medium at 25 °C and 65% humidity. All experiments were performed with 5- to 7-d-old male wild-type Oregon R flies. All mutant flies were back-crossed onto the Oregon R background to limit background effects. In particular, white mutant flies are very sensitive to S. pneumoniae and do not elicit a primed response (unpublished data). Mutant lines used in this study include PGRP-SAseml (from P. Ligoxygakis), Dif, imd10191, and Tak12527. Molecular information for imd10191 is included below.
The imd10191 line has a 26-nucleotide deletion that frameshifts the protein at amino acid 179, which is the beginning of the death domain.
Microbial strains used in this study include S. pneumoniae strain SP1, E. coli DH5α, M. luteus, B. bassiana, L. monocytogenes strain 10403S, S. typhimurium strain SL1344, and M. marinum strain M. S. pneumoniae cultures were grown standing at 37 °C 5% CO2 in brain heart infusion broth (BHI) (BD Bioscience, http://www.bdbioscience.com) to an OD600 of 0.15, and aliquots were frozen at −80 °C in 10% glycerol. For infection, an aliquot of S. pneumoniae was thawed, diluted 1:3 in fresh BHI, and allowed to adjust for 2 h at 37 °C 5% CO2. E. coli, S. typhimurium, and L. monocytogenes cultures were grown standing overnight in BHI at 37 °C. M. luteus was cultured standing at 29 °C in BHI for 1 wk or until a sufficient density was reached. M. marinum was cultured standing at 29 °C in Middlebrook 7H9 broth (BD Bioscience) supplemented with Middlebrook OADC (BD Bioscience) and 0.2% Tween. B. bassiana spores were grown on malt agar (BD Bioscience) at 29 °C for 2 wk or until a sufficient density was reached.
For injection, flies were anesthetized with CO2 and injected with a total volume of 50 nl using individually calibrated pulled glass needles attached to a Picospritzer III injector (Parker Hannifin, http://www.parker.com). Flies were always injected in the abdomen, close to the junction with the thorax and just ventral to the junction between the ventral and dorsal cuticles. Flies were never anesthetized for longer than 10 min. After each injection, all flies were transferred to a new vial and maintained at 29 °C and 65% humidity.
To prepare priming doses of microbes, concentrated cultures were boiled for 30 min, centrifuged 5 min at 2,000g, and washed three times in PBS. Bacterial cultures were diluted in PBS to an OD600 of 0.1 and stored at −80 °C. Heat-killed B. bassiana spores were counted on a hemocytometer, adjusted to a concentration of 1 × 107/ml in PBS, and stored at −80 °C. Aliquots were plated on the appropriate media to verify that the microbes had been heat-killed. To prime flies, 50 nl of these solutions was injected into the fly. Flies were incubated at 29 °C and 65% humidity until they received their second challenge.
Fresh cultures were washed three times in PBS and diluted to the appropriate OD600 in PBS. For the different concentrations of S. pneumoniae, appropriate bacterial load corresponding to different optical densities was experimentally determined. For reference, an OD600 of 0.1 corresponds to 3,000 CFU. Lethal doses of other bacterial species are as follows: S. typhimurium, OD 0.1 (10,000 CFU); L. monocytogenes. OD 0.01 (6,500 CFU); and M. marinum, OD 0.05 (500 CFU). Bacterial load after injection was verified for all strains except M. marinum by plating on the appropriate media (blood agar for S. pneumoniae, BHI agar for all other strains). For lethal B. bassiana challenges, flies were anesthetized in groups of 20 and shaken on a plate of spores for exactly 30 s. All infections were carried out at 29 °C.
Individual flies were homogenized in 100 μl of PBS, diluted serially, and spotted onto appropriate plates. S. pneumoniae were grown on blood agar supplemented with 500 μg/ml streptomycin (Sigma, http://www.sigmaaldrich.com), 10 μg/ml colistin (Sigma), and 5 μg/ml oxolinic acid (Sigma) to eliminate the growth of bacterial contaminants from the fly. Plates were incubated overnight at 37 °C 5% CO2. E. coli colonies were grown on LB and incubated overnight at 37 °C.
Flies were challenged as described above and incubated at 29 °C for the indicated time points. At the given times, triplicates of three flies were anesthetized, placed in 1.5-ml tubes, and homogenized in 100 μl of Trizol-LS (Invitrogen, http://www.invitrogen.com). RNA was extracted using the standard Trizol-LS protocol, and remaining genomic DNA was degraded with DNase I treatment. RT-PCR was carried out with a Bio-Rad iCycler (http://www.bio-rad.com) using TaqMan probes and rTth polymerase (Perkin-Elmer, http://www.perkinelmer.com) as directed by the manufacturer. The following primers below were used. Relative RNA quantities were determined with respect to Drosophila ribosomal protein 15a, and all levels were normalized with respect to the zero time point for media injection: defensin TTCTCGTGGCTATCGCTTTT (left primer), GGAGAGTAGGTCGCATGTGG (right primer), AGGATCATGTCCTGGTGCATGAGGA (Taqman probe); attacin CAATGGCAGACACAATCTGG (left primer), ATTCCTGGGAAGTTGCTGTG (right primer), AATGGTTTCGAGTTCCAGCGGAATG (Taqman probe); diptericin ACCGCAGTACCCACTCAATC (left primer), CCCAAGTGCTGTCCATATCC (right primer), CAGTCCAGGGTCACCAGAAGGTGTG (Taqman probe); and ribosomal protein 15a TGGACCACGAGGAGGCTAGG (left primer), GTTGGTTGCATGGTCGGTGA (right primer), TGGGAGGCAAAATTCTCGGCTTC (Taqman probe).
Carboxylate-modified blue fluorescent 0.2-μm-diameter polystyrene beads (Molecular Probes, http://www.invitrogen.com) were injected to block phagocytosis essentially as previously described [29]. Briefly, beads were washed twice in sterile water and resuspended in one fourth of the original volume. Flies were injected with 50 nl of bead solution or water as an injection control. To confirm that phagocytosis was inhibited, the in vivo phagocytosis assay was performed as described previously with FITC-conjugated E. coli or FITC-conjugated Staphylococcus aureus. Phagocytic inhibition was confirmed each time bead-injected flies were manipulated.
All experiments were performed at least three times. For survival analysis, a minimum of 45 flies were injected for each condition. Dead flies were counted daily, and survival data were graphed and analyzed using GraphPad Prism (GraphPad Software, http://www.graphpad.com). Mean survival time with standard error was calculated using R (http://www.r-project.org). |
10.1371/journal.pgen.1005976 | Role of Double-Strand Break End-Tethering during Gene Conversion in Saccharomyces cerevisiae | Correct repair of DNA double-strand breaks (DSBs) is critical for maintaining genome stability. Whereas gene conversion (GC)-mediated repair is mostly error-free, repair by break-induced replication (BIR) is associated with non-reciprocal translocations and loss of heterozygosity. We have previously shown that a Recombination Execution Checkpoint (REC) mediates this competition by preventing the BIR pathway from acting on DSBs that can be repaired by GC. Here, we asked if the REC can also determine whether the ends that are engaged in a GC-compatible configuration belong to the same break, since repair involving ends from different breaks will produce potentially deleterious translocations. We report that the kinetics of repair are markedly delayed when the two DSB ends that participate in GC belong to different DSBs (termed Trans) compared to the case when both DSB ends come from the same break (Cis). However, repair in Trans still occurs by GC rather than BIR, and the overall efficiency of repair is comparable. Hence, the REC is not sensitive to the “origin” of the DSB ends. When the homologous ends for GC are in Trans, the delay in repair appears to reflect their tethering to sequences on the other side of the DSB that themselves recombine with other genomic locations with which they share sequence homology. These data support previous observations that the two ends of a DSB are usually tethered to each other and that this tethering facilitates both ends encountering the same donor sequence. We also found that the presence of homeologous/repetitive sequences in the vicinity of a DSB can distract the DSB end from finding its bona fide homologous donor, and that inhibition of GC by such homeologous sequences is markedly increased upon deleting Sgs1 but not Msh6.
| In budding yeast, DNA double-strand breaks (DSBs) are mostly repaired by gene conversion (GC) in which an intact region with homology to both ends serves as template for repair. If only one of the DSB ends shares homology with another region, repair proceeds via break-induced replication (BIR), which, unlike GC, can lead to nonreciprocal translocations. We previously showed that a Recombination Execution Checkpoint (REC) delays BIR thereby preventing it from acting on breaks that can be repaired by GC. Here, we induced two DSBs such that only one end from each break shares homology with an ectopic donor so that although repair can proceed via GC, it will result in DNA rearrangement. We asked whether REC would inhibit GC-mediated repair in this scenario. We found that REC does not restrict such repair; however, repair by GC is faster when both ends are derived from a single DSB. We suggest that the DSB ends engage in a coordinated search for homology that favors both ends engaging the same donor, thus avoiding complex chromosomal rearrangements. Moreover, when the two ends of a DSB share homology with different loci, recombination of one end with its partner impedes repair of the other end.
| Correct and efficient repair of DNA double-strand breaks (DSBs) is crucial for cell survival. DSBs can be repaired either by nonhomologous end joining (NHEJ), which involves simple re-ligation of the broken ends with little or no homology, or by homologous recombination (HR) in which an intact homologous sequence serves as template for repair [1–3]. Of all repair pathways, gene conversion (GC) is most extensively used to repair DSBs in Saccharomyces cerevisiae [4,5]. GC is a homology-driven process in which the DSB ends are resected to produce 3’-ended single-stranded DNA tails, which become coated with the Rad51 recombinase protein. Rad51 nucleoprotein filaments search for and strand-invade homologous sequences, which then serve as the template for synthesis of a short patch of DNA required to seal the break. During GC, new DNA synthesis is detected within ~30 minutes of strand-invasion, as measured by a PCR-based primer extension assay that requires only ~45 nucleotides to be added to the synapsed 3’ DSB end [6,7]. Repair by GC does not require components of the lagging strand DNA synthesis machinery such as Polα and primase [8]. Even though GC is associated with an elevated rate of mutations, it is the least error-prone mode of DSB repair [9–11]. However, repair can proceed via GC only if both ends of the break share homology with the donor.
If homology to only one of the ends is present, repair proceeds via a different pathway called break-induced replication (BIR) [12–17]. Although BIR involves both leading- and lagging-strand synthesis it does not proceed by a classical replication fork. Strand-invasion by the homologous end leads to the establishment of a migrating D-loop that can copy all sequences distal to the site of homology, while the DSB end that lacks homology is lost by degradation [18–20]. Compared to GC, the efficiency of BIR is quite variable, ranging from ~10% to as much as >90%, depending upon the position of the DSB and its homologous partner, the length of homology and the extent of DNA synthesis required to complete repair [13–15,18]. BIR is kinetically slower than simple gene conversion, and new DNA synthesis, as assessed by the primer-extension assay, does not initiate until ~3 h after Rad51-mediated strand-invasion of the donor sequence by the homologous end in cycling cells [13–15]. Given that the primer-extension assay requires synthesis of at least ~45 nucleotides to yield a PCR signal, it remains possible that shorter stretches of DNA may be added to the synapsed 3’ end during this 3-h period. BIR requires all three major DNA polymerases and associated DNA replication factors such as Cdc7, Cdt1 and the Cdc45-GINS-MCM helicase complex [14,21] although the MCM helicase was found to be less important when a DSB end shares extensive homology with a donor, such as a homologous chromosome [20]. Only components specifically needed for origin-dependent DNA replication (Cdc6 and the ORC proteins) appear to be completely dispensable for BIR. In addition, BIR requires the 5’ to 3’ helicase, Pif1 [19,20], as well as Pol32 [13,14], the nonessential subunit of DNA polymerase δ [22]. Furthermore, BIR is blocked by mutations in the PCNA replication clamp protein that have little effect on replication or gene conversion [21].
When a DSB could be repaired by either GC or BIR, the GC outcomes prevail [15]. We have previously shown that a Recombination Execution Checkpoint (REC) mediates the choice between the GC and BIR pathways prior to the actual initiation of stable repair synthesis [13]. This choice is based on the topology of the engaged ends such that even when homology to both ends is present, if the homologies lie far from each other or if they are close together but in the wrong orientation, REC restricts the rapid initiation of new DNA synthesis from the synapsed ends. In the absence of such a restriction, both ends could initiate independent/uncoordinated BIR-like repair events, some of which may get resolved by single-strand annealing to produce a gap-repair GC outcome; but in many instances repair would yield a nonreciprocal translocation. If a DSB occurred within a repeated sequence, the two ends could engage two different templates and their uncoordinated repair could produce two non-reciprocal translocations. Therefore, REC may play an important role in maintaining genome integrity and preventing chromosomal translocations by imposing a delay in the initiation of BIR-mediated repair, presumably providing more time for the DSB ends to find homology in the vicinity of each other, while ensuring quick and efficient repair when ends are engaged in a GC-compatible configuration. This delay in the initiation of BIR is partially suppressed by deletion of the Sgs1 helicase [13]. Recently we have shown that the kinetics of BIR becomes as rapid as repair of a short 1.2 kb gap when both Sgs1 and Mph1 helicases are deleted [23].
Since REC is able to sense how the DSB ends are engaged in terms of their orientation and distance with respect to each other [13], we wished to examine if it can also sense the origin of the two ends. In other words, would REC impede repair if ends from two different breaks were involved in repair, even if the two ends could synapse close to each other at the donor in the correct orientation? It is plausible that impeding such repair would restrict deleterious chromosomal translocations that can arise from repair of ends belonging to two different breaks. Besides REC, DSB end-tethering may also play a role in preventing such chromosomal translocations. It has been shown that the ends of a DSB remain tethered to each other for up to several hours after the induction of a break [24–27]. This end-tethering is dependent on several factors including the Mre11-Rad50-Xrs2 complex (MRX), Tel1, Sae2, γ-H2AX and Rad52 such that deletion of either of these factors results in un-tethering of the DSB ends in ~10–25% of the cells. End-tethering has been shown to play an important role in preventing NHEJ-mediated chromosomal translocations, presumably by limiting the interaction between ends from different breaks. However, the importance of DSB end-tethering in HR-mediated repair has not been carefully examined. To address these questions, we compared the efficiency and kinetics of repair between strains in which the DSB ends involved in repair either arose from the same break (Cis) and were tethered together or originated from two different breaks (Trans) and therefore, were not tethered to each other. We found that REC cannot sense the origin of the engaged ends per se and does not play a role in restricting GC-mediated repair in Trans. Nevertheless, the kinetics of repair in Trans are much slower relative to repair in Cis; however, this delay is due to the involvement of the opposite end of one of the DSBs in another, competing repair event. We suggest that DSB ends often remain tethered and travel together during the homology search, such that if one end engages with its homologous partner, it influences the fate of the other end by dragging it along.
In the course of this work we also found that neighboring sequences can seriously affect the efficiency of homologous recombination. We observed a sequence-context related diminution in repair both during a gene conversion event as well as during single-strand annealing.
Since the distance separating the strand-invaded DSB ends and their orientation are important signaling parameters during DSB repair, we wanted to test whether GC-mediated quick and efficient repair also required the two ends to come from the same break. Strain, YSJ379, referred to as Trans (Fig 1A) was designed to receive two DSBs, each on a different chromosome, created by galactose-inducible HO endonuclease. In this configuration one end from each of these DSBs can pair with a homologous donor on a third chromosome in a GC-compatible configuration. Specifically, the two ends of a LEU2 gene (LE on Chr XI and U2 on Chr V) can synapse with the LEU2 donor on Chr III, and repair can be mediated by an ectopic recombination event–analogous to a gene conversion mediated by synthesis-dependent strand annealing (SDSA)–that results in a translocation between chromosomes V and XI. The other two ends–the URA3 end from the break on Chr XI and the ura3 end from the break on Chr V–can be repaired efficiently by interchromosomal single-strand annealing (SSA), also resulting in a chromosomal translocation and loss of one of the URA3 repeats (Fig 1A). For comparison, we built another strain (YSJ357, designated Cis) (Fig 1B), which also harbors two HO-inducible breaks, one of which can be repaired by GC and the other by SSA. Here, the LE and U2 ends that participate in the GC event originate from the same DSB on Chr V, while the second DSB on Chr XI, which is flanked by ura3 and URA3 sequences, is repaired by intrachromosomal SSA. We have shown previously and also confirm below that interchromosomal SSA is as efficient, and only slightly slower than intrachromosomal SSA [28]. Hence, using these two strains, we can compare the efficiency and kinetics of the repair events in which the LE and U2 ends arise from a single break (Cis arrangement) or from two different breaks (Trans arrangement).
We first examined the efficiency of repair in these two strains by a viability assay and found that while ~73% of the cells were able to survive the breaks when the ends were in the Cis arrangement, ~61% of the cells survived the breaks in the Trans arrangement (Fig 2A). We then studied the kinetics of repair in these strains using a PCR assay with primer pairs as shown in Fig 1. We found that overall, the SSA product accumulated more rapidly than the GC product in both cases (Fig 2B); however, there was an hour-long delay in the appearance of the interchromosomal SSA product (Trans-SSA) relative to the intrachromosomal SSA product (Cis-SSA) (Fig 2B). This delay likely reflects the difference in the time it takes to find homology on the same chromosome (in Cis) as opposed to finding it on an altogether different chromosome (in Trans) [29,30], but may also be attributed to the competition in homology searching from the adjacent ends in the Trans configuration (see below). In contrast, the kinetics of LEU2 repair were markedly different between Cis and Trans, with the Trans-LEU2 product appearing at least 2 h after the appearance of the Cis product (Fig 2B), even though LEU2 repair in both cases is an interchromosomal event. These differences in repair kinetics cannot be attributed to differential kinetics of HO break formation as our Southern assays show >90% cutting at the HO sites within an hour of HO endonuclease induction in both configurations (S1 Fig).
One of the differences between the Cis and Trans strains is that while the U2 end originates from a break on Chr V in both cases, the LE end originates from a break on Chr XI in the Trans case as opposed to Chr V in the Cis case. Since the distal arm of Chr V does not contain any essential genes and can be lost without causing lethality, U2-mediated BIR events could also contribute to the overall efficiency of repair in the Cis arrangement (S2 Fig). However, in the Trans case, BIR will not yield a viable outcome; and this difference could explain the somewhat higher efficiency of repair in Cis relative to Trans. To address this possibility, we used a PCR-based approach to identify the proportion of colonies that had repaired the break in Cis by BIR (S2 Fig). We found that only ~2% of the repaired Cis colonies gave a product consistent with the BIR outcome indicating that the latter is not responsible for the higher efficiency of repair in Cis.
To further test whether the difference in efficiency and kinetics of LEU2 repair between the Cis and Trans arrangements could be due to an effect of chromosomal context relating to the origin of the LE and U2 ends in the two cases, we reversed the positions of the DSB cassettes in both configurations. We first built a Reverse-Cis strain, tNS2614, (Fig 1C) in which the positions of the LE-HOcs-U2 and the 5’ truncated ura3-HOcs-URA3 cassettes are reversed relative to the Cis strain such that the LE end is now present on Chr XI (just as in the Trans strain). We found that the overall efficiency of repair was somewhat reduced relative to the Cis strain, and only ~61% of the cells were able to survive the breaks in the Reverse-Cis configuration (Fig 2A). However, the kinetics of LEU2 repair were quite comparable between the Cis and Reverse-Cis strains (Fig 2C) with the repair product appearing at least 2 h earlier in these configurations relative to the Trans arrangement. We also constructed an analogous Reverse-Trans strain, tNS2638, (Fig 1D) in which the positions of the LE-HOcs-URA3 and the 5’ truncated ura3-HOcs-U2 cassettes were reversed relative to the Trans strain. We found that the efficiency (Fig 2A) as well as the kinetics of repair (Fig 2D) were indistinguishable between the Trans and Reverse-Trans configurations.
Since the viability of the Reverse-Cis strain was comparable to that of the Trans and Reverse-Trans strains, we conclude that the configuration of the DSB ends per se does not affect the overall efficiency of repair. However, the slower kinetics of LEU2 repair in the Trans and Reverse-Trans configurations is due to the involvement of DSB ends originating from two different breaks, as opposed to the Cis and Reverse-Cis cases where repair involves ends belonging to the same break.
Because there is an hour-long delay in the kinetics of ura3 SSA repair in the Trans strain relative to the Cis strain (Fig 2B), we wondered if this delay could indirectly contribute to the slower kinetics of LEU2 repair in this setting, perhaps by soaking up one or more repair factor(s). To address this possibility we studied the kinetics of LEU2 repair in a modified Cis strain, which lacks one of the SSA substrates and therefore suffers an unrepairable break on Chr XI, in addition to the LE-HOcs-U2 break on Chr V (Fig 3A). As a further control, we also compared the kinetics of LEU2 repair in these strains with a strain that harbors just the single LE-HOcs-U2 break on Chr V (Fig 3B, [13]). We found that although the overall repair efficiencies of these strains were different (Fig 3C), the kinetics of LEU2 repair, when normalized to the amount of total product formed at 15 h, remained the same irrespective of the presence of an unrepairable break, or the lack of a second break altogether (Fig 3D). Overall, these data clearly demonstrate that GC-like repair involving ends from different breaks is kinetically slower than repair involving ends from the same break.
The reduced efficiency and kinetics of LEU2 repair in the Trans arrangement could be due a defect in signaling between the ends because they are unlinked to each other. Thus, even though the LE and U2 ends will synapse next to each other in the correct orientation (with the LEU2 donor), they might fail to generate a signal for quick repair because they don’t belong to the same break. In the absence of this signaling, the ends might slowly initiate two independent BIR events, whose products may subsequently anneal with each other giving rise to a GC/SDSA-like outcome [31,32]. To address this possibility, we deleted POL32, which is required for BIR but has only a modest effect on GC [14], to test if the kinetically slow Trans repair occurred by BIR rather than GC. Deleting POL32 had a mild effect on the efficiency of repair in both Cis and Trans (Fig 2A), consistent with previous studies of its effect on GC [14]. The absence of a much greater defect in the Trans case strongly argues against the likelihood of repair in Trans being initiated by two independent BIR events. Hence, even though LEU2 repair in Trans is kinetically slower, it does not appear to be mechanistically different from repair in Cis.
The MRX complex, Tel1, Sae2 γ-H2AX and Rad52 proteins have all been implicated in holding the ends of a single DSB together for up to several hours after break formation [24–27]. We wondered whether this end-tethering might be responsible for the quicker kinetics of repair in the Cis case. Synapse formation may happen more quickly and efficiently in the Cis strain where the LE and U2 ends can travel together compared to the Trans strain where the LE and U2 ends must independently seek out and pair with the LEU2 donor. To examine the possible effect/s of end-tethering, we deleted RAD50 to see if releasing the ends in the Cis arrangement would make them behave more like ends in the Trans arrangement, resulting in similar LEU2 repair kinetics. We note that deleting RAD50 or its partners MRE11 and XRS2 –in addition to reducing end-tethering [25,26]–also causes defects in both 5’ to 3’ resection of DSB ends and the DNA damage checkpoint [33–36], but those effects should be identical in our Cis and Trans arrangements and therefore, should not contribute toward equalizing repair in Cis and Trans. Deleting RAD50 reduced the repair efficiency in both Cis and Trans to ~55% of their respective WT levels (Fig 4A). Although the kinetics of Cis and Trans repair in the absence of Rad50 looked rather similar at a first glance (S3 Fig), when we normalized the PCR assays to the total amount of product formed, we found that rad50Δ did not affect the overall kinetics of repair in either case (Fig 4B). We note that disruption of the MRX complex (as well as deletion of other tethering factors such as Sae2, γ-H2AX and Tel1) has been shown to eliminate end-tethering in only ~10–25% of the cells [25,26]; it is possible that this partial un-tethering of the DBS ends may not be sufficient to alter the repair kinetics in our assay.
In addition to facilitating the homology search by the DSB ends in Cis (as suggested above), DSB end-tethering might specifically impede repair in Trans, wherein the ability of LE and U2 ends to locate their donor templates might be affected by the SSA event involving the URA3 sequences on the other side of each DSB. For example, the URA3 end from the break on Chr XI might drag its associated LE end to the ura3 sequences on Chr V, thereby preventing LE from finding its homologous LEU2 donor on Chr III. To address this possibility, we modified the Trans strain such that it no longer harbors the URA3 SSA substrate on Chr XI (Fig 4C). Even though this strain does not generate viable repair outcomes, the LE end from Chr XI and U2 end from Chr V can still repair by GC using the LEU2 donor on Chr III, an event that can be monitored by PCR. We found that in the absence of a URA3 substrate, the kinetics of LEU2 repair in Trans were indistinguishable from those of Cis (Fig 4D). Hence, we conclude that in the original Trans strain, the LE and U2 ends are inhibited from interacting with their donor sequences because of their association, across the “divide” of the DSB, with the URA3 sequences, which are involved in a separate repair event. Conversely, the association of URA3 sequences with both the LE and U2 ends might be partially responsible for the slower kinetics of SSA repair in the Trans arrangement (Fig 2B).
In addition to the Cis and Trans strains used in the above experiments in which the 5’ truncated ura3-HOcs-URA3 cassette and the LE-HOcs-URA3 cassette for Cis and Trans strains, respectively, were inserted ~265 kb from the left end of Chr XI (referred to hereafter as the 265- Cis and Trans strains), we had constructed another set of strains in which these cassettes were inserted distal to the YKL162C locus ~147 kb from the left end of Chr XI (referred to hereafter as the 147- Cis and Trans strains). While the repair efficiencies of the 265- and 147- Cis strains were comparable, the viability of the 147-Trans strain was only ~36% as opposed to the ~61% viability of the 265-Trans strain (Fig 5A). We speculated that the much reduced repair efficiency of the 147-Trans strain could be attributed to chromosomal context, wherein the neighboring sequences might interfere with repair. Indeed, a closer examination of the YKL162C locus (the site of insertion of the LE-HOcs-URA3 cassette on chromosome XI) revealed that ~1.8 kb upstream of the LE sequences (and ~2.5 kb from the DSB end) lies the PIR3 gene, which shares about 70% homology with two other genes–PIR1 and PIR2. PIR2 is present on Chr X, but PIR1 lies ~1.6 kb further upstream of PIR3 in an inverted orientation (Fig 5B). This arrangement of PIR3 and PIR1 plus the presence of two regions (77 and 81 bp) of perfect homology between these sequences raised the possibility that if 5’ to 3’ resection from the DSB end extends ~2.5 kb, it could promote recombination between PIR3 and its homeologs; moreover, if resection extended ~6 kb, PIR3 could fold back upon PIR1 to form a stem-loop structure [37,38] and/or form SSA-mediated inter-chromatid dimers [38], both of which could preclude the LE-mediated repair in Trans. Therefore, it seemed possible that the PIR3 sequences might interact with PIR1 and/or PIR2, thereby interfering with LEU2 repair in Trans.
To test whether PIR3 sequences are indeed responsible for impeding LEU2 repair in Trans, we re-inserted the LE-HOcs-URA3 cassette at the same locus on Chr XI while deleting the adjacent PIR3 sequences. An equivalent Cis strain was made by re-inserting the 5’ truncated ura3-HOcs-URA3 cassette on Chr XI with a similar deletion of PIR3. In these pir3Δ strains, the viability of the Cis strain was ~69% and that of the Trans strain was ~54% (Fig 5A). Hence, while the PIR3 deletion had no effect on repair in Cis, it significantly improved the efficiency of repair in Trans. Furthermore, double deletion of PIR1 and PIR3 increased the viability of the 147-Trans strain to ~61%, which was indistinguishable from that of the 265-Trans strain (Fig 5A).
It has been well established that Sgs1 plays a key role in preventing homeologous recombination [39–41]. We found that deletion of SGS1 severely compromised repair in the 147-Trans arrangement (Fig 5A), presumably by further sequestering the LE end as a result of increased homeologous interactions between the PIR3 gene and its homologues. When we deleted PIR3 in the context of sgs1Δ, the efficiency of repair in 147-Trans increased from <10% in the sgs1Δ strain to ~30% in the sgs1Δ pir3Δ strain, which is still only ~50% of the SGS1 pir3Δ strain (Fig 5A). We suggest that in the pir3Δ strain, Sgs1 still limits homeologous interactions between PIR1 and PIR2 sequences. To investigate whether Sgs1 exerts its effect through the mismatch repair pathway, we deleted MSH6 in the 147-Trans strain. Unlike sgs1Δ, deletion of Msh6 only mildly affected the efficiency of 147-Trans repair, and the latter was completely rescued by the deletion of PIR3 (Fig 5A). Overall these results argue that interference from the neighboring PIR3 and PIR1 sequences is largely responsible for the much-reduced efficiency of repair in the 147-Trans strain. Surprisingly, strand-invasion of the LEU2 donor by the LE and U2 ends (as assessed by a Rad51 chromatin immunoprecipitation assay) appeared to occur with similar efficiency and kinetics in Cis and Trans (Fig 5C) suggesting that the PIR sequences may not affect synapse formation per se. Instead, this result argues that 5’ to 3’ resection continues even after the initial contact of the DSB end with its homologous donor sequences. A similar interpretation can be drawn from the data of Wu et al. [42] or Agmon et al. [29], who found that ectopic gene conversion could be followed by an intrachromosomal SSA event in which 5’ to 3’ resection continued into the flanking homologous region, past the sequences involved in GC.
Although the adjacent PIR3 and PIR1 sequences interfere with GC-mediated LEU2 repair in the Trans strain, they do not affect the overall efficiency of repair in the 147-Cis strain (Fig 5A). Since the PIR3 and PIR1 sequences are present at the same distance away from the ura3-HOcs-URA3 cassette in the 147-Cis strain as they are from the LE-HOcs-URA3 cassette in the 147-Trans strain, the data from WT and pir3Δ 147-Cis strains imply that these PIR sequences do not interfere with intrachromosomal SSA. This difference is perhaps due to the much faster kinetics of SSA compared to GC, which may not allow enough time for resection to continue past the PIR3 sequences before the completion of SSA.
The results presented above as well as previous studies have suggested that sequences distant from the DSB can participate in competing recombination events thereby interfering with recombination of sequences closer to the break [43]. In another assay system, we found that the presence of a nearby copy of a family of dispersed, repeated sequences impaired the efficiency of an HO-induced SSA event. We have previously described strain YMV80, in which an HO-induced DSB is created within the leu2 gene (at its original location on Chr III). This break is repaired by SSA between the U2 sequence next to the cleavage site and a ~1 kb U2 sequence inserted at the his4 locus ~25 kb distal to the HOcs, resulting in a 25 kb deletion of the nonessential sequences between the U2 repeats [30]. Repair takes approximately 6 h, consistent with the amount of time required to resect 25 kb of DNA at ~4 kb/h rate of resection. Distal to the LEU2 locus and between the flanking U2 segments lies a complete Ty2 retrotransposon element plus several LTRs (long terminal repeats), as well as two tRNA genes (Fig 5D). In the presence of these repeated sequences, the efficiency of U2-mediated SSA repair, as measured by a viability assay, was found to be 76±10%. However, when these intervening repeated sequences were deleted (Fig 5D, TyΔ), the repair efficiency increased to 100±15%.
To explore this further, we also deleted a short region to the left of Ty2 containing an inverted repeat (IR) of the Ty1 LTRs and a tRNA gene. However, this deletion (Fig 5D, IRΔ) did not affect the efficiency of repair (81±3%) relative to the parent strain. These observations suggest that resection into the Ty2 element allows it to recombine with some of the other ~30 copies of Ty elements and even more LTRs dispersed throughout the genome, most of which would yield inviable outcomes. Thus, removing these potentially competing sequences greatly enhances repair by SSA and provides another example of impediment of DSB repair by nearby repeated sequences.
We have previously shown that a Recombination Execution Checkpoint regulates the choice of the HR pathway based on whether the two ends of a DSB strand-invade closely positioned and correctly oriented homologous donor sequences [13]. We reported that increasing the distance between the homologous donors (while the site of DSB induction was kept constant) resulted in REC-mediated shift from GC to BIR mode of repair. Here we tested whether REC can also detect the origin of the ends that are engaged in repair, since repair involving ends from different breaks can produce deleterious chromosomal translocations. We therefore examined the effect of separating the origin of the DSB ends while the donors were kept constant and positioned next to each other in the correct GC-compatible configuration. We found that involvement of ends from two different breaks did not compromise the overall efficiency of repair (Fig 2A). However, it slowed down the kinetics of repair quite dramatically as we observed a ~2 h delay in the appearance of product in the Trans and Reverse-Trans case relative to the Cis and Reverse-Cis configurations (Fig 2B–2D). Contrary to our expectation, we found that just like the Cis case, repair in Trans was largely Pol32-independent (Fig 2A) indicating that although it is kinetically slower, repair in Trans still proceeds via GC. Hence, the REC cannot distinguish between the origins of the synapsed ends and does not specifically restrict GC-mediated repair of ends belonging to two different breaks. This finding argues that the while physical distance separating the synapsed ends is a critical parameter that governs the REC-mediated switch between GC and BIR, if the DSB ends are synapsed close to each other (in the correct orientation) on the donor template, the GC requirement is fulfilled and the origin of the ends has no relevance. This observation supports our previous hypothesis that the REC acts after the strand-invasion step whereby it prevents loading of the GC machinery to single-end invasions, or if the ends have invaded far away from each other, or in the wrong orientation. We conclude that when two DSB ends, irrespective of their origin, are synapsed close to each other in the correct orientation, the REC allows the GC machinery to carry out repair before BIR can come into play.
However, we did observe slower kinetics of repair in Trans. Our data indicate that this delay is attributable to DSB end-tethering. Using fluorescently tagged arrays, it has previously been shown that DSB ends remain tethered together up to several hours after induction of an irreparable break in a vast majority of the cells [24–27]. This tethering is partially dependent upon the MRX complex, as well as Sae2, Tel1 and Rad52 proteins. While end-tethering has been shown to play a very important role in preventing NHEJ-mediated chromosomal translocations, its role in GC has remained elusive because these proteins are pleiotropic and affect many other aspects of repair as well, such as G2/M checkpoint activation, 5’ to 3’ end resection and DSB-induced sister-chromatid cohesion [33–36]. Therefore, it is difficult to ascertain whether or not a repair defect observed upon disruption of these proteins is attributable to the loss of end-tethering per se. We reasoned that our Trans strain would overcome this problem because the ends that are involved in GC-mediated repair don’t belong to the same break, and therefore, would not be tethered to each other (although they would presumably be tethered to the URA3 sequences on the opposite sides of the DSBs). Since the cells are able to survive a break in the Trans and Reverse-Trans configuration as efficiently as a break in the Reverse-Cis configuration (in which the ends involved in repair should be tethered) (Fig 2A), DSB end-tethering seems to be largely dispensable for the successful completion of GC-mediated repair. Nevertheless, the delay in the kinetics of repair in Trans and Reverse-Trans might reflect the lack of association between LE and U2 ends, which would have to independently seek out their homologous sequences, as opposed to repair in the Cis and Reverse-Cis configurations where the LE and U2 ends would be tethered together and might therefore engage in a coordinated search for their homologous partners. Alternatively, the slower kinetics of LEU2 repair in Trans and Reverse-Trans could be an indirect effect of tethering of the LE and U2 ends to the corresponding URA3 sequences on the other side of the DSBs, which are involved in a separate SSA repair event and might therefore drag the associated LE and U2 ends away from their homologous donor. While RAD50 deletion reduced the efficiency of both Cis and Trans repair, it did not alter their relative kinetics (Fig 4A and 4B). We suspect that this is due to only partial un-tethering of the DSB ends in the absence of Rad50 [25,26]. Indeed, when we removed the influence of association to the other end by eliminating the SSA event in the Trans setting, the kinetics of Trans GC became as rapid as in the Cis case (Fig 4D). These data argue that in the original Trans strain, the URA3 sequences that are involved in SSA-mediated repair, potentially drag their associated LE and U2 ends away from the LEU2 donor on Chr III. This, in turn, would compromise the probability of LE and U2 ends simultaneously engaging the LEU2 donor, resulting in a smaller proportion of cells initiating GC-mediated repair at any given time. Overall these results suggest that end-tethering may play a substantial role in preventing chromosomal translocations that could arise from GC repair involving ends from different breaks. This might become particularly important if a break occurs within a repeated sequence such that the ends could either travel together and repair using donors present in the vicinity of each other or engage in separate GC events involving unlinked donors.
We did not see a difference in the kinetics or efficiency of strand-invasion of the LEU2 donor between our 147- Cis and Trans strains (Fig 5C); however, ChIP gives a bulk estimate and does not tell us the proportion of cells in which both ends are simultaneously synapsed with the homologous donor–a parameter that seems to play a critical role in regulating the initiation of repair synthesis.
We also found that sequences adjacent to those engaged in an HR event have a profound effect on the outcome. This is most evident in the case of the 147-Trans strain in which the adjacent PIR3 sequences greatly interfere with GC-mediated repair (Fig 5A). If 5’ to 3’ resection is extensive before (or maybe even after) LE succeeds in finding its homologous donor, single-stranded DNA at PIR3 may fold back upon PIR1 (which lies just ~1.6 kb upstream of PIR3 in the inverted orientation) and/or it may engage in some other homeologous interactions with PIR1 or PIR2, with which it shares ~70% homology, thereby driving the LE end away from its GC template. These PIR3 interactions can at least partially be disrupted by Sgs1 (Fig 5A), which has previously been shown to prevent gross chromosomal rearrangements (GCRs) between homeologous sequences [39,41].
Because Sgs1 has roles in processes other than mismatch repair, we also deleted MSH6 to impair heteroduplex rejection during interactions between PIR3 and PIR1/PIR2. The much weaker effect of msh6Δ on the viability of the 147-Trans strain may be due to its redundancy with Msh2-Msh3 in heteroduplex rejection. We did not test msh2Δ and msh3Δ mutants due to their essential role in removal of the non-homologous tails during SSA. Moreover, Sgs1 has been similarly reported to discourage spontaneous interchromosomal recombination among a set of highly diverged genes (CAN1, LYP1 and ALP1) independently of MSH6 and MSH2. The overall divergence of CAN1 and LYP1 or ALP1 is very high (64% sequence identity) and many recombination events occurred in a stretch of DNA that shared ~74% sequence identity, which is comparable to the sequence similarity between PIR genes. Hence, we conclude that homeologous interactions between PIR genes are responsible for impeding LEU2 repair in the 147-Trans strain.
Nearby repeated sequences also affect long-range SSA. Unlike GC, SSA will delete the repeated sequences in the intervening region in surviving cells; yet they still influence the outcome. It is possible to envision several possible modes of impairment. The two ends of the DSB may be tethered so that when the Ty sequences engage another repeated sequence, the U2 sequences may be dragged along, thus restricting its interaction with the second U2. It is also possible that the Ty recombines with another Ty and generates an inviable chromosomal configuration. Alternatively, the repeated sequences may generate an intermediate that prevents DNA resection or in some manner interferes with SSA. We note that deletion of the Ty element results in net deletion of ~4.7 kb in the 25-kb region between the U2 repeats (see Methods). We think it is unlikely that this change in the separation of the U2 sequences could be responsible for the increase in repair efficiency from ~76% to ~100%, because increasing the distance between the U2 repeats from ~25 kb to ~30 kb was not found to reduce the efficiency of SSA [30].
The effect of chromosomal context may also explain the difference in repair between the Cis and Reverse-Cis strains. Recent studies by Agmon et al. [44], emphasize that sequences at some chromosome locations recombine more readily with other sequences located at equivalent sites in the yeast nucleus. Interference from Ty elements has been seen before [38,45,46] and it has been shown that sequences further away from a break can drive repair as efficiently as the sequences immediately flanking the DSB [43]. Such interference from neighboring sequences can pose a big challenge for HR-mediated repair, especially in mammalian cells, which possess repeated sequences throughout their genome.
All strains were derived from YSJ119 (ho hmlΔ::ADE1 mataΔ::hisG hmrΔ::ADE1 leu2::KAN ade3::GAL::HO ade1 lys5 ura3-52 trp1 can1::LE-HOcs-U2 LEU2 at position 41400 of Chr III [13]. YSJ159 was constructed by replacing the endogenous ura3-52 in YSJ119 with HPH, and YSJ352 was obtained by replacing leu2::KAN in YSJ159 with leu2::LYS5. The Cis strain YSJ357 was constructed by inserting a 5’ truncated ura3-HOcs-URA3 cassette (from pSJ20) at position 265602 on Chr XI in YSJ352. The Trans strain YSJ379 was constructed in several steps. First, the LE portion of the LE-HOcs-U2 cassette at the can1 locus on Chr V in YSJ159 was replaced with a NAT-5’ truncated ura3 sequence (from pSJ18) to obtain YSJ165. Next, leu2::KAN in YSJ165 was replaced with leu2::LYS5 to obtain strain YSJ353. Finally, an LE-HOcs-URA3 cassette (from pSJ17) was inserted at position 265602 on Chr XI in YSJ353 to obtain YSJ379. Plasmid pSJ18 was constructed by inserting an 800 bp long 5’ truncated ura3 sequence between the SacI and SpeI sites in pAG25. pSJ20 was constructed by inserting an HOcs-URA3 fragment (from pSJ17) at the SpeI site in pSJ18. pSJ17 was obtained by inserting URA3 at the AgeI site in pJH1386 (which carries the XhoI-SalI leu2 fragment containing the 117 bp HOcs at the Asp718 site). The Reverse-Cis strain (tNS2614) was derived from YSJ352 in several steps. First, the can1::LE-HOcs-U2 cassette on Chr V was replaced with the 5’ truncated ura3-HOcs-URA3 cassette (from pSJ20) to obtain tNS2563. A NAT-LE-HOcs-U2 cassette derived from plasmid pNSU283-3 was inserted at position 265602 between BUD2 and MBR1 on Chr XI in tNS2563 to obtain tNS2605. Subsequently, the NAT marker was swapped with TRP1, and the trp1 allele on Chr IV was replaced with trp1::NAT to create the Reverse-Cis strain. Plasmid pNSU283-3 was constructed by inserting LE-HOcs-U2 into pAG25 [47] between the EcoRI and SpeI sites and chr XI sequences were added to each side of the LE-HOcs-U2 cassette using the Gibson Assembly method [48]. The reverse-trans strain, tNS2638, was constructed by first PCR amplifying the NAT-ura3-HOcs-U2 cassette from YSJ379 and inserting it at position 265602 between BUD2 and MBR1 on Chr XI in YSJ352. The second cassette, LE-HOcs-URA3, was amplified from YSJ379 and inserted at the can1 locus on Chr V. The modified Cis strain, tNS2607, was derived from YSJ159 by inserting the HOcs-URA3 cassette at position 165957 between MBR1 and BUD2 on Chr XI. To construct the modified Trans strain, tNS2628, we replaced the NAT-ura3 portion of the NAT-ura3-HOcs-U2 cassette on ChrV in YSJ379 with TRP1 and subsequently replaced the trp1 allele on Chr IV with trp1::NAT. The 147-Cis strain (YSJ176) was derived from YSJ159 by inserting a 5’ truncated ura3-HOcs-URA3 cassette (from pSJ20) at position 147172 on ChrXI. The 147-Trans strain (YSJ175) was constructed by inserting an LE-HOcs-URA3 cassette (EcoRI-BamHI fragment from pSJ19) at position 147172 on ChrXI in YSJ165. pSJ19 was constructed by inserting the LE-HOcs-URA3 fragment (from pSJ17) at the HindIII site in pSJ16 (pBR322 carrying ChrXI sequences from positions 146537 to 147378 between EcoRI and BamHI sites). YSJ194 and YSJ195 were derived from YSJ175 and YSJ176, respectively, by replacing the leu2::KAN cassette with LYS5. YSJ363 (pir3Δ version of the 147-Cis strain) was made by inserting the 5’ truncated ura3-HOcs-URA3 cassette (from pSJ20) between positions 144273 and 147176 on ChrXI in YSJ352. YSJ377 (pir3Δ version of the 147-Trans strain) was made by inserting the LE-HOcs-URA3 cassette (from pSJ17) between positions 144273 and 147192 on ChrXI in YSJ353. tNS2646 (pir1Δ pir3Δ version of the 147-Trans strain) was made by inserting the LE-HOcs-URA3 cassette (from pNSU290) between positions 141790 and 145384 on ChrXI in YSJ353. pol32Δ, rad50, and sgs1Δ strains were made by the standard PCR-based gene disruption method to obtain strains YSJ398 (YSJ379 pol32Δ), YSJ399 (YSJ357 pol32Δ), YSJ382 (YSJ379 rad50Δ), YSJ383 (YSJ357 rad50Δ), YSJ226 (YSJ194 sgs1Δ), YSJ227 (YSJ195 sgs1Δ), YSJ384 (YSJ377 sgs1Δ) and YSJ385 (YSJ363 sgs1Δ). The msh6Δ mutants were made by oligo-directed transplacement to create strains tNS2648 (YSJ194 msh6Δ) and tNS2650 (YSJ377 msh6Δ). YMV80 has been described before [30]. We first deleted the sequences between KCC4 and LE-HOcs-U2 in the YMV80 parent strain, YFP17 [30], using plasmid pNSU262 to obtain strain tNS2326. pNSU262 was derived from pAG32 [47] and harbors sequences flanking KCC4 and LEU2 to target a 8.8 kb deletion of Chr III encompassing the Ty element. Transplacement of pNSU262 results in addition of 4.1 kb of vector sequences resulting in a net deletion of 4.7 kb. A TRP1-U2 was then inserted at the at the his4 locus in this strain to create the strain tNS2333. tNS2427, possessing the deletion of the inverted repeat sequence, was obtained by oligo-directed transplacement of a PCR product using pNSU262 (kcc4-HPH) as the template and a target sequence located to the left of YCL021W-A.
Yeast cells were grown in YEP containing 2% raffinose to a density of ~1x107 cells/ml. Equal volumes of appropriate dilutions were plated on YEP containing 2% galacotse (YEPGal; to induce the HO break) and YEP containing 2% dextrose (YEPD; no DSB control). Viability was determined from the ratio of CFUs able to survive the break (number of colonies that grew on YEPGal) to the total number of CFUs plated (number of colonies that appeared on YEPD). Data represent mean ± S.D. (n ≥ 5), and p values were calculated using the Student’s t-test.
Yeast cells were grown in YEP containing 2% raffinose to a density of ~1x107 cells/ml and HO endonuclease was induced by adding galactose to a final concentration of 2%. Samples were collected for DNA analysis just prior to and at different time points following addition of galactose, as described before [49].
Equal amounts of genomic DNA isolated from samples collected at different time points were PCR-amplified (primer pairs are listed in S1 Table) within linear range as described before [50]. A pair of primers 300 bp upstream of ura3 and 700 bp (for Cis and Trans) or 256 bp (for Reverse-Cis and Reverse-Trans) downstream of URA3 was used to analyze the URA3 repair kinetics. A primer pair with one primer downstream of HOcs-U2 and another primer 150 bp (for Cis) or 650 bp (for Trans) upstream of the LE-HOcs sequence was used to analyze the kinetics of LEU2 repair. For the Reverse-Cis strain the primer pairs were located 22 bp upstream of LE-HOcs and 100 bp downstream of U2-HOcs, and for the Reverse-Trans strain, the primers were located 150 bp upstream of LE-HOcs and 100 bp downstream of U2-HOcs. The PCR reactions were run on agarose gels and the repair product was quantified using Bio-Rad Quantity One software. For the Cis and Reverse-Cis strains, the primer pairs used to study LEU2 and URA3 repair give bands at 0 h time points as well; however, those bands are bigger than the ones obtained after repair (owing to presence of the 117 bp HOcs and the HOcs plus an additional URA3 repeat, respectively) and were not used in our analysis. PCR signal from an independent locus (SLX4) was used to normalize for input DNA. The ratio of test and reference signals obtained from the DNA of a repaired colony was set to 100%.
A small quantity of cells taken from colonies growing on YEPGal (from the viability assay described above) were heated at 100°C for 10 minutes and the lysates were amplified with primers listed in S2 Table to assay for the loss of the distal arm of Chr V and appearance of the BIR product (as diagramed in S2 Fig).
Rad51 ChIPs were performed as described [7]. The IP signal from the donor locus was normalized to the IP signal from the CEN8 locus, which was immunoprecipitated using anti-Mif2 antibody.
|
10.1371/journal.pntd.0004147 | What Is Needed to Eradicate Lymphatic Filariasis? A Model-Based Assessment on the Impact of Scaling Up Mass Drug Administration Programs | Lymphatic filariasis (LF) is a neglected tropical disease for which more than a billion people in 73 countries are thought to be at-risk. At a global level, the efforts against LF are designed as an elimination program. However, current efforts appear to aim for elimination in some but not all endemic areas. With the 2020 goal of elimination looming, we set out to develop plausible scale-up scenarios to reach global elimination and eradication. We predict the duration of mass drug administration (MDA) necessary to reach local elimination for a variety of transmission archetypes using an existing model of LF transmission, estimate the number of treatments required for each scenario, and consider implications of rapid scale-up.
We have defined four scenarios that differ in their geographic coverage and rate of scale-up. For each scenario, country-specific simulations and calculations were performed that took into account the pre-intervention transmission intensity, the different vector genera, drug regimen, achieved level of population coverage, previous progress toward elimination, and potential programmatic delays due to mapping, operations, and administration.
Our results indicate that eliminating LF by 2020 is unlikely. If MDA programs are drastically scaled up and expanded, the final round of MDA for LF eradication could be delivered in 2028 after 4,159 million treatments. However, if the current rate of scale-up is maintained, the final round of MDA to eradicate LF may not occur until 2050.
Rapid scale-up of MDA will decrease the amount of time and treatments required to reach LF eradication. It may also propel the program towards success, as the risk of failure is likely to increase with extended program duration.
| Lymphatic filariasis (LF) is a disease caused by filarial worms transmitted by different types of mosquitos that can lead to massive disability, including elephantiasis and hydrocele. LF has no significant zoonotic reservoir and is thought to be a potentially eradicable disease through once yearly treatment distributed by mass drug administration (MDA). In this study, we set out to determine how many treatments and over how much time it might take to globally eliminate and eradicate LF under different levels of treatment intensities. We created a model that took into account country-specific and disease-specific variables, and found that if the current intensity of MDA is maintained, 3,409 million treatments distributed over the next 37 years will be required. However, if treatment is rapidly expanded to the entire at-risk population in all endemic countries, eradication could be achieved with 4,159 million treatments and in less than half the time. While our estimates suggest more time may be needed to reach LF elimination than what is currently projected, with continued commitment, eradicating LF is within reach.
| Lymphatic filariasis (LF) is a neglected tropical disease (NTD) primarily prevalent in poor populations in 73 countries [1]. LF is caused by infection with Wuchereria bancrofti, Brugia malayi, or B. timori transmitted by a variety of mosquito genera [2]. Infection with the filarial nematodes can damage the lymphatic vessels, the main clinical manifestations being lymphedema, hydrocele, and elephantiasis [3]. In addition to disfigurement and disability, people affected by LF face stigma, social adversity, and economic hardship [4–6].
LF is spread by mosquitoes that take up circulating microfilarae (mf) in the peripheral blood of infected humans [7]. Administration of albendazole with ivermectin or diethylcarbamazine citrate (DEC) has been shown to reduce circulating mf to such low levels that transmission cannot be sustained [8]. For this reason, LF is one of six diseases considered to be potentially eradicable [9]. Accordingly, in 1997 the World Health Assembly (WHA) adopted resolution WHA 50.29, which calls for the elimination of LF as a public health problem and, in 2000, the World Health Organization (WHO) established the Global Program to Eliminate Lymphatic Filariasis (GPELF). The GPELF aims to eliminate LF in all endemic countries by 2020 through annual mass drug administration (MDA) maintained over multiple years [8]. The program benefits through donations from Merck & Co. and GlaxoSmithKline (GSK), which have pledged to provide enough ivermectin and albendazole, respectively, to achieve elimination, as well as from Eisai, which in 2010, pledged 2.2 billion DEC tablets [10, 11].
The GPELF has scaled up rapidly and is among the fastest growing disease elimination programs in the world [12]. By the end of 2013, 56 LF-endemic countries had carried out MDA, of which 15 are now undertaking post-MDA surveillance. In 2013 alone, more than 410 million anti-filarial treatments were distributed under the GPELF. However, the program is not without its challenges: mapping is incomplete in 12 countries, 14 countries requiring MDA are yet to begin, and many of the other endemic countries are targeting relatively small proportions of their at-risk populations [13]. Issues with compliance, contraindications of ivermectin and DEC in areas with hyper Loa loa-endemicity, and interruptions in funding also plague the program [14, 15]. At a global level, the efforts against LF could be considered a global elimination program (elimination of infection in some but not all countries) as the name suggests, or an eradication program (permanent reduction to zero of the worldwide incidence of infection) as implied by the stated aims of the program [13, 16, 17].
In order to assist decision makers in determining whether efforts for LF should be scaled up to try to achieve eradication, it has been proposed to use an analytic and deliberate methodology to produce evidence-based guidance on the rationale for investing [18, 19]. As part of this endeavor, we herein predict the duration of MDA necessary to reach local elimination for a variety of transmission archetypes using an existing model of LF transmission, outline plausible scale-up scenarios leading to global elimination and eradication, and estimate the number of treatments required under each scenario. Potential delays in implementation, previous progress, and different intensities of infection and transmission are also taken into account. Studies on the economic and financial costs, the impact on disease burden, and cost-effectiveness of these scenarios are to be published as companion papers.
We have defined four hypothetical scenarios that differ in their geographic coverage and rate of scale-up. The global elimination scenario represents the case whereby countries continue with current practices. As such, it serves as the comparator against all other scenarios. The other three scenarios aim at reaching LF eradication through varying levels of MDA scale-up. Key assumptions and differences between the scenarios are outlined in Table 1. The number of years that each endemic country exceeded the minimum effective coverage rate of 65% in previous rounds of MDA, as well as the geographic coverage and rates of scale-up are provided in Table 2 (countries without previous rounds of MDA for LF) and Table 3 (countries that previously carried out MDA for LF). All scenarios were assumed to begin in 2014 and run until the final round of MDA has been distributed in each country under consideration. Though coverage rates above 65% are considered to be the lowest threshold necessary to be effective, the average programmatic coverage for countries that had previously achieved effective coverage was over 80%. Therefore, we presume that prospective MDA will continue to be performed at higher levels, and therefore assume MDA coverage to be fixed at 85%.
Scenarios were developed by first reviewing the WHO preventive chemotherapy (PCT) databank to assess progress made towards LF elimination as of 2012 [13]. The scenarios were further refined, with key assumptions agreed upon in a series of technical advisory group meetings, including stakeholders from WHO, Centers for Disease Control and Prevention (CDC), funders, pharmaceutical companies, and program managers from endemic countries.
In the global elimination scenario, countries that have not yet started will not start, and countries that have started continue according to their assigned level of scale-up (see: Rate of scale-up). In the eradication I scenario, countries that have already started MDA continue as in the global elimination scenario and countries that have not yet started implement MDA following an ‘average’ level of scale-up. The eradication II scenario represents the case in which all countries scale-up MDA more quickly (fast). Eradication III serves as the ‘best case’ scenario, whereby all endemic countries provide MDA to their entire at-risk populations immediately. Thus, this analysis provides insight into the differences in the amount of time and treatments required to extend elimination efforts to all endemic countries (eradication I), increase MDA intensity (eradication II) and, most ideally, scale-up instantaneously (eradication III).
An important assumption underlying this study is that annual MDA using DEC with albendazole, or, in onchocerciasis-endemic countries, ivermectin and albendazole, will be sufficient to reduce circulating mf enough to interrupt the transmission cycle of LF if maintained for an appropriate number of years. Therefore, hardly predictable features that could undermine success, including systematic non-compliance with MDA, but particularly events such as civil unrest and humanitarian emergencies (e.g. earthquakes in Haiti and Nepal; Ebola epidemic in West Africa) that could compromise the health system’s capacity, could not be accounted for. We also assume that countries undertake MDA without interruption.
Administration of ivermectin to communities with high prevalence (>40%) of L. loa is contraindicated, as the microfilaracidal actions of the drug poses an unjustifiably high risk of causing severe adverse events. As such, the WHO provisionally recommends the LF program to instead treat these areas with albendazole monotherapy distributed bi-annually and vector control [20]. Here we assume that this strategy will be equally efficacious as annual albendazole-ivermectin, and thereby assume the number of years of MDA required in areas co-endemic with L. loa to be equivalent to the number of years required with albendazole-ivermectin.
The GPELF advises LF endemic countries to conduct MDA for 4–6 years [8]. This duration only holds at a country level if all endemic areas are treated simultaneously. To incorporate scaling-up of geographic coverage for each scenario, we divided each country’s at-risk population into deciles, and assumed MDA to start in subsequent deciles after varying durations according to four schedules of scale-up. In schedule I (fast), 20% of the at-risk population is added to the MDA schedule annually. In schedule II (average), one decile is added each year, in schedule III (slow) one decile is added every two years and in schedule IV (very slow) this period is three years.
In the global elimination scenario, scale-up is based upon the proportion of the at-risk population each country previously targeted. In order to be allocated to schedule I, the at-risk population targeted in the most recent round of MDA had to exceed 50%. Schedule II has been assigned to countries previously targeting 30–50%, schedule III to those targeting 20–29.9%, and schedule IV to those targeting <20%. Rather than attempting to recreate the progress of each country exactly, we used these categories to incorporate a range of scale-up levels encountered. Previous progress made towards local elimination was further taken into account by counting the number of previously effective years of MDA, which was considered as any year in which program coverage within the targeted area (regardless of the at-risk population targeted) exceeded 65%. We then subtracted the number of effective years previously achieved from the number of years of MDA deemed necessary (see below: Transmission Archetypes; Table 4) in order to determine the number of years of MDA remaining.
The number of rounds corresponds to the minimum at which at least 97.5% of simulations went to elimination.
For all scenarios, we assume that countries that have finished mapping but not begun MDA have a 1-year delay, whereas countries that have not completed mapping nor begun MDA have a 4-year delay. While countries face challenges of different magnitudes and require different durations to map, the 4-year delay assumed corresponds to the average number of years that mapping took in countries with available data to support the calculation [13].
To account for heterogeneity in transmission intensity within countries, we obtained paired baseline circulating filarial antigenaemia prevalence, measured through immunochromatographic tests (ICTs), and mf prevalence data from sentinel site surveys from program countries across the AFRO region. As specified by the WHO, these surveys involve collecting fingertip blood, between 10 p.m. and 2 am. from at least 300 participants aged five years and above [21]. We gained additional access to ICT prevalence data from mapping studies in 17 African countries. The relationship between mf and antigenaemia prevalence was estimated using the non-parametric regression proposed by Passing and Bablock, which assumes linearity and uncertainties in both variables [22]. The regression equation calculated from the paired prevalence data was then used to infer mf prevalence from the ICT mapping data.
We determined the percentage of the at-risk population that fell into prevalence quartiles: <5%, 5–10%, 10.1–15%, >15%, for each country that provided district level prevalence data. To account for uncertainties in this approach, we took 500 random draws from a multinomial distribution with probabilities based on weighted averages from the dataset and assumed these to be the possible ranges of pre-intervention prevalence distributions for all countries in our analysis.
It has been theoretically demonstrated that the required duration of MDA is region-specific and dependent on various factors, including drug regimen and level of coverage, vector species, and pre-intervention transmission intensity [23–25]. In order to broadly capture the heterogeneous transmission patterns of LF, we defined transmission archetypes (Table 4). In addition to prevalence levels and drug regimens, we accounted for differences in transmission between Anopheles spp. and Culex spp., which notably differ in their mf-density dependent likelihood of becoming infected [26]. Predicting regional anopheline- or culicine-mediated LF transmission has been shown to require different model formulations and parameterizations [27]. For our analysis we made several simplifications: we assumed transmission of W. bancrofti by Aedes spp. was similar to transmission efficacy by Culex spp., while transmission of Brugia spp. was assumed to be comparable to W. bancrofti transmission by Anopheles spp. Where the primary vector was unclear, infection by Culex spp. was assumed in order to avoid underestimating the number of MDA rounds required.
The duration of MDA required to eliminate LF was predicted for the transmission archetypes using a deterministic model of LF transmission, EpiFil [28]. The model used for the current analysis has been described in detail, validated against multiple data sets for both transmission settings with Anopheles spp. and Culex spp., and used extensively to predict LF intervention outcomes [28–31]. Details on model structure, equations, and the approach to obtaining parameter estimates are provided in Supporting Text 1: LF model description.
For all transmission archetypes, we ran 500 simulations of once-yearly MDA of varying total durations, drawing from a range of parameter estimates. The lowest number of rounds at which the 95th percentile range of the simulations resulted in an mf prevalence below 1% 50 years after the start of the MDA program was taken as a conservative measure of the number of rounds required to ensure elimination.
Population at-risk figures were taken from the WHO PCT database for 2012 and adjusted for population growth using country-specific 2012 United Nations estimates [13, 32]. MDA coverage rates were assumed to be 85% for all countries. Except for areas co-endemic with L. loa, treatments are assumed to occur annually. Based on the pre-intervention prevalence distributions, we developed 500 estimates of the number of treatments needed for each country and scenario. Results are reported as the mean number of treatments by region and scenario, along with 95% credible intervals (CI).
Our results indicate that interrupting LF transmission in all countries by 2020 is unlikely, though if MDA is drastically scaled-up and expanded, the final round of MDA to eradicate LF could be carried out by 2028 (eradication III; Fig 1). If scale-up continues at the current rate, as modeled in our global elimination and eradication I scenarios, the last round of MDA will not be given until 2050, largely due to slow scale-up in areas where transmission occurs through Culex spp. The eradication II scenario reaches the last round of MDA by 2032. As this scenario assumes that all countries add 20% of their at-risk populations to MDA annually, the last countries to reach local elimination are those that were delayed due to mapping, and whose vector and treatment combination included Anopheles spp. and ivermectin or Culex spp. and DEC, including: Angola, Chad, the Democratic Republic of Congo, South Sudan, Sudan, Zambia, and Zimbabwe. Fig 2 provides a visual representation of the impact different intensities of scale-up and expansion have on time to reach local elimination for each country.
Since the scenarios take into account population growth, rapid scale-up of MDA also decreases the number of treatments required. As depicted in Fig 3, the eradication III scenario initially requires substantially more treatments, but by 2024, the treatments under this scenario are projected to be less than that required under all other scenarios. The global elimination scenario is projected to require approximately 3,409 million treatments (95% CI: 3,185m–3,538 million). Expanding the program to all endemic countries will increase the number of treatments to 4,666 million (95% CI: 4,419m–4,904 million). Scaling up MDA more rapidly, as under the eradication II scenario, results in savings of nearly 300 million treatments compared to the eradication I scenario. Under the most optimistic scenario (eradication III), eradication could be achieved with 4,159 million treatments (95% CI: 3,924m–4,382 million). As shown in Fig 1, this represents nearly 750 million treatments more than the global elimination scenario but 210 million treatments less than the intensified eradication scenario (eradication II). Owing to the largest burden, the AFRO region requires the majority of treatments, followed by Southeast Asia. With the shift from global elimination to eradication, the number of treatments required in the Eastern Mediterranean region increases by more than 380 fold due to treatments required for Sudan, which is not considered under the elimination scenario (Table 5).
As not all LF endemic countries are considered under the global elimination (comparator) scenario, any eradication campaign will require a massive increase in treatments. However, if LF is to be eliminated in all endemic countries, then rapid scale-up as soon as possible will lead to increased savings—both in terms of time and treatments. Accelerated MDA may also propel the program towards success, as the risk of failure (due to lapses in funding, donor fatigue, or occurrence of calamitous events) potentially increases with extended program duration [33]. It is conceivable that a decrease in program duration may also decrease the likelihood of drug resistance evolution [34].
Noticeably missing from our analysis is India. While India has the greatest burden of LF [35], it has made substantial progress against the disease, having distributed nearly 3.5 billion antifilarial treatments since 2001 [13]. As such, our model suggests that further rounds may not be necessary for India. However, previous studies have found pockets of systematic non-compliance in India, leading to MDA coverage in those areas to fall below effective coverage [36]. It is therefore possible that transmission of LF may still occur in India. However, in order to remain consistent in our approach, and in recognizing that to provide global estimates we cannot take into account all eventualities, additional treatments for India have not been considered.
We sought data from a number of diverse sources. Due to the inherent structure of the LF program, however, our analysis relies heavily on data that have been collected and reported directly by each country. While this arrangement raises a number of issues, discrepancies in the data could also decrease the validity of our estimates. Inconsistencies in coverage data may affect the number of years required to interrupt transmission, while inaccuracies in at-risk estimations would directly impact the number of treatments projected to reach our scenario endpoints. Whether these issues would result in underestimates or overestimates is dependent upon the direction and magnitude of the error.
While we avoided underestimating scale-up potential through our eradication III scenario, it is possible that we overestimated the capacity of some countries to scale-up. It is possible that we also overestimated the effectiveness and ability to proceed with rapid scale-up in areas co-endemic with L. loa. While WHO has provisional guidelines for dealing with LF and L. loa co-endemicity, no such areas have been broadly targeted for LF elimination as yet, and thus the effectiveness and feasibility of the strategy remains unclear. At the same time, the mass distribution of long-lasting insecticidal nets (LLINs) in many malaria endemic sites is likely to have a large impact on LF transmission by anophelines [37, 38]. Because the impact remains difficult to quantify, and uncertainty remains regarding the duration LLINs have to remain in place, we have not included this here. The time and treatment estimates in this study are based on data and model formulations and parameterizations currently available to the authors. Many of the assumptions and simplifications inherent to our scenarios are in need of closer investigation. Ideally, models would be fit to specific transmission settings within and between countries, as parameter values have been shown to differ by region [29]. Other aspects equally deserving of more attention, but likewise beyond the scope of this project, are the effectiveness of twice-yearly albendazole in concert with vector control for areas co-endemic with L. loa, and the consequences of mid-program delays, [39, 40]. Care should thus be taken when interpreting these results, particularly at a country-specific level.
Our duration estimates are considerably longer than those proposed under the GPELF, which envisages all endemic countries to reach full geographic coverage by 2016, with post-MDA surveillance in all countries anticipated by 2020 [17]. While this level of scale-up is similar to that proposed under our eradication III scenario, we project the last round of MDA to occur nearly a decade later, in 2028. This divergence arises from differences in the assumed number of rounds of MDA required to interrupt transmission. Depending on baseline prevalence and vector-treatment combinations, our model estimates interruption in transmission to occur after 6–15 rounds of MDA (Table 4). In contrast, the GPELF assumes five years of MDA in all areas [17]. It is worth noting that the durations in this study represent a potentially conservative measure, as they were based on the 95th percentile range of simulations leading to elimination, accounting for the uncertainty in our parameter estimates. This measure was taken to represent the time that could guarantee elimination with a reasonable level of certainty, but does not preclude that shorter durations may be sufficient in many areas. However, the discrepancy between predicted MDA durations and those advocated by GPELF was also evident in previous estimates with both deterministic and stochastic LF transmission models [41]. While aggressive goals for disease elimination and eradication potentially propel campaigns forward, overly optimistic projections could stifle innovations and further investment, ultimately hindering the initiative.
This study adds to the growing body of evidence on the feasibility of eradicating LF. While our estimates suggest more time may be needed to reach LF elimination than what is currently projected, the treatment estimates for our scenarios represent 66–89% of that which has already been distributed under the GPELF. Thus, our analysis indicates that with continued commitment, eradicating LF is within reach.
|
10.1371/journal.pgen.1006284 | Whole Exome Sequencing in Atrial Fibrillation | Atrial fibrillation (AF) is a morbid and heritable arrhythmia. Over 35 genes have been reported to underlie AF, most of which were described in small candidate gene association studies. Replication remains lacking for most, and therefore the contribution of coding variation to AF susceptibility remains poorly understood. We examined whole exome sequencing data in a large community-based sample of 1,734 individuals with and 9,423 without AF from the Framingham Heart Study, Cardiovascular Health Study, Atherosclerosis Risk in Communities Study, and NHLBI-GO Exome Sequencing Project and meta-analyzed the results. We also examined whether genetic variation was enriched in suspected AF genes (N = 37) in AF cases versus controls. The mean age ranged from 59 to 73 years; 8,656 (78%) were of European ancestry. None of the 99,404 common variants evaluated was significantly associated after adjusting for multiple testing. Among the most significantly associated variants was a common (allele frequency = 86%) missense variant in SYNPO2L (rs3812629, p.Pro707Leu, [odds ratio 1.27, 95% confidence interval 1.13–1.43, P = 6.6x10-5]) which lies at a known AF susceptibility locus and is in linkage disequilibrium with a top marker from prior analyses at the locus. We did not observe significant associations between rare variants and AF in gene-based tests. Individuals with AF did not display any statistically significant enrichment for common or rare coding variation in previously implicated AF genes. In conclusion, we did not observe associations between coding genetic variants and AF, suggesting that large-effect coding variation is not the predominant mechanism underlying AF. A coding variant in SYNPO2L requires further evaluation to determine whether it is causally related to AF. Efforts to identify biologically meaningful coding variation underlying AF may require large sample sizes or populations enriched for large genetic effects.
| Atrial fibrillation is a common and morbid cardiac arrhythmia. Atrial fibrillation is heritable, and numerous genome-wide susceptibility loci have been identified, predominantly in non-coding regions. Over 35 genes also have been implicated in atrial fibrillation pathogenesis mostly through prior smaller scale candidate gene association studies, which generally did not have robust replication to support the associations. Therefore, the role of coding variation in the biology of atrial fibrillation is unclear. We examined whole exome sequencing data from 1,734 individuals with and 9,423 without atrial fibrillation, and did not observe any significant associations between coding variation and the arrhythmia. Furthermore, we did not observe any enrichment for association in previously implicated atrial fibrillation genes. In aggregate, our findings suggest that large effect coding variation is unlikely to be a predominant mechanism of common forms of atrial fibrillation encountered in the community.
| Atrial fibrillation (AF) is a common [1, 2] arrhythmia associated with substantial morbidity [3–7]. Current treatments for AF have limited efficacy and can cause significant adverse effects [8, 9]. AF is heritable and approximately one in four individuals with AF has a first-degree relative with the condition [10].
In recent years a large number of genes have been implicated in AF risk using both genome-wide association studies and candidate gene screening approaches. Large-scale genome-wide association studies have identified multiple AF susceptibility loci [11–15], and the top variants at discovered loci have largely been localized to noncoding regions of the genome. In contrast, there have been over 35 genes implicated in AF in candidate gene studies [16]. These studies have had a number of limitations including small sample sizes, consideration of only one or a small number of genes, and the lack of suitable control populations. To date, large-scale studies to determine whether these genes are truly related to AF have not been performed.
Since the discovery of genes causally related to AF may enable a better understanding of AF pathogenesis and potentially inform the development of therapies for AF, there is a critical need to systematically identify the genetic basis of AF. We therefore sought to assess the relations between coding variation and AF in a large sample of individuals who underwent whole exome sequencing. We further sought to determine whether coding variation in genes implicated in AF was enriched among AF cases.
The current analysis included 6,737 participants of European ancestry (n = 1,155 AF events) and 1,246 participants of African ancestry (n = 246 AF cases) from a Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) exome sequencing effort and 1,919 participants of European ancestry (n = 233 cases) and 1,255 participants of African ancestry (n = 100 AF events) from the NHLBI-GO Exome Sequencing Project (ESP). The clinical characteristics of studied participants are listed in Table 1. Sequencing coverage for the subset of AF genes is provided in S2 Table.
A total of 99,404 common variants (MAF≥0.01) were included in our study. Approximately 99.7% of the variants were already reported in dbSNP (version 142) or the 1000 Genomes Project. The Manhattan plot representing the primary pooled ancestry analysis is displayed in Fig 1 and the QQ plot is shown in S1 Fig. No inflation of Type I error was observed (genomic control λ = 0.91).
The top 15 variants most significantly associated with AF are listed in Table 2. No common variants were significantly associated with AF after Bonferroni correction for multiple testing (all P>0.05/99,404 = 5.0x10-7). The most significantly associated variant was rs56025621 (P = 1.6x10-5), which is located in the first intron of HFE2, a gene encoding the hemochromatosis type 2 peptide. The SNP was not genotyped in HapMap phase II, thus the association between rs56025621 and AF was not assessed in previous genome-wide association studies.
The SNP rs3812629, a missense variant encoding a proline to leucine amino acid substitution at amino acid 707 of SYNPO2L, occurs at a genome-wide significant disease susceptibility locus for AF [15]. The variant is in moderate linkage disequilibrium (r2 = 0.69 European ancestry, 1000 Genomes Project) with the top SNP (rs10824026) associated with AF at the locus in a prior genome-wide association study [15]. In the subset of individuals with both genome-wide genotyping data and exome sequence data available from ARIC (n = 6,630), CHS (n = 671), and FHS (n = 1,256) we examined associations between the top noncoding SNP (rs10824026) and the p.Pro707Leu (rs3812629) variant with AF after adjustment for one another (S3 Table). Adjustment for the coding variant attenuated the signal of the lead GWAS SNP in the analysis, and vice versa, suggesting the two variants represent the same AF susceptibility signal.
Previously reported top SNPs for AF derived from genome-wide association studies [15, 17], which are located in noncoding regions, were not assayed using the capture arrays in this study. As such, they were not analyzed in the current analysis.
We collapsed rare variants (MAF<1%) into gene regions and performed association testing between each gene region with AF. Our primary analysis was restricted to nonsynonymous and splice-site variants. We excluded gene regions with a cumulative MAF less than 1%. In total, we tested 8,879 gene regions. None of the gene regions were significantly associated with AF after adjusting for multiple testing (all p>0.05/8,879 = 5.6x10-6). The most significantly associated gene region was IL17REL (p = 1.3x10-5), a gene encoding interleukin 17 receptor E-like (S4 Table). The most significant single variant in IL17REL in this analysis was rs200958270 (OR 6.92, 95% CI 3.38–14.15, p = 1.2x10-7), a missense (p.Glu151Gly) variant that has a minor allele frequency of 0.004%. The variant did not meet our prespecified significance criteria for association. In a secondary analysis restricted to damaging variants, no specific gene regions were significantly associated with AF (S5 Table). Again, the most significantly associated gene in the damaging analysis was again IL17REL (p = 1.9x10-6). Variants in IL17REL have been implicated in inflammatory bowel disease [18, 19] though the relations between variation in IL17REL and cardiac function are unclear.
We also examined the associations between rare coding variants and AF within reported AF-susceptibility genes (Table 3). None were significantly associated with AF after adjusting for multiple testing.
In a post-hoc exploratory analyses, we included all rare variants (<1%) within each gene region, irrespective of annotation, and tested them for association with AF using an adjusted significance threshold of p = 2.5x10-6 (0.05/19913 genes). The results are summarized in S6 Table. The lead gene associated with AF was ACY3 (p = 2.2x10-7), which encodes aminoacylase 3. No relation between ACY3 and cardiac function or arrhythmias has been described previously.
With the current sample size, we estimated the statistical power to identify genetic variants with α = 5x10-7, assuming 100,000 independent tests. As shown in Fig 2, we had limited statistical power to identify genetic variants with allele frequencies as low as 1% unless the genetic relative risk was higher than two. In contrast, the statistical power increased significantly for relatively common variants with allele frequencies of at least 5%.
We subsequently assessed whether genetic variation in pre-specified gene sets was enriched among individuals with AF. We did not observe enrichment for common (FDR = 0.38) or rare (FDR = 0.91) variation in reported AF-related genes among individuals with AF (Table 4).
In our sample of 1,734 individuals with and 9,423 without AF who underwent whole exome sequencing, we did not observe any rare coding variation significantly associated with AF. Our observations suggest that coding variation with large effect sizes is unlikely to be the predominant mechanism underlying common forms of AF.
Our results extend prior literature focusing on coding variation underlying AF. Numerous reports propose coding variation as a mechanism underlying AF (S1 Table). However, much of the prior literature was generated via candidate gene association studies. Such discoveries have not been routinely replicated, and the studies were of small size, potentially favoring spurious results. Indeed, we previously observed that most findings from prior AF candidate gene association studies were not replicated when tested in additional study samples [17].
In the context of prior literature and our sample size, our study has two major implications for understanding AF pathogenesis. First, the lack of observed association between coding variation and AF implies that large effect coding variation is not likely to be common in typical forms of AF. In contrast, both noncoding variation, and coding variation with smaller effect sizes, may contribute to AF pathogenesis. Genome-wide association studies have identified highly associated common genetic variants near ion channels, cardiac and pulmonary transcription factors, and other genes in individuals with AF [11–15], underscoring the polygenic nature of AF. Nevertheless, the causal variants and genes underlying the arrhythmia remain unknown. Future whole-genome sequencing efforts may help to clarify the genetic contributions to AF.
Second, our findings suggest that efforts to identify potential therapeutic targets for AF through exome sequencing analyses will require much larger sample sizes or populations enriched for large genetic effects. Such populations might include those with early onset AF or consanguineous populations with the propensity to homozygous loss of function alleles in genes. Nevertheless, the additional cost required to sequence such large populations must be balanced against the potentially more cost-efficient approach of performing GWAS genotyping, imputation, and subsequent functional characterization for genetic discovery. The lack of observation of any prominent coding variation underlying AF is consistent with other whole exome sequencing efforts of complex diseases such as coronary disease and diabetes [20], which generally have not identified coding variation as the major mechanisms underlying these conditions.
Our study should be interpreted in the context of the study design. Our study was predominantly comprised of individuals of European ancestry, and therefore the findings may not be generalizable to other ancestral groups. The individuals with AF may have had multiple etiologies for the condition, and may not have been enriched for genetic forms of the arrhythmia. We cannot exclude that AF may have been misclassified, especially since AF may be paroxysmal and asymptomatic at times. Such misclassification is expected to bias the results toward the null. Furthermore, our study had limited power to assess the role of many coding variants, particularly because classifying missense variants as pathogenic or not remains challenging despite the routine use of bioinformatic algorithms. An earlier report of whole exome sequencing in 6 families with AF has summarized some of the bioinformatics challenges of utilizing whole exome sequencing data [21]. The size of our study sample limited our ability to detect potentially functional rare variants. Additionally, we utilized a Bonferroni significance threshold, which may be overly conservative for genetic discovery.
In conclusion, we observed that coding variation is not a major contributor to AF in a sample of individuals predominantly of European ancestry. Efforts to identify coding variation underlying AF will require much larger study samples. Future analyses that integrate coding and noncoding variation, such as whole genome sequencing, are warranted.
The current study included participants from three population-based cohorts that participated in the CHARGE exome sequencing effort (N = 15,459 individuals of either European or African ancestry): the Atherosclerosis Risk in Communities study (ARIC), Cardiovascular Health Study (CHS), and Framingham Heart Study (FHS). In ARIC, a random subset of 4000 European ancestry control subjects and 1000 African ancestry subjects were chosen without regard for age or sex matching. Each cohort has been described in detail previously [22–25].
We also included individuals from ESP (N = 6823 individuals of European or African ancestry) in whom AF data were ascertained (cohorts included ARIC, CHS, FHS and the Women's Health Initiative) [26]. We omitted from analysis samples for whom phenotypic data for AF were missing (N = 2593 CHARGE, N = 3689 ESP). Individuals in ESP that overlapped with individuals from the CHARGE effort (n = 40) were omitted to avoid duplicate individuals in analyses. Institutional Review Boards or Ethics Committees approved each contributing study. All participants provided written informed consent to participate in genetic research on cardiovascular disease.
We performed a combined analysis of exome sequencing conducted in the CHARGE consortium [27] and ESP [26]. In CHARGE, the exome was captured using NimbleGen SeqCap EZ VCRome (Roche, Basel, Switzerland). The enriched library was then sequenced by Illumina HiSeq platform at Human Genome Sequencing Center at Baylor College of Medicine. The Mercury pipeline [28] was used to process sequencing data, whereas the raw short reads were aligned to the reference human genome (NCBI Genome Build 37, 2009) by Burrows-Wheeler Aligner [29], and the variants were called by Atlas [30]. The mean read depth was 92x, and more than 92% of target regions were covered by at least 20 unique reads. Rigorous quality control was performed to exclude low-quality variants or samples. We excluded variants that were multi-allelic or monomorphic, had a missing rate higher than 20%, had mean depth higher than 500, or had Hardy-Weinberg equilibrium p-value less than 5x10-6 within ancestry groups. For individual samples, we calculated four quality metrics: mean depth, transition to transversion (Ti/Tv) ratio, number of singletons, and heterozygote to homozygote ratio. Samples with any metric exceeding 6 standard deviations in the respective study were omitted from analyses.
ESP included samples from 6823 individuals of European or African ancestry. The details of library construction, sequencing and alignment have been described previously [31–33]. Briefly, the exome was captured using either Agilent SureSelect Human All Exon 50Mb (Agilent, Santa Clara, CA) or NimbleGen SeqCap EZ VCRome (Roche, Basel, Switzerland). The sequencing was performed at the University of Washington and at the Broad Institute of MIT and Harvard. The mean depth was 127x. Variants with mean depth greater than 500, or with missing rate greater than 20% were excluded.
Ascertainment of AF in each cohort has been described previously [15]. Briefly, ascertainment of AF was standardized at each participating study and included the presence of either atrial fibrillation or flutter observed on a study electrocardiogram, within obtained medical encounters, or indicated by billing codes. Both incident and prevalent AF were treated together as AF cases for the purposes of this analysis. For ESP, AF information was obtained from the phenotype file (“ESP6800_Phenotype_Update_061212_final.xlsx”), from which individual level phenotypic data was provided.
Each cohort from CHARGE performed separate analyses and shared results for downstream meta-analysis. For ESP, samples from all cohorts were treated as a single sample for analyses, and adjusted for study sites and capture kits.
For common variants with minor allele frequency (MAF) at least 1%, the association of variants with AF was tested by multivariable logistic regression (ARIC, CHS, and ESP) or logistic generalized estimating equation to account for familial correlation (FHS). In common variant association analyses, we also included noncoding variants in regions flanking exons that were captured by the exome arrays. For rare variants (MAF<1%), we pooled all rare variants based on RefSeq gene regions, and jointly tested their associations with AF with the Sequence Kernel Association Test (SKAT) [34]. To circumvent the dilution of signals by variants with unknown functions, our primary analysis of rare variants focused on nonsynonymous and splice-site variants. In secondary analyses, we limited the analysis to damaging variants, defined as nonsense variants or variants predicted to be damaging by PolyPhen [35] or SIFT [36].
For both common and rare variant analyses, models adjusted for age and sex, and stratified by ancestry (European or African American). ARIC and CHS additionally adjusted for their clinical sites, FHS accounted for family structure. The association analyses were performed using the R package seqMeta (http://cran.r-project.org/web/packages/seqMeta/). Each cohort provided single variant score tests as well as genotype covariance matrices for all variants. We meta-analyzed the individual-cohort results using the inverse-variance weighted fixed effects model in seqMeta. Bonferroni correction was used to adjust for multiple testing, and the significance was defined as 0.05/N, where N is the total number of tests.
Pathway analyses were used to investigate the collective effects of multiple genetic variants on AF risk. Each common variant was assigned a score to indicate its association with AF. The score was calculated as –log10(P-value), where the P-value was derived from the common variant test described above. The genetic variant was then mapped back to RefSeq genes (August 23, 2015). A gene score was defined as the highest score of variants within 110kb upstream and 40kb downstream of the gene’s most extreme transcript boundaries, which was anticipated to include the majority of cis-regulatory gene elements [37]. For rare variants, each gene was assigned a score equivalent to –log10(P-value), in which the P-value was derived from the SKAT test described previously.
We examined the enrichment of AF-related variants in an AF gene set comprised of 37 genes previously implicated in AF (S1 Table). Genes identified on the basis of GWAS results were selected on the basis of proximity to the AF susceptibility signal, biological literature supporting a putative functional role in AF pathogenesis, or using GRAIL [38]. Gene set enrichment analysis [39] was used to estimate the enrichment, and the significant gene sets were defined as those with P-value less than 0.05/3 = 0.017.
|
10.1371/journal.pbio.1001539 | Subcellular Localization Determines the Stability and Axon Protective Capacity of Axon Survival Factor Nmnat2 | Axons require a constant supply of the labile axon survival factor Nmnat2 from their cell bodies to avoid spontaneous axon degeneration. Here we investigate the mechanism of fast axonal transport of Nmnat2 and its site of action for axon maintenance. Using dual-colour live-cell imaging of axonal transport in SCG primary culture neurons, we find that Nmnat2 is bidirectionally trafficked in axons together with markers of the trans-Golgi network and synaptic vesicles. In contrast, there is little co-migration with mitochondria, lysosomes, and active zone precursor vesicles. Residues encoded by the small, centrally located exon 6 are necessary and sufficient for stable membrane association and vesicular axonal transport of Nmnat2. Within this sequence, a double cysteine palmitoylation motif shared with GAP43 and surrounding basic residues are all required for efficient palmitoylation and stable association with axonal transport vesicles. Interestingly, however, disrupting this membrane association increases the ability of axonally localized Nmnat2 to preserve transected neurites in primary culture, while re-targeting the strongly protective cytosolic mutants back to membranes abolishes this increase. Larger deletions within the central domain including exon 6 further enhance Nmnat2 axon protective capacity to levels that exceed that of the slow Wallerian degeneration protein, WldS. The mechanism underlying the increase in axon protection appears to involve an increased half-life of the cytosolic forms, suggesting a role for palmitoylation and membrane attachment in Nmnat2 turnover. We conclude that Nmnat2 activity supports axon survival through a site of action distinct from Nmnat2 transport vesicles and that protein stability, a key determinant of axon protection, is enhanced by mutations that disrupt palmitoylation and dissociate Nmnat2 from these vesicles.
| Neurons are polarized cells that rely on bidirectional transport to deliver thousands of cargos between the cell body and the most distal ends of their axons. One cargo that is of particular importance is the NAD-synthesising enzyme Nmnat2. This surprisingly unstable protein is produced in the cell body and its constant supply into axons is required to keep them alive. If this supply is interrupted, Nmnat2 levels in the distal axon drop below a critical threshold, leading to axon degeneration. The rapid turnover of Nmnat2 contributes critically to the time course of axon degeneration. If its half-life could be extended, axons may be able to survive transient interruptions of its supply. In this study, we find that disruption of Nmnat2 localization to axonal transport vesicles increases both its half-life and its capacity to protect injured neurites. Specifically, association of Nmnat2 with transport vesicles reduces it stability by making it vulnerable to ubiquitination and proteasome-mediated degradation. These findings suggest that modulation of the subcellular localization of Nmnat2 on transport vesicles could serve as a potential avenue for therapeutic treatment of axon degeneration.
| The chimeric fusion protein WldS (Entrez Gene ID 22406) affords robust protection of injured axons in vitro and in vivo [1],[2] and extends axon survival in several disease models [3]–[9]. The WldS protein incorporates full-length Nmnat1 (Entrez Gene ID 66454), an NAD synthesising enzyme whose enzymatic activity is necessary for the protective effect of WldS [10],[11]. Additionally WldS harbours an N-terminal region that causes its partial re-distribution from the nucleus to an axoplasmic localization that is necessary for axon protection in vivo [1],[10]–[12]. We recently identified the related NAD-synthetic enzyme Nmnat2 (Entrez Gene ID 226518) as a labile, endogenous axon survival factor whose constant supply from cell bodies into axons is required for axon survival in primary culture. Specific depletion of Nmnat2 causes neurite degeneration without injury. After injury, endogenous Nmnat2 is depleted rapidly in the distal stump of neurites, initiating the process of Wallerian degeneration. If the more long-lived WldS protein is present in the axon, Nmnat2 is still depleted at the same rate; however, the NAD-synthetic enzyme activity of the stable WldS protein substitutes for that of Nmnat2, resulting in a significant delay of Wallerian degeneration [13].
Supporting a role of Nmnat2 in axon survival, strong overexpression of Nmnat2 delays Wallerian degeneration in vitro, and this protective effect is dependent on its enzymatic activity [13],[14]. Furthermore, Nmnat2 overexpression delays injury-induced axon degeneration in zebrafish [15] and alleviates neurodegeneration in the P301L mouse model of tauopathy [16]. Spontaneous synapse and axon loss in Drosophila lacking dNmnat, which can be partially rescued by murine Nmnat2, also supports a key role for endogenous axonal Nmnat activity in axon survival [17].
Neurons are extremely polarized cells with processes extending up to centimetres or even meters beyond the cell body. Moreover, in some neurons the axon constitutes over 99% of total cytoplasmic volume [18]. Despite growing evidence for the local, axonal synthesis and regulation of some proteins [19]–[21], many others appear to be synthesized only in the cell body and rely on axonal transport to reach their site of action in the axon or synapse. This supply process is a huge logistical challenge, and not surprisingly, any impairment affects axonal function or survival. Indeed, recent work has illustrated that there is a significant, early impairment in axonal transport in many neurodegenerative conditions and, for at least some of these, impaired axonal transport appears to cause the degenerative process [22].
Given that Nmnat2 is essential for axon maintenance [13] and its mRNA has not been found in axons [23]–[26], the delivery of this short-lived protein into axons is likely to limit axon survival when axonal transport is impaired through injury, aging, or disease. Although the WldS protein can overcome the need for Nmnat2, its clinical application faces the problem that this gain-of-function chimeric fusion protein is expressed only in WldS mice and a few strains of transgenic organisms. Instead, understanding and manipulating the delivery, turnover, or intra-axonal targeting of the endogenous survival factor, Nmnat2, is a more promising route to influence axon survival. Thus, it is important to understand how healthy neurons ensure a steady supply of Nmnat2 and the mechanisms that regulate its delivery, turnover, and activity in axons.
Nmnat2 localizes to vesicular structures and undergoes fast axonal transport in the neurites of primary culture neurons [13],[27]. Previous work showed that association of Nmnat2 with Golgi membranes in HeLa cells requires palmitoylation of the two adjacent cysteine residues C164/165. In the absence of these residues, Nmnat2 adopts a more diffuse, cytosolic localization [27],[28]. This palmitoylation site is located within the isoform-specific targeting and interaction domain (ISTID) of Nmnat2. ISTID regions are found in all three mammalian Nmnat isoforms and, in contrast to the more conserved core catalytic domains that make up the remainder of the proteins, are highly divergent between isoforms and are thus thought to account for the differential subcellular localizations of the Nmnats [28].
Here we report that the small, centrally located exon 6 is both necessary and sufficient for palmitoylation, stable membrane association, and vesicle-mediated delivery of Nmnat2 into axons. By manipulating its localization, we then test the hypotheses that these transport vesicles are the sites of Nmnat2 axon-protective action and that vesicular Nmnat2 is somehow protected from rapid turnover, enabling it to reach the ends of long axons before being degraded. Surprisingly, we find that a diffuse, nonvesicular localization enhances Nmnat2 axon protection through increased protein stability. Our results support a model in which Nmnat2 subcellular localization regulates its turnover and protective capacity and suggest a site of action distinct from its transport vesicles.
Previously, we reported that, in primary culture neurons, Nmnat2-EGFP is transported in particulate structures with an anterograde bias and at velocities in the range of fast axonal transport [13]. To shed light on the identity of this trafficking organelle, we utilized dual-colour, live-cell imaging of primary culture superior cervical ganglia (SCG) neurons to visualize the co-transport of Nmnat2 with established axonal transport cargos and organelle markers. Co-migration of two fluorescent markers was quantified using kymographs (see Materials and Methods), counting only moving particles. We detected significant co-migration of Nmnat2 with several Golgi network markers (Golga2, Syntaxin6, TGN38; Figure 1D–F, L). Similarly, high levels of co-migration were observed for synaptic vesicle markers (SNAP25, Synaptophysin, SynaptotagminI; Figure 1I–K, L). Interestingly, however, no significant co-migration was found for mitochondria (Figure 1A, L), lysosomes (lamp1; Figure 1B, L), or active zone precursor vesicles (Bassoon; Figure 1G, L), and only partial co-migration was detected for a marker of the ER-Golgi intermediate compartment (ERGIC) (ERGIC53/LmanI; Figure 1C, L). Thus, Nmnat2 undergoes fast axonal transport in a Golgi-derived vesicle population that overlaps with synaptic vesicle precursors and is distinct from mitochondria, lysosomes, active zone precursor vesicles, and the ERGIC.
Next, we tested whether the reported mechanism of Golgi-targeting in HeLa cells [27],[28] applies to neuronal cell bodies and vesicle targeting in axons. In order to define the requirements for membrane association in neurons, we used a photoactivation assay in SCG cell bodies. Photoactivatable GFP (PA_GFP [29]) was fused to variant Nmnat2 sequences and microinjected along with an mCherry marker to identify injected cells. PA_GFP was activated in a small region of the cell body and the pool of activated PA_GFP followed over time. For quantification, we compared the fraction of fluorescence intensity that remained in the activated area with a non activated area elsewhere in the cell body (excluding the nucleus). Activated PA_GFP alone diffused very rapidly throughout the cell body, resulting in an even distribution of fluorescence after about 5–10 s (Figure 2). In contrast, Nmnat2-PA_GFP was retained within the originally activated area, suggesting strong membrane association that was stable over the course of the experiment (Figure 2; Table 1). This indicates that the majority of Nmnat2 is stably membrane-bound and most of the spread of fluorescence that did occur was slow and resulted from transport of vesicle-bound Nmnat2-PA_GFP out of the activated region (see Figure 2A).
The C164/165 palmitoylation site is located at the centre of the 27 amino acids encoded by exon 6 of Nmnat2 (see Figure S1A for Nmnat2 primary structure with relevant regions highlighted). To test whether this exon is sufficient for membrane targeting, we created an exon6-PA_GFP construct and subjected it to the membrane association assay described above. Interestingly, we observed a very strong membrane association that was indistinguishable from that seen with full-length Nmnat2-PA_GFP (Figure 2; Table 1). This same sequence also targeted EGFP to transport vesicles in neurites that co-migrated with full-length Nmnat2-mCherry. The degree of co-migration was similar to that between Nmnat2-EGFP and Nmnat2-mCherry (Figure 3B, C, and H). Together, these results indicate that exon 6 encodes residues sufficient for efficient, stable membrane association and the resulting vesicular fast axonal transport of Nmnat2.
We then confirmed the requirement for the C164/165 palmitoylation site for membrane association in neurons, using a C164S/C165S construct (Nmnat2ΔPS-PA_GFP; see Figure S1B for an overview of all mutant constructs used in this study). Photoactivation in SCG cell bodies resulted in a rapid spread of fluorescence, similar to PA_GFP alone (Figure 2), although quantification revealed that the spread was slightly slower than for PA_GFP, even when accounting for differences in molecular size (see k values in Table 1) [30], suggesting that other residues contribute weakly to membrane association.
GAP43 (Entrez Gene ID 14432), which associates with membranes through palmitoylation of a similar double-cysteine motif, also requires an adjacent group of basic residues for efficient and stable membrane association [31]–[33] and shows partial co-migration with Nmnat2 (Figure 1H, L). We tested whether a similar mechanism applies to Nmnat2 by mutating the five basic residues encoded by exon 6 (K151A, K155A, R162A, R167A, and R172A). This construct, Nmnat2ΔBR-PA_GFP, showed intermediate membrane association. A pool of diffusible material spread quickly throughout the cell (as seen with PA_GFP), but a significant portion of the signal remained in the originally activated area for the duration of the experiment, suggesting the presence of a pool of strongly membrane-bound material (as seen with Nmnat2-PA_GFP and exon6-PA_GFP) (Figure 2). These results suggest that, in addition to the palmitoylated cysteines themselves, basic residues that surround C164/165 are also required for efficient palmitoylation and membrane association. However, the mobility of the diffusible portion of Nmnat2ΔBR-PA_GFP was not significantly different from that of Nmnat2ΔPS-PA_GFP (Figure 2; Table 1). Accordingly, a double mutant, Nmnat2ΔPSΔBR-PA_GFP, also showed the same mobility as either of the single mutants, thus still exhibiting slower diffusion than PA_GFP alone.
To confirm that these changes in membrane association reflect the degree of palmitoylation of Nmnat2, we used radiolabelling to measure palmitate incorporation into wild-type and mutant Nmnat2 (Figure S2). In agreement with previous findings [27],[28] we found that FLAG-Nmnat2ΔPS loses all detectable palmitoylation. Interestingly, however, a small but significant portion of palmitate incorporation was maintained in FLAG-Nmnat2ΔBR as predicted by the membrane association assay, further supporting the idea that exon 6 basic residues are necessary to enable efficient palmitoylation and only a small amount of palmitoylation can occur in their absence. To further substantiate the role of palmitoylation in Nmnat2 membrane association, we treated SCG neurons with 2-Bromopalmitate (2-BP), a lipid-based inhibitor of palmitoylation. As predicted, treatment with 2-BP substantially reduced membrane association of wild-type Nmnat2-PA_GFP in the photoactivation assay (Figure S6A, B).
Next, we sought to test the effect of exon 6 mutations on axonal transport of Nmnat2. As expected, mutation of the palmitoylation site in Nmnat2ΔPS-EGFP led to a diffuse, nonvesicular distribution in neurites. We detected little co-migration with Nmnat2-mCherry (Figure 3E), which was not significantly different from EGFP alone (Figure 3A, H). Like other cytosolic proteins, nonspecific or transient membrane association may help deliver this protein to neurites in this system [34]. For Nmnat2ΔBR-EGFP, the amount of diffuse, nonvesicular fluorescence signal was also greatly increased, but we still observed significant co-migration with Nmnat2-mCherry (Figure 3D, H), consistent with the residual palmitoylated, membrane-bound component inferred from the photoactivation and palmitate labelling experiments. Nmnat2ΔPSΔBR-EGFP and Nmnat2Δex6-EGFP (lacking all of exon 6) were both similar to Nmnat2ΔPS-EGFP with no further reduction in co-migration with Nmnat2-mCherry (Figure 3F–H). Taken together, these results suggest that, within exon 6, both the palmitoylation site and surrounding basic residues are necessary for efficient, stable membrane association, and vesicular axonal transport of Nmnat2.
We next tested the hypothesis that vesicle targeting is necessary for Nmnat2-mediated axon protection. We injected wild-type or variant Nmnat2-EGFP together with a DsRed2 fluorescent marker into SCG cell bodies and transected their neurites 48 h later. Due to its short half-life, Nmnat2-EGFP protects neurites for 24 h after transection only if strongly overexpressed. Surprisingly, however, both Nmnat2ΔPS-EGFP and Nmnat2ΔBR-EGFP preserved transected neurites significantly more strongly when only 0.001 µg/µl of DNA was injected (Figure 4A, C). Interestingly, the protective effects of these two mutations were additive. At 0.002 µg/µl, Nmnat2ΔPSΔBR-EGFP showed strongly preserved neurites up to 72 h, significantly more than either single mutant (Figure 4B, D). To rule out any influence of the EGFP tag on these results, we also found that untagged Nmnat2ΔPSΔBR protects neurites significantly better than untagged wild-type Nmnat2 (0.01 µg/µl; Figure S3). Thus, vesicle association is dispensable for Nmnat2-mediated neurite protection in primary culture, and missense mutations disrupting vesicle association boost Nmnat2 axon protective capacity to a modest but significant degree.
These surprising findings prompted us to investigate the effects on axon protection of deletions within and around exon 6. In particular, three deletion mutants were recently reported to retain enzyme activity (Nmnat2Δ32-EGFP, Nmnat2Δ43-EGFP, Nmnat2Δ69-EGFP [35]). Remarkably, all these deletion mutants and a mutant lacking exon 6 only (Nmnat2Δex6-EGFP) protected neurites far more strongly than any of the missense constructs above. Even microinjection of very low DNA concentrations (0.0005 µg/µl) preserved around 80% of neurites for 72 h after transection (Figure 5), compared with less than 10% for Nmnat2-EGFP or Nmnat2ΔPSΔBR-EGFP at this concentration. Surprisingly, this even exceeds the level of protection achieved by WldS-EGFP at this concentration (around 30% intact neurites at 72 h), illustrating the very strong enhancement of axon protective capacity in these mutants (Figure 5). The presence of the EGFP tag was not necessary for the observed increase in protection (Figure S4). To rule out the possibility that deletion of exon 6 induces a novel gain-of-function in Nmnat2 that is independent of its NAD-synthesis activity, we introduced an enzyme-dead mutation into Nmnat2Δex6. His24 is conserved in all three mammalian Nmnats and is critical for Nmnat2 NAD-synthesis activity as well as its ability to protect axons after cut [14],[36],[37]. Convincingly, a Nmnat2Δex6H24D enzyme-dead mutant did not protect neurites after injury (Figure S5), although we found that this mutation also reduces the stability of Nmnat2Δex6H24D (unpublished data). Together, these findings suggest that, in addition to the palmitoylation site and surrounding basic residues, other exon 6 residues influence the axon protective capacity of Nmnat2, while the additional sequences deleted outside of exon 6 in Nmnat2Δ32, Nmnat2Δ43, and Nmnat2Δ69 appear to have little further effect.
Next, we sought to identify the mechanism by which these mutations increase axon protective capacity. As the short half-life of Nmnat2 limits survival of injured axons [13], we decided to investigate protein stability using an emetine chase assay. HEK293 cells transiently expressing FLAG-Nmnat2 or one of its mutant forms were treated with 10 µM emetine to inhibit protein synthesis. Nmnat2 turnover was then measured as the rate of decline of the FLAG-Nmnat2 signal over time relative to a more stable control (FLAG-WldS) [13]. We had envisaged that vesicular Nmnat2 may be relatively stable, allowing it to reach the ends of long axons, and in dynamic equilibrium with a less stable, but perhaps more active cytosolic form. However, all mutations disrupting membrane targeting were found to increase protein stability (Figure 6A, B; Table 2). Moreover, the FLAG-Nmnat2ΔPSΔBR double mutant showed significantly higher protein stability than either FLAG-Nmnat2ΔPS or FLAG-Nmnat2ΔBR, supporting a model in which an increase in protein stability contributes to the increase in protective capacity observed for these missense mutations. Intriguingly, however, the very strongly protective Nmnat2Δex6 construct was no more stable than Nmnat2ΔPSΔBR (Figure 6A, B; Table 2), suggesting that factors other than protein stability underlie the further increase in its protective capacity.
As Nmnat2 degradation is blocked by proteasome inhibitor MG132 [13], we then asked whether Nmnat2 becomes ubiquitinated and whether these stabilizing mutations reduce ubiquitination. Wild-type and mutant FLAG-Nmnat2 were overexpressed in HEK293 cells, and the K48-specific ubiquitin binding domain of Dsk2 was bound to Glutathione-Sepharose beads and used to immunoprecipitate ubiquitinated proteins. An inactive mutant form of the Dsk2 UBA was used as a control [38],[39], and 20 µM MG132 was added 6 h prior to cell lysis to increase the abundance of ubiquitinated proteins. FLAG-Nmnat2 produced a clear ladder of ubiquitinated products (Figure 6C), and all missense mutants and Nmnat2Δex6 showed significantly less ubiquitination (Figure 6C, D). Thus, reduced ubiquitination is likely to contribute to the increased stability and protective capacity of these mutants.
Based on these results, we hypothesized that palmitoylation and vesicle association cause wild-type Nmnat2 to become destabilised through increased levels of ubiquitination. In contrast, the nonvesicular, cytosolic location of the Nmnat2 mutants reduces ubiquitination and increases protein stability and protective capacity. The Nmnat2ΔPS data, and strongly enhanced protective capacity of Nmnat2Δex6 over Nmnat2ΔPSΔBR, indicate that the observed effects do not just reflect removal of lysine residues. In agreement with this model, inhibiting palmitoylation directly with 2-BP resulted in reduced levels of ubiquitination on FLAG-Nmnat2 (Figure S6C, D). Furthermore, 2-BP treatment enhanced the half-life of FLAG-Nmnat2 in the emetine chase assay (Figure S6E, F). Based on this, we predicted that a cytosolic, stable Nmnat2 mutant with increased protective capacity (such as Nmnat2ΔPSΔBR) would revert to a wild-type behaviour when re-targeted to vesicles. To test this, we first attached sequences to the N-terminus of Nmnat2ΔPSΔBR-PA_GFP that would re-target it to the same Golgi-derived vesicle population as wild-type Nmnat2. For this, we used exon 6 of Nmnat2 (Nmnat2ΔPSΔBR-Nterex6-PA_GFP) or the signal peptide and transmembrane domain of TGN38 (Nmnat2ΔPSΔBR-NterTGN-PA_GFP). These constructs were stably targeted to membranes as assessed by the photoactivation assay (Figure S7A, B), and as predicted, re-targeting significantly reduced the ability to protect injured neurites (Figure 7A, B). We also confirmed that Nmnat2ΔPSΔBR-NterTGN showed substantial levels of ubiquitination that were not detectable in Nmnat2ΔPSΔBR (Figure 7E) and that the protein stability of Nmnat2ΔPSΔBR-NterTGN was significantly reduced as expected (Figure 7C, D).
While these results suggest that re-targeting cytosolic Nmnat2 to membranes reverts its stability and protective capacity to lower levels as expected, we cannot rule out the possibility that the N-terminal sequences have a direct effect on Nmnat2 stability. To address this issue, we used a commercially available heterodimerisation system (iDimerize, Clontech), in which two proteins tagged with DmrC and DmrA domains, respectively, undergo heterodimerisation after addition of a soluble “A/C heterodimeriser” compound [40]. We used TGN38-DmrC-HA to provide the membrane anchor for re-targeting of cytosolic DmrA-Nmnat2ΔPSΔBR-PA_GFP based on the strong co-migration of TGN38 with Nmnat2 (see Figure 1). Thus, this system overcomes the abovementioned limitation of the N-terminal targeting sequence as all that is required to induce membrane re-targeting is addition of the small molecule heterodimeriser compound. The photoactivation assay confirmed successful re-targeting, as mobility of DmrA-Nmnat2ΔPSΔBR-PA_GFP was significantly reduced after addition of heterodimeriser, albeit less strongly than using the N-terminal targeting sequence above (Figure S8). Correspondingly, addition of heterodimeriser also significantly reduced the axon protective capacity of DmrA-Nmnat2ΔPSΔBR-PA_GFP, but only in the presence of TGN38-DmrC-HA (Figure 8A, B). Even though the reduction in protective capacity observed in response to N-terminal re-targeting was stronger than that achieved by heterodimerisation, this was reflected in a lower degree of membrane re-targeting in the heterodimerisation system (compare Figure S7A, B and Figure S8). Additionally, we confirmed that re-targeting of cytosolic Nmnat2ΔPSΔBR to membranes through heterodimerisation resulted in increased levels of ubiquitination (Figure 8F, G) and a decreased protein half-life (Figure 8C–E), confirming the results obtained with N-terminally re-targeted mutants above.
At this point, it is interesting to ask whether the observed changes after Nmnat2 re-targeting arise from a special property of the vesicle membranes that Nmnat2 exon 6 and TGN38 target to, or whether they reflect a more general effect of Nmnat2 membrane association. To test this, we attached the mitochondrial outer membrane anchor (a.a. 1–37) of TOM20 to the N-terminus of Nmnat2ΔPSΔBR (Nmnat2ΔPSΔBR-NterMOM). Note that, as with Nmnat2ΔPSΔBR-NterTGN, the Nmnat2 portion of this construct faces the cytosol. The NterMOM tag led to stable membrane association in the photoactivation assay (Figure S7A, B). To confirm targeting to mitochondria, we co-stained neurons expressing Nmnat2ΔPSΔBR-NterMOM-PA_GFP with MitoTracker dye and observed largely overlapping staining patterns (Figure S7C). We then subjected Nmnat2ΔPSΔBR-NterMOM to the ubiquitination assay and found no evidence for increased levels of ubiquitination relative to Nmnat2ΔPSΔBR (Figure 7E). This suggests that the induction of ubiquitination after membrane attachment is indeed specific for transport vesicle membranes. Despite the absence of detectable ubiquitination, however, we found Nmnat2ΔPSΔBR-NterMOM to be destabilized relative to Nmnat2ΔPSΔBR (Figure 7C, D), which resulted in a loss of protective capacity (Figure 7A, B). These findings suggest that targeting to the mitochondrial outer membrane destabilizes Nmnat2 through a mechanism distinct from that operating on transport vesicle membranes.
As described above, deletion of exon 6 led to a very strong increase in Nmnat2 protective capacity without any further changes in stability with respect to Nmnat2ΔPSΔBR. To further explore this dissociation between protective capacity and protein stability, we re-targeted Nmnat2Δex6 to vesicle membranes using the N-terminal TGN38 tag. Nmnat2Δex6-NterTGN was stably targeted to membranes in the photoactivation assay (Figure S9) and showed increased ubiquitination (Figure 9C) and reduced protein stability (Figure 9D, E) with respect to Nmnat2Δex6. Interestingly, however, this did not affect its protective capacity, which remained indistinguishable from Nmnat2Δex6 up to 72 h after cut (Figure 9A, B). This finding suggests that the increase in protective capacity resulting from loss of exon 6 is sufficient to strongly delay degeneration even when only a low, residual level of protein remains.
Nmnat2 is required for axon survival and is the only confirmed endogenous Nmnat isoform in axons. However, its ability to promote axon survival is limited by its short half-life. We have identified a series of mutations that extend Nmnat2 half-life without disrupting enzyme activity and which significantly increase axon protection. For deletion mutants lacking exon 6 the efficacy even surpasses that of WldS. Surprisingly, these changes arise when Nmnat2 targeting to a population of post-Golgi axonal transport vesicles is disrupted and are reversed when vesicle targeting is restored, indicating a nonvesicular site for the axon survival function of Nmnat2. The more stable and protective variants are less prone to ubiquitination through a mechanism likely to involve subcellular targeting and not just lysine availability. We identify cysteine-linked palmitoylation as the vesicle targeting mechanism and propose modulation of this targeting as a promising, novel therapeutic strategy for axonopathies.
We show that Nmnat2 axonal transport vesicles carry Golgi markers as well as synaptic vesicle markers. In contrast, we find no evidence of Nmnat2 undergoing co-transport with mitochondria. This suggests that Nmnat2 is involved in the regulation of cytosolic and not mitochondrial NAD metabolism in the axon, especially since mitochondria are not thought to take up cytosolic NAD under normal conditions [41]. Thus, any potential influence of Nmnat2 on mitochondrial function is likely to be mediated through indirect mechanisms. Furthermore, given that removal of Nmnat2 from its vesicles does not impair its protective capacity, it seems unlikely that Nmnat2 exists in close proximity to important downstream targets on its own transport vesicles. If this were the case, removal of Nmnat2 from this microenvironment would be expected to result in reduced levels of neurite protection. Instead, it appears that the regulation of overall cytosolic NAD metabolism by Nmnat2 is critical for axon survival.
Given the importance of axonal transport, surprisingly little is known about sequences targeting proteins to axons. The mechanism of Nmnat2 membrane association appears very similar to that reported for another axonally transported protein, GAP43. Both proteins lack a transmembrane domain or alternative membrane targeting structures and depend fully on palmitoylation of a double-cysteine motif for membrane association [27],[28],[42],[43]. For GAP43, it was reported that, in addition to the palmitoylated cysteine residues themselves, several adjacent basic amino acid residues are also necessary for efficient membrane association [31]–[33]. We found a similar mechanism to operate for Nmnat2. Our results suggest that basic amino acid residues surrounding the palmitoylation site in exon 6 are involved in mediating initial membrane contact and allow palmitoylation to establish stable membrane anchoring. In their absence, the level of palmitoylation is strongly reduced and membrane targeting of Nmnat2 is less efficient, resulting in a higher level of diffuse, soluble protein. However, once palmitoylation has occurred, membrane association seems to be as stable as for wild-type Nmnat2 over the time-scale of our analysis. This view is also supported by our finding that Nmnat2ΔBR, despite its increased level of diffuse fluorescence in neurites, is still associated with the correct population of transport vesicles. These similarities between GAP43 and Nmnat2 suggest a common underlying axonal targeting motif based on a dual-cysteine palmitoylation site and adjacent or surrounding lysine and arginine residues.
Palmitoylation regulates the axonal transport and subcellular sorting of several axonally delivered proteins [44] and is unique among the fatty-acid modifications in that it is readily reversible. It is now well established that palmitate cycling, the recurrent addition and removal of palmitate groups to a target protein, is an important regulatory mechanism in many cases [45]. Thus it is possible that endogenous Nmnat2 undergoes similar cycles of palmitoylation and depalmitoylation. If this is the case, a model could be envisaged in which Nmnat2 is palmitoylated and vesicle bound for the purpose of delivery over long distances into axons. Once in the axon, depalmitoylation could cause Nmnat2 to detach from vesicles and instead assume a diffuse, cytosolic localization in order to carry out its function in regulating axoplasmic NAD metabolism. Thus, palmitate cycling would effectively regulate the ratio of vesicle-bound to diffuse Nmnat2, and hence ultimately the axon protective capacity of Nmnat2. Modulation of the Nmnat2 palmitoylation-depalmitoylation cycle by targeting the relevant palmitoyltransferase and thioesterase enzyme(s) might hence present a useful tool to alter the course of axon degeneration.
We also found support for the hypothesis that increased protein stability underlies the mechanism by which diffuse, cytosolic Nmnat2 becomes more highly protective. Wild-type Nmnat2, which is mainly vesicle-bound, is very short-lived both in cell lines and in the neurites of primary culture neurons [13]. Here we found it to have a protein half-life of around 40 min, whereas soluble, cytosolic Nmnat2 mutants have a longer half-life. This means that while wild-type Nmnat2 is very rapidly depleted when its supply stops after neurite transection, cytosolic mutants with increased protein stability have an increased potency to protect neurites against degeneration, due to a combination of higher steady-state levels in the neurites before transection and a longer protein half-life after transection. Moreover, the reduced level of ubiquitination in these cytosolic mutants suggests that increased protein stability is a direct result of reduced turnover of Nmnat2 by the ubiquitin-proteasome system. This fits with our previous findings that inhibition of the ubiquitin proteasome system, which was shown to delay Wallerian degeneration [46], stabilizes endogenous Nmnat2 [13]. Thus our data support a model in which vesicle-bound Nmnat2 is unstable due to its high levels of ubiquitination, which in turn results in its rapid turnover and short protein half-life. Releasing Nmnat2 from its vesicles reduces ubiquitination and leads to a more stable protein with a higher axon protective capacity. Our results indicate that Nmnat2 can undergo ubiquitination on lysine residues outside of exon 6 as Nmnat2 mutants lacking exon 6 lysines (Nmnat2ΔPSΔBR and Nmnat2Δex6) can still be ubiquitinated when re-targeted to membranes.
Interestingly, this destabilising effect of palmitoylation-mediated membrane attachment contrasts with findings for several palmitoylated transmembrane domain proteins, including cell-surface receptors [47]–[50] and SNARE proteins [51], for which it was found that palmitoylation increases protein half-life through reduction of ubiquitin-proteasome mediated degradation. Our results indicate that the effect of palmitoylation on ubiquitination and protein stability might differ for proteins lacking transmembrane regions (such as Nmnat2). The mechanism by which palmitoylation-mediated vesicle-association causes high levels of ubiquitination in Nmnat2 is as yet unclear and will be the object of future studies. One interesting possibility is the localization of a relevant ubiquitin ligase to the surface of Nmnat2 transport vesicles. Such a vesicle-specific mechanism is supported by our finding that re-targeting cytosolic Nmnat2 mutants to mitochondrial outer membranes does not induce detectable ubiquitination. Alternatively, vesicular axonal transport may deliver Nmnat2 to parts of the axon with higher levels or activity of relevant elements of the ubiquitin proteasome system. A recently published study reported one such element regulating the turnover of Drosophila Nmnat (dNmnat) in axons. The Drosophila E3 ubiquitin ligase Highwire was found to be necessary and sufficient for rapid turnover of dNmnat, and of ectopically expressed mammalian Nmnat2, in the distal stump of injured axons. In its absence, dNmnat persists and degeneration is delayed [52]. Together with our findings, this suggests that ubiquitination regulates the course of axon degeneration both in mammals and Drosophila and that Nmnat2 axonal transport vesicles play an important role in bringing together dNmnat or Nmnat2 with their respective ubiquitin ligases.
Furthermore, our results indicate that subcellular localization and protein stability are not the only determinants of Nmnat2 axon protective capacity. Deletion of exon 6 dramatically increases Nmnat2-mediated neurite protection without any further increase in protein stability. Our findings with the enzyme-dead exon 6 deletion mutant suggest that this strong increase in protective capacity depends on Nmnat2 enzymatic activity. However, the reduced stability of this mutant means we cannot completely rule out the possibility that other, nonenzymatic mechanisms contribute to the rise in protective capacity. Interestingly, deletion of exon 6 overcomes the reduction in Nmnat2 protective capacity upon membrane re-targeting that was observed for point mutants. This rescue occurred despite the destabilising effects of membrane attachment, which were unchanged by the removal of exon 6. This suggests that exon 6 regulates Nmnat2 axon protective function through various mechanisms, which could include protein-protein interactions or additional posttranslational modifications.
In summary, we show that Nmnat2, normally the least axon protective of the three endogenous Nmnat isoforms due to its short half-life, can be converted to a highly protective molecule by disrupting its targeting to axonal transport vesicles. While the importance of these vesicles for long-range axonal trafficking is clear, we suggest that Nmnat2 must dissociate to carry out its axon survival function optimally. We also propose that cytosolic NAD metabolism is central to the axon survival mechanism. Our data establish the principle that Nmnat2 can be modified to promote axon survival and highlight modulation of its palmitoylation state as a route to achieve this. Unlike WldS or other Nmnats, this approach utilizes a protein already identified in wild-type axons, raising the attractive prospect of converting an endogenous axonal protein into one with a protective capacity that matches or even exceeds that of WldS.
Nmnat2-EGFP, FLAG-Nmnat2, and FLAG-WldS constructs were described previously [13]. WldS-EGFP was created by insertion of the WldS coding sequence into the MCS of pEGFP-NI vector (Clontech). Nmnat2-mCherry was created by replacing the EGFP coding sequence of Nmnat2-EGFP with the mCherry coding sequence from pmCherry-NI (Clontech). Nmnat2ΔPS-EGFP, Nmnat2ΔBR-EGFP, and Nmnat2ΔPSΔBR-EGFP were created from Nmnat2-EGFP using the QuikChange II Site Directed Mutagenesis Kit (Stratagene) according to the manufacturer's instructions. Nmnat2Δex6-EGFP was created by PCR amplification of the Nmnat2-EGFP vector excluding exon 6 and introduction of a SacII restriction site to allow vector re-ligation. FLAG-tagged, untagged, and PA_GFP tagged Nmnat2 wild-type or mutants were created by insertion of the appropriate Nmnat2 mutant into the MCS of pCMV-Tag2A (Stratagene), pCMV-Tag4A (Stratagene), and pPAGFP-NI [29] (Addgene plasmid 11909) vectors, respectively. Exon6-EGFP and Exon6-PA_GFP constructs were created by PCR of exon6 from Nmnat2 and insertion into the MCS of pEGFP-NI and pPAGFP-NI vectors. DmrA-Nmnat2ΔPSΔBR-PA_GFP was created by PCR amplification of the DmrA coding sequence from pHet-NucI vector (Clontech) and insertion at the N-terminus of Nmnat2ΔPSΔBR-PA_GFP. For the TGN38-DmrC-HA construct, the DmrC-HA coding sequence was PCR amplified from the pHET-1 vector (Clontech) and inserted in place of the GFP coding sequence of TGN38-EGFP. Nmnat2ΔPSΔBR-Nterex6-PA_GFP was created by PCR amplification of Nmnat2 exon 6 and insertion at the N-terminus of Nmnat2ΔPSΔBR-PA_GFP. The N-terminal TGN38 targeting sequence was made by fusing the TGN38 signal peptide sequence (a.a. 1–20) to its transmembrane domain surrounded by linker sequences (a.a. 281–330). This sequence was then inserted at the N-termini of Nmnat2ΔPSΔBR and Nmnat2Δex6 to create Nmnat2ΔPSΔBR-NterTGN and Nmnat2Δex6-NterTGN, respectively. Nmnat2ΔPSΔBR-NterMOM was created by addition of amino acids 1–37 of Mus musculus TOM20 to the N-terminus of Nmnat2ΔPSΔBR. Nmnat2Δex6H24D was created by site-directed mutagenesis of Nmnat2Δex6.
For organelle markers, the following accession numbers were used for PCR primer design. Constructs were amplified from mouse brain cDNA and inserted into the MCS of pEGFP-NI (Clontech) or ptagRFP-NI (Evrogen) vectors. TGN38-EGFP, NM_009443; Syntaxin6-EGFP, NM_021433; Synaptophysin-EGFP, NM_009305; mito-tagRFP, AK003116 (bp 1–72); LmanI-EGFP, AK011495; Golga2-EGFP, NM_133852; GAP43-EGFP, BC080758; SynaptotagminI-EGFP, NM_001252341. All constructs were verified by DNA sequencing (Beckman Coulter Genomics).
Nmnat2 deletion mutants (Δ32, Δ43, and Δ69) were a gift from Prof. Giulio Magni (Ancona, Italy). GST-Dsk2 UBA was kindly provided by Dr. Simon Cook (Cambridge, UK). The SNAP25-EGFP construct was a gift from Dr. Luke Chamberlain (Glasgow, UK). GFP-Bassoon was a gift from Prof. Eckart Gundelfinger (Magdeburg, Germany). Lamp1-RFP [53] was from Addgene (plasmid 1817).
All animal work was carried out in accordance with the Animals (Scientific Procedures) Act, 1986, under Project License 80/2254. C57BL/6JOlaHsd mice were obtained from Harlan UK (Bicester, UK).
Dissociated superior cervical ganglia cultures were prepared and maintained in culture as described previously [13]. For live-cell imaging of axonal transport, cells were transferred (on the day of imaging) into imaging medium in order to improve performance and detectability of fluorescent proteins [54]. Cells were viable and appeared morphologically normal in imaging medium for at least 3 days. Imaging medium consisted of 1.80 mM CaCl2 (Sigma), 0.25 µM Fe(NO3)3 (Sigma), 0.81 mM MgSO4 (AnalaR), 5.33 mM KCl (AnalaR), 44.05 mM NaHCO3 (Sigma), 110.34 mM NaCl (AnalaR), 0.92 mM NaH2PO4 (Sigma), 4,500 mg/l glucose (AnalaR), 110 mg/l sodium pyruvate(PAA), 2 mM glutamine (Invitrogen), 100 ng/ml 7S NGF (Invitrogen), 1% penicillin/streptomycin (Invitrogen), 4 µM aphidicolin (Calbiochem), 1× MEM amino acids (PAA), 30 mg/l glycine (AnalaR), and 42 mg/l serine (Sigma) in sterile distilled water. For live-cell imaging of photoactivatable GFP (PA_GFP), cells were transferred (on the day of imaging) into Hibernate-E medium (Invitrogen) with added 2 mM glutamine (Invitrogen), 100 ng/ml 7S NGF (Invitrogen), 1% penicillin/streptomycin (Invitrogen), 2% B27 supplement (Invitrogen), and 4 µM aphidicolin (Calbiochem). For heterodimerisation, 500 nM A/C heterodimeriser (Clontech) was added 8 h before imaging. Where indicated, 40 µM 2-BP (Sigma) was added immediately after microinjection.
DNA microinjections into the nuclei of primary culture SCG neurons were performed as described [13]. For dual-labelling live cell imaging, both DNA constructs were used at a concentration of 0.03 µg/µl in the injection mix. For single-color axonal transport imaging, constructs were used at 0.05 µg/µl in the injection mix. For photoactivation experiments, 0.05 µg/µl of the photoactivation construct was co-injected with 0.01 µg/µl mCherry expression construct except for experiments involving heterodimerisation where 0.03 µg/µl each of DmrA- and DmrC-tagged constructs were co-injected with 0.01 µg/µl mCherry expression construct. Seventy-five cells were injected per dish, and imaging was performed 24 h after microinjection. For experiments on neurite degeneration after cut, 0.01 µg/µl DsRed2 expression vector was co-injected with the relevant Nmnat2 constructs at varying concentrations (see text). We injected 75–100 cells in each dish and neurites were cut 48 h after microinjection. MitoTracker Red CMXRos (Invitrogen) was used according to the manufacturer's instructions.
Time-lapse imaging of axonal transport was performed on an Olympus CellR imaging system (IX81 microscope, Hamamatsu ORCA ER camera, 100×1.45 NA apochromat objective, 485 and 561 nm laser excitation). During imaging, cell cultures were maintained at 37°C in an environment chamber (Solent Scientific). Images were captured at 4 (single-color) or 2.5 (dual-colour) frames per second for 1–2 min.
The extent of axonal co-migration of two fluorescent protein markers was analysed in time-lapse recordings of individual neurites. Using ImageJ software version 1.44 (Rasband, W.S., ImageJ, NIH, Bethesda, Maryland, USA, http://imagej.nih.gov/ij/, 1997–2011), the same neurite was straightened, contrast adjusted, and projected as a kymograph for both image stacks. The kymographs were then merged to create an overlay. Co-migration was scored for each particle trace according to whether it was detectable in only one or both of the kymograph channels. Co-migration was determined only for moving particles (i.e., traces that were not exclusively vertical in kymographs).
Parameters of axonal transport of fluorescently labelled proteins were determined from straightened time-lapse recordings of individual neurites using the Difference Tracker ImageJ software plugin [55]. Additionally, the percentage of moving and stationary particles was scored manually on kymographs.
Photoactivation imaging was carried out on an Olympus FV1000 point scanning confocal microscope system (IX81 microscope, 60×1.35 NA plan super apochromat objective, 488 and 561 nm laser excitation). Microinjected cell bodies were identified based on their mCherry fluorescence. Imaging settings were adjusted to standard settings (5× zoom, scan rate 8 µm/s, frame rate 3 s/frame). After taking a pre-activation image, PA_GFP was activated by a 100 ms pulse of a 405 nm laser at 50% intensity in a 100 pixel region of interest in the cell body. Images were then taken every 3 s for a total of 5 min. For analysis, two circular regions of interest of identical size (50 pixel diameter) were selected in the cell body. One was placed in the originally activated area, while the other one was placed 10–20 µm away in an area that was not activated by the original laser pulse. For quantification, the percentage of combined fluorescence in these two areas that remained in the originally activated area was determined for each time point. Data were fitted for exponential decay, and decay constant (k) was calculated using GraphPad Prism 5.04.
Degeneration of ds-Red2 labelled neurites was determined for the same field of view at indicated time points after neurite transection. The percentage of neurites remaining continuous and morphologically normal compared to the initial time point was scored for each field. For experiments involving heterodimerisation, A/C heterodimeriser (Clontech) was added to the relevant dishes 8 h before cut at a final concentration of 500 nM. Fresh medium (with heterodimeriser where appropriate) was added 24 and 48 h after cut.
HEK 293 cells were maintained in culture as described [13]. For transfection, cells were grown to 80% confluency in 10 cm or 24-well dishes and transfected using lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. For emetine chase experiments, HEK293 cells in 24-well dishes were co-transfected with WldS-FLAG and the appropriate FLAG-Nmnat2 construct (turnover of Nmnat2 mutants) or with Flag-Nmnat1, TGN38-DmrC-HA, and DmrA-Nmnat2ΔPSΔBR-FLAG (turnover of re-targeted mutants). For relevant experiments, 100 µM 2-BP was added 6 h post-transfection. Twenty-four hours after transfection, cells were treated with 10 µM emetine (Sigma) and samples were taken at indicated time points.
For ubiquitination experiments, HEK293 cells in 10 cm dishes were transfected with the appropriate FLAG-Nmnat2 construct and, for re-targeting experiments, with empty pCMV-Tag4A or TGN38-DmrC-HA constructs. Twenty-four hours after transfection, 20 µM MG132 (Sigma) was added to the medium. Six hours later, cells were lysed and subjected to GST-Dsk2 UBA (wild-type or mutant) pulldown assay as described [38],[39]. Where indicated, 100 µM 2-BP was added 6 h post transfection.
SDS-PAGE analysis and Western blotting analysis were performed as described [13],[56]. Mouse monoclonal anti-FLAG (Sigma, M2) was used at 1∶3,000. AlexaFluor680-conjugated anti-mouse secondary antibody (Molecular Probes, Eugene, OR, USA) was used at 1∶5,000. Blots were scanned and quantified using the Odyssey imaging system (LI-COR Biosciences, Lincoln, NC, USA).
HEK293 cells in six-well dishes were transfected with the appropriate FLAG-Nmnat2 construct. Twenty-four hours after transfection, 0.5 mCi/ml [9,103H]-palmitate (Perkin Elmer) was added to the medium. After 6 h, cells were washed in PBS, lysed in 500 µl lysis buffer (20 mM Tris pH 7.5, 137 mM NaCl, 1 mM EGTA, 1% TritonX-100, 10% glycerol, 1.5 mM MgCl2, 50 mM NaF, 1 mM Na3VO4, and protease inhibitor mix (Roche); all chemicals AnalaR unless stated otherwise). The lysate was centrifuged for 10 min, 13,000 rpm. Following overnight incubation of the lysate with 5 µg of anti-FLAG antibody (Sigma), 50 µl of washed Sepharose beads (GE Healthcare) was added and mixed for another 3 h at 4°C. Beads were washed thrice in lysis buffer and twice in wash buffer (50 mM Tris, pH 8.0). Bound protein was eluted with Laemmli sample buffer (BioRad) and boiling for 5 min and processed for SDS-PAGE. After transfer to PVDF membrane, blots were dried and radiolabel was detected by exposure on Tritium phosphor screen (Fuji) for 14 d.
Statistical analyses and graph fitting were performed using GraphPad Prism 5.04 (GraphPad Software Inc.) and SPSS Statistics 19 (IBM).
|
10.1371/journal.pntd.0005941 | Multi-site cholera surveillance within the African Cholera Surveillance Network shows endemicity in Mozambique, 2011–2015 | Mozambique suffers recurrent annual cholera outbreaks especially during the rainy season between October to March. The African Cholera Surveillance Network (Africhol) was implemented in Mozambique in 2011 to generate accurate detailed surveillance data to support appropriate interventions for cholera control and prevention in the country.
Africhol was implemented in enhanced surveillance zones located in the provinces of Sofala (Beira), Zambézia (District Mocuba), and Cabo Delgado (Pemba City). Data were also analyzed from the three outbreak areas that experienced the greatest number of cases during the time period under observation (in the districts of Cuamba, Montepuez, and Nampula). Rectal swabs were collected from suspected cases for identification of Vibrio cholerae, as well as clinical, behavioral, and socio-demographic variables. We analyzed factors associated with confirmed, hospitalized, and fatal cholera using multivariate logistic regression models.
A total of 1,863 suspected cases and 23 deaths (case fatality ratio (CFR), 1.2%) were reported from October 2011 to December 2015. Among these suspected cases, 52.2% were tested of which 23.5% were positive for Vibrio cholerae O1 Ogawa. Risk factors independently associated with the occurrence of confirmed cholera were living in Nampula city district, the year 2014, human immunodeficiency virus infection, and the primary water source for drinking.
Cholera was endemic in Mozambique during the study period with a high CFR and identifiable risk factors. The study reinforces the importance of continued cholera surveillance, including a strong laboratory component. The results enhanced our understanding of the need to target priority areas and at-risk populations for interventions including oral cholera vaccine (OCV) use, and assess the impact of prevention and control strategies. Our data were instrumental in informing integrated prevention and control efforts during major cholera outbreaks in recent years.
| Cholera is a major public health problem in many countries in sub-Saharan Africa. In Mozambique, annual outbreaks occur but the place and time may vary. Africhol was implemented in Mozambique in 2011 to generate more detailed information on disease burden and characteristics to support appropriate interventions for cholera control and prevention in the country. The study was conducted in six different zones, where patients with cholera symptoms seeking care at a health facility were asked questions on socio-demographic characteristics, their symptoms, and behaviors that may increase cholera risk. Stool samples were also taken to test for the presence of cholera infection (Vibrio cholerae). Among the 1,863 patients, more than half were tested for cholera, and among those tested, less than one in four was infected with the pathogen. About 1% of patients died from cholera. Our study helps to understand the burden of cholera in different areas of the country, and the characteristics of the people infected. It is important to continue the surveillance of this disease to choose the most appropriate control and preventive interventions, and to apply them in precisely the right place.
| Cholera is an acute, diarrheal illness caused by infection of the intestine with the bacterium Vibrio cholerae. Its epidemics have been continuously reported from Southern Africa since its reintroduction on the continent in 1970 [1, 2]. Today, cholera remains a significant cause of morbidity and mortality, and a key indicator of lack of adequate infrastructure and structural development, specifically an insufficient supply of drinking water and inadequate sanitation [3, 4]. While there has been significant work done in Africa to quantify the magnitude of cholera as a public health problem in recent years, individual and population characteristics in specific settings remain ill defined.
Since 1973, Mozambique has reported cholera almost yearly. Cholera in Mozambique is an endemic disease with seasonal epidemic peaks. The usual cholera season is between January and March with annual incidences ranging from 0 to 211 per 100,000 population with periodically high case-fatality ratios (CFRs) [5–9].
Disease burden data are critical for making evidence-based decisions on public health interventions, including water, sanitation, and hygiene (WASH) and use of oral cholera vaccine (OCV). In Mozambique, epidemiological surveillance data are used to identify districts at risk of cholera and to identify contributing factors such as an insufficient supply of drinking water, sanitation practices harmful to health (such as open defecation, using dirty or contaminated water, and handling and selling unhygienic food), and natural disaster [5].
Africhol is a multi-centric project that was launched in 2009 with funding from the Bill & Melinda Gates Foundation. It consists of a consortium of 11 African countries and non-governmental organizations, seeking to collect epidemiological and microbiological information regarding the occurrence of cholera in Africa to better inform intervention strategies (http://amp-vaccinology.org/activity/cholera-in-africa)[10].
In Mozambique, the Africhol project was implemented in 2011, to conduct prospective surveillance in dedicated surveillance zones with the main objective of assessing cholera burden. The current study presents the results from five years of surveillance in terms of disease burden, population characteristics, and risks factors in specific surveillance zones to support appropriate interventions for cholera control and prevention.
The surveillance protocol was approved by the Mozambican National Bioethics Committee for Health. The informed oral consent was obtained from all participants and documented by a specific question on the case report form. The Ministry of Health (MOH) considered that the Africhol project was integrated in their national epidemic disease surveillance and response and subject to the laws and regulations, and thus did not require written consent. For minors below 18 years, informed oral consent was obtained from parents, guardians, or next of kin, on behalf of the child. Oral consent information provided enough details to the study participants about the stool and blood sample collection process as well as the planned use of these specimens.
In 2011, the project was first implemented in two enhanced surveillance zones: Beira city in Sofala province, and Mocuba district in Zambézia province. In 2013, a third surveillance zone was established in Pemba city, Cabo Delgado province, following several cholera outbreaks there. These three areas were among the 8 districts out of the country's 145 districts with the highest number of cholera cases (>1000) reported to the Ministry of Health from 2009 to 2011 [5].
Beira city is the third largest city in Mozambique with a population of 463.442. Mocuba district includes 404.749 inhabitants and is rural area. Pemba City, with a population of 218.152, is the capital city of northernmost Cabo Delgado province. It has preserved largely rural characteristics with low population density.
The enhanced surveillance zones were selected based on: a high incidence of cholera during the previous five years; a sufficient large population denominator to determine incidence; a history of reliably providing surveillance reports on suspected cholera cases to the National MOH—Surveillance System; reliable health care access for the local population and diagnostic facilities to allow case identification. In these areas we included all health centers that treated the local population for severe diarrhea and collected additional data through community investigations. Usually, specific cholera treatment centers (CTC) were open only during cholera outbreak periods when declared by health provincial authorities. According to the area, the delay to open CTCs varied. Outside outbreaks patients with severe diarrhea were treated at general health centers or hospitals.
Africhol was integrated into the routine cholera surveillance and provided additional support beyond that supported by the MOH. For example, Africhol provided funding for specific additional staff such as a country focal point, a coordinator and a surveillance officer for each zone who ensured the proper conduct of the study, data collection, data quality and supervision of data entry. Also, in the enhanced surveillance zones, whenever possible most suspected cases were culture-tested, while in the routine system, only a few suspected cases at the start, during and at the end of an epidemic were usually tested for cholera. Africhol provided repeated refresher training for laboratory staff. Finally, Africhol provided standard case-report forms (CRFs) for data collection, laboratory standard operating procedures, support for specimen transport, support for serotyping, and technical assistance from the Africhol international team.
Data was collected prospectively through Africhol since its implementation in October 2011 until December 2015 in the three surveillance zones of Beira, Mocuba, and Pemba (Fig 1). This information was collected year round [11]. In addition, outside these enhanced surveillance zones, and using the same Africhol tools, we also investigated individual cholera outbreaks in three sites that experienced the greatest number of cases during the same time period, in Montepuez district (Cabo Delgado province) during February 2012, Cuamba district (Niassa province) during January-February 2012, and Nampula city (Nampula Province) every year from 2012 to 2015, as well as a few cases from other zones (mainly from Tete, Quelimane and Lichinga districts during outbreaks occurring in 2015). Those cases were investigated mainly from CTCs.
For Africhol, we defined a suspected case as an individual at least one year of age or older, with acute watery diarrhea and dehydration or must have died from acute watery diarrhea. A cholera case was considered confirmed when Vibrio cholerae O1 was isolated from a stool sample or rectal swab specimen by microbiological culture.
These definitions differed slightly from those used by the national MOH routine surveillance system for which a suspected case is defined as any patient aged two years or older, with acute watery diarrhea and abdominal pain, and profuse watery stools (type rice water stools), with vomiting or rapid dehydration. According to MOH definitions, once a case of cholera in a neighborhood has been confirmed, suspected cases are then considered cholera for a varying period of time, ranging from two to six months depending on the district.
Africhol data were collected through a CRF which included demographic, clinical, behavioral, and laboratory information. The rainy season was defined as the months of October to March, and the dry season from April to September.
Rectal swabs from suspected cholera patients were collected by health staff for laboratory investigation, and then transported from the health unit to the regional or provincial laboratories in Cary Blair transport media at room temperature conditions by a private courier services company once a week. In local laboratories samples were pre-enriched in alkaline peptone water and pre-incubated at 35–37°C for six to eight hours, followed by V. cholerae diagnostic tests carried out by sub-culturing on thiosulphate citrate bile salts sucrose agar (TCBS). After pre-processing at provincial level, the isolates cultured positive (showing yellow colonies) on TCBS were sent to the National Microbiology Laboratory at the Instituto Nacional de Saúde (INS) in the capital Maputo for further confirmation and quality control by standard biochemical tests and serology using polyvalent, anti-Ogawa, and anti-Inaba antisera. An aliquot of V. cholerae isolates was stored in the INS lab at -80°C and a copy was sent to the National Institute for Communicable Diseases (NICD) in South Africa for quality control and molecular subtyping and results were previously published elsewhere [12].
Blood samples were systematically taken for malaria and HIV testing in all sites from those patients that were in conditions to consent and accepted to have an HIV test. HIV counseling and testing were performed according to Mozambique’s national guidelines, which include confidentiality, counseling, and informed consent. Current guidelines for rapid testing call for a two test serial testing algorithm that screens with Determine HIV-1/2 (Alere, USA), and confirmation with Uni-Gold HIV (Trinity Biotech, Ireland). Malaria tests were performed by a health professional according to the national guidelines using one of several rapid tests available to the National Health Service.
From October 2011 to December 2015, a total of 1,863 suspected cholera cases were reported through the Africhol surveillance system. Among them, 1,010 (54.2%) were reported from the surveillance zones (Fig 2) and 853 (45.8%) from cholera outbreaks investigated in Cuamba, Montepuez, and Nampula (Fig 3). Additionally, a few cases were reported episodically from other zones, including mainly from Tete, Quelimane, and Lichinga during outbreaks occurring in 2015 (n = 151). Overall, the majority of cases (81.7%) were reported during the rainy season (October-March), especially in outbreak zones compared to surveillance sites (96.0% vs. 69.7%, p<0.001). Male-to-female sex ratio was 1.14 and varied between sites, from 1.49 in Nampula to 0.75 in Montepuez. The median age was 20 years (interquartile range (IQR), 9–32). There was a similar age distribution in all zones, except in Beira, which had the lowest median age (nine years) with the highest proportion of suspected cases under or equal to five years (38.4%, p<0.001) (Table 1). There were 101 children aged 12–23 months–mainly in Beira, thus representing 33% of the overall ‘below 5 years-old” age category (308).
The annual district-level incidence of suspected cases in surveillance zones ranged from 0.04 per 10,000 in Beira (2015) to 29.7 in Pemba (2013). In outbreak sites, the maximum attack rate was seen in Nampula city with 16.1 cases per 10,000 during the 2015 outbreak (Table 2).
A total of 972 suspected cases (52.2%) were tested and 228 (23.5%) of them were confirmed (Table 3; Fig 4). The rate of confirmed cholera varied between age groups (p<0.001), with the lowest rate among the age group ≤5 years (26/216, 12%), compared with confirmation rates ranging from 21% to 31% for other age groups. However, after adjusting these results on the surveillance zone, the confirmation rate was no longer associated with age. Only 4.5% of the children 12–23 months who were tested (4/89) were confirmed to be cholera by culture. Among the 207 children aged 2–5 years, 17.3% of those tested (22/127) were culture positive.
A total of 23 deaths were identified (CFR suspected, 1.2%), all of which occurred during the rainy season. The CFR was significantly higher in outbreak zones than surveillance sites (2.0% vs 0.6%, p<0.01), and in males compared to females (1.8% vs. 0.6%, p = 0.02). A total of 3 deaths occurred out of 228 cholera confirmed cases (CFR confirmed cases = 1.3%), all of them in Pemba city (Table 3).
In outbreak zones, most suspected cases were notified from CTCs, whereas in surveillance zones it varied (Table 1). The majority of cases consulted the health facility the same day or the day following symptom onset (70.4%). Most suspected cases were hospitalized (66.1%), except in Beira (5.6%). The majority of cases reported symptoms of watery stools (78.0%), dehydration (66.6%), and vomiting (59.9%), but this varied between sites (S1 Table).
Among all suspected cholera cases, 5% tested HIV positive, but most cases did not have HIV status determined (70%). The highest positivity rate was seen in Beira city (14%). For malaria, 8% of all suspected cholera cases had a positive rapid test result, with a high proportion in Mocuba district (24%), while 62% had unknown malaria status.
Most of the suspected cases had no contact with another suspected case, nor had they attended a funeral or a social event in the seven days prior. The public tap was the most common source of drinking water at home (48.2%), followed by a shallow well (18.3%). The majority of cases reported drinking untreated water (62.3%). Among those who treated water, bleach/chlorine was the most common treatment procedure (38.2%), followed by boiling water (22.3%) (S2 Table).
Factors independently associated with confirmed cholera in the multivariate analysis were: living in Nampula city district, the year 2014, HIV positive status, and the primary water source for drinking (Table 4). Factors associated with hospitalization included: male gender; young age; location; longer duration between date of onset and consultation; presence of rice water stools; vomiting; abdominal pains; leg cramps; HIV positive status; and receiving IV fluids before the consultation (S3 Table). Factors associated with death of suspected cholera cases included: male gender; short duration between disease onset and consultation; rice water stools; abdominal pain; and leg cramps (S4 Table).
Our analysis is in line with previous published data showing that cholera in Mozambique is marked by spatial heterogeneity and seasonality, with a high concentration of cases during the rainy period between January and March [5] and inter-epidemic periods largely free of confirmed cases. Each surveillance zone showed a different pattern of suspected cases distributed over time, although there were clear similarities in the seasonality of suspected cases between the surveillance and outbreaks zones. This heterogeneity may be explained by the hypothesis that in Mozambique outbreaks do not evolve locally, but rather follow cholera re-introduction from distant epidemic regions [7]. Our study allowed for the measurement of cholera incidence prospectively. Incidence of suspected cases varied widely between sites and between years within the same site. The incidence based on suspected cases only may be overestimated, given the low culture confirmation rate, as was shown previously in other Africhol countries [9].
There were 23 cholera deaths during the enhanced surveillance period (CFR, 1.2%), similar to what was found previously in Mozambique (0.9% during reported outbreaks in the period 2009–2011). Other studies showed the CFR was likely much higher under non-research conditions when immediate rehydration and transportation to hospitals may not be available [13]. Previous studies in Mozambique using surveillance data showed that 90% of deaths and 70% of cases occurred during the first six weeks of the outbreak [5].
Case distribution by sex and age group showed common patterns in all zones, except for Beira, where a high proportion of suspected cases occurred in persons under five years of age. In these areas, young children with symptoms of acute diarrhea are often brought to healthcare facilities by their mothers for treatment, while adults may be more hesitant to visit healthcare facilities. The proportion of confirmed cases in this age group in Beira was low (3.8%), but was similar to the other age groups from this zone (ranging from 0 to 9%). In another study in Beira, cholera incidence was higher among children below five years of age compared to older age groups [14].
Overall, the proportion of confirmed cases was lower in the one- to five-year-old age group than in older age groups. As we did not have the mandate or resources to study other pathogens that might also be associated with acute diarrhea in young children we cannot ascertain the attributable fraction of cholera to all-cause diarrhea. In the Global Enteric Multicenter Study (GEMS) conducted in four African sites and three Asian sites among children below 5 years, most attributable cases of moderate-to-severe diarrhea were due to four pathogens: rotavirus, Cryptosporidium, enterotoxigenic Escherichia coli producing heat-stable toxin (ST-ETEC), and Shigella [15]. Nonetheless, in one of the study sites located in southern Mozambique (Manhica), with traditionally low cholera incidence, the main pathogens identified among children 2–5 years-old were Shigella (14.9% of all moderate-to-severe diarrhea) and Vibrio cholerae (8.3%).
Our results indicate the suspected cases tended to rapidly seek care. This might be explained by the fact that people have the perception that cholera is a serious disease likely to result in death if left untreated. In Beira, there was a very low cholera confirmation and hospitalization. One of the possibilities for this finding is that some of the large outbreaks in Beira might be attributed to other pathogens.
Of the expected known exposures or risk factors, the only one that was independently associated with confirmed cholera in our study was the primary source of drinking water. Drinking water from unknown sources posed a greater risk for cholera. This could indicate wide-spread circulation of V.cholerae in the drinking water sources of the affected areas during outbreak periods resulting in a higher proportion of common source versus person-to-person transmission. In line with this, most cases occurred during the rainy season (82%), especially in outbreak sites with high incidence, with heavy floods possibly deteriorating water and sanitation system and triggering water borne transmission. Further studies using molecular biology methods, innovative approach to evaluate risk factors for cholera infection (eg. including community controls); small scale spatial epidemiological analysis and other studies such as description of the secondary infection at household and neighbor level should further elucidate the modes of cholera transmission. In parallel, majority of suspected cases reported that they had not attended a mass gathering or market in the seven days before symptom onset. This would underline the over proportional importance of continuous exposure through the water source at the place of residence rather than at isolated mass gathering events. Looking at all factors associated with confirmed cholera, we can see that risk factors are not that different for suspected and confirmed cases. This would reflect the common risk factors for cholera and other water-borne diarrheal diseases.
Although some association between HIV status and confirmed cholera was shown (adjusted OR, 95% CI: 4.5 [1.6–12.2]), this result should be interpreted with caution, because the majority of suspected cases (70%) and confirmed cases (54%) had unknown HIV status. Also, most HIV tests that were carried out and produced positive results came from one site (Beira). Additionally, we could not differentiate any HIV infection from HIV infection with immunosuppression, and HIV infection may simply be a marker for persons with less access to clean water and sanitation. A previous study in Beira indicate that persons with HIV infection had an increased risk of cholera compared to those without HIV infection, although this risk did not reach statistical significance (p = 0.08), and no information on immunosuppression status of enrolled patients was provided in this study [16].
The different cholera characteristics between zones highlight the need to tailor intervention strategies to the specific local setting and at-risk populations and are useful for evidence-based decision making. The data presented here were used to direct interventions to prevent/treat cholera such as timely and solid cholera surveillance system, improved environmental management in particular continued access to safe water and proper sanitation, and the adequate use of cholera vaccines as a complementary immediate measure. The high-risk zones where cholera outbreaks repeatedly occur (e.g., Nampula city in Nampula province, Mocuba district in Zambezia, and Pemba city in Cabo Delgado province) are so-called “cholera hotspots” and may benefit from preventive or reactive cholera immunization campaigns in combination with other cholera control activities (such as WASH activities), as recommended by the World Health Organization (WHO) [17]. The persistence of cholera over decades and the wide-spread risk of drinking contaminated water in those areas will require a decisive comprehensive effort from all stakeholders to improve the situation. No plans for such interventions are yet known and they will likely take years. In contrast OCV can provide rapid protection. A previous OCV campaign had been conducted in Beira city in 2003–2004 and showed a high vaccine effectiveness of one or more doses (78%, 95% CI: 39–92), even in a setting with high HIV prevalence [6]. More recently, in October 2016, another OCV campaign was conducted in Mozambique in Nampula city, targeting high-risk neighborhoods; monitoring and evaluation of this campaign is currently on-going.
Although our study provides valuable information about cholera surveillance in Mozambique, there are some limitations. Only 52.2% of suspected cases were tested and 23.5% were microbiologically confirmed. This limited capacity to confirm cases by culture combined with limited sensitivity of culture lowered the overall sensitivity of our surveillance [9]. Our diagnostic procedures relied exclusively upon culture results, which can be influenced by various factors including collection, transport and storage conditions, training of human resources or previous antibiotic exposure. Also, culture has a limited sensitivity which can translate in a low negative predictive value. It would have been preferable to use PCR testing, however, given resource limitations, we were unable to do so and thus our burden estimations need to be interpreted accordingly. Conversely, acute diarrhea cases might also have been reported as cholera while being caused by other pathogens. There was a high proportion of certain variables missing, which made it difficult to analyze the risk factors associated with confirmed cholera. In addition, data gaps limited our ability to conduct a robust examination of the association of cholera with malaria or HIV, including any differences in associations that may be due to immunosuppression status. The selection criteria for our surveillance included functional surveillance system and functional health structures with access to appropriate case management. Therefore the mortality related estimates cannot be generalized for the entire country which would likely underestimate mortality and CFRs. Moreover, our study did not account for the cholera cases and deaths in the community when the patients did not visit the health facilities. In this analysis we presented cholera cases in Africhol surveillance sites and some cholera outbreaks from 2011–2015. Compared with simultaneous official figures reported by the National Surveillance System, our surveillance system showed fewer cases reported in some of the districts, indicating that the Africhol surveillance was not exhaustive since it was limited to certain surveillance zones. However, the introduction of case-based surveillance methodology into the public-health system has allowed national staff to appreciate its value and expand this approach to other infectious diseases thereby reinforcing national surveillance capacity overall.
Our study provides the most comprehensive information on cholera in Mozambique available in recent years. This study does not aim to replace the national system but to assess the characteristics of populations at risk of cholera and risk factors in specific sites through an enhanced case-based surveillance. There is a need for continued surveillance, detailed data and stronger laboratory capacity to target prevention and control efforts, including locally adapted WASH interventions and preemptive use of OCV. In order to improve quality and access to safe water and sanitation, the Mozambican government has established investment funds and water supply assets which shall be used for public investment into the water supply system and the management of several companies across Mozambique.
The use of burden data in smaller-scale geographic units will help target higher risk neighborhoods (bairros) within the identified districts, and this work is ongoing.
|
10.1371/journal.ppat.1004045 | Coxsackievirus B Exits the Host Cell in Shed Microvesicles Displaying Autophagosomal Markers | Coxsackievirus B3 (CVB3), a member of the picornavirus family and enterovirus genus, causes viral myocarditis, aseptic meningitis, and pancreatitis in humans. We genetically engineered a unique molecular marker, “fluorescent timer” protein, within our infectious CVB3 clone and isolated a high-titer recombinant viral stock (Timer-CVB3) following transfection in HeLa cells. “Fluorescent timer” protein undergoes slow conversion of fluorescence from green to red over time, and Timer-CVB3 can be utilized to track virus infection and dissemination in real time. Upon infection with Timer-CVB3, HeLa cells, neural progenitor and stem cells (NPSCs), and C2C12 myoblast cells slowly changed fluorescence from green to red over 72 hours as determined by fluorescence microscopy or flow cytometric analysis. The conversion of “fluorescent timer” protein in HeLa cells infected with Timer-CVB3 could be interrupted by fixation, suggesting that the fluorophore was stabilized by formaldehyde cross-linking reactions. Induction of a type I interferon response or ribavirin treatment reduced the progression of cell-to-cell virus spread in HeLa cells or NPSCs infected with Timer-CVB3. Time lapse photography of partially differentiated NPSCs infected with Timer-CVB3 revealed substantial intracellular membrane remodeling and the assembly of discrete virus replication organelles which changed fluorescence color in an asynchronous fashion within the cell. “Fluorescent timer” protein colocalized closely with viral 3A protein within virus replication organelles. Intriguingly, infection of partially differentiated NPSCs or C2C12 myoblast cells induced the release of abundant extracellular microvesicles (EMVs) containing matured “fluorescent timer” protein and infectious virus representing a novel route of virus dissemination. CVB3 virions were readily observed within purified EMVs by transmission electron microscopy, and infectious virus was identified within low-density isopycnic iodixanol gradient fractions consistent with membrane association. The preferential detection of the lipidated form of LC3 protein (LC3 II) in released EMVs harboring infectious virus suggests that the autophagy pathway plays a crucial role in microvesicle shedding and virus release, similar to a process previously described as autophagosome-mediated exit without lysis (AWOL) observed during poliovirus replication. Through the use of this novel recombinant virus which provides more dynamic information from static fluorescent images, we hope to gain a better understanding of CVB3 tropism, intracellular membrane reorganization, and virus-associated microvesicle dissemination within the host.
| Enteroviruses are significant human pathogens, causing myocarditis, aseptic meningitis and encephalitis. The mechanisms of enterovirus dissemination in the host and cell-to-cell spread may be critical factors influencing viral pathogenesis. Here, we have generated a recombinant coxsackievirus expressing “fluorescence timer” protein (Timer-CVB3) which assists in following the progression of infection within the host. Unexpectedly, we observed the shedding of microvesicles containing virus in partially-differentiated progenitor cells infected with Timer-CVB3. These extracellular microvesicles (EMVs) were released in high levels following cellular differentiation, and may play a role in virus dissemination. Timer-CVB3 will be a valuable tool in monitoring virus spread in the infected host.
| Enteroviruses (EV) are among the most common and medically-important human pathogens, and a frequent cause of central nervous system (CNS) disease [1]. Worldwide distribution of EV infection is revealed by the detection of EV-specific antibodies in the sera of approximately 75% of individuals within developed countries. For example, in 1996, approximately 10–15 million diagnosed cases of EV infection occurred in the US alone [2].
Coxsackieviruses (CV), members of the enterovirus genus, are significant human pathogens, and the neonatal central nervous system (CNS) and heart are major targets for infection. CV infection causes severe morbidity and mortality, particularly in the very young. CV infection during pregnancy has been linked to an increase in spontaneous abortions, fetal myocarditis [3], and neurodevelopmental delays in the newborn [4]. Infants infected with CV have been shown to be extremely susceptible to myocarditis, meningitis and encephalitis with a subsequent mortality rate as high as 10%. Adult infection and subsequent viral myocarditis has also been described, and a substantial proportion of patients suffering from chronic viral myocarditis eventually develop dilated cardiomyopathy, a condition underlying almost half of all heart transplants. Severe demyelinating diseases may also occur following infection, including acute disseminated encephalomyelitis [5] and acute transverse myelitis [6]. Also, a number of delayed neuropathologies have been associated with previous CV infection, including schizophrenia [7] [8], encephalitis lethargica [9], and amyotrophic lateral sclerosis [10] [11]. Previously, we have shown that CVB3 preferentially targets neural progenitor and stem cells (NPSCs) in the CNS [12] [13] [14] [15] [16] [17]. Lasting consequences have be observed in the CNS following CVB3 infection [18], and NPSCs may represent a site of virus persistence in surviving mice infected shortly after birth [19] [20]. Also, CVB3 can infect the bone marrow and reduce hematopoietic progenitor cell populations [21].
We wished to more carefully observe CVB3 infection in the context of an ongoing Type I interferon response in order to visualize the dynamics of virus dissemination simultaneously with counteracting and protective antiviral responses generated in neighboring cells. Virus dissemination within the host may be an important consideration in predicting eventual pathogenesis in the host. Although challenging, identifying the sequence of infection upon initial virus exposure could be critical in preventing the natural disease course. For example, CVB3 has been shown to target the pancreas at early time points following infection. In addition to initiating acute pancreatitis, early virus replication in the pancreas may shed additional virions which disseminate in the host eventually reaching the heart or CNS. Organ-specific expression of interferon-γ within the pancreas was previously shown to reduce initial CVB3 replication and limit acute myocarditis in the host [22]. Therefore, hindering the normal progression of virus dissemination by targeting early sites of infection may be a possible strategy to protect the host.
In order to track virus dissemination both in cultured cells and in vivo, we generated a recombinant CVB3 expressing “fluorescent timer” protein (Timer-CVB3). “Fluorescent timer” protein encodes a mutated form (V105A, S197T) of the red fluorescent protein, drFP583 [23]. drFP583 protein has been shown to fluoresce green immediately following translation. The red fluorescence of the native protein, drFP583, develops over a periods of a few hours following autocatalytic modification. The engineered mutations in “fluorescent timer” protein resulted in this autocatalysis process being substantially delayed and hence fluoresces green for a greater period of time following translation. Immediately after translation, the mutant “fluorescent timer” protein shows strong green fluorescence, but over a period of ∼48 hours is gradually converted to red fluorescence most likely due to a conformational change. This progressive conversion from green to red fluorescence assisted in determining the sequence of infection and virus spread in real time within the host. In general, more dynamic information regarding viral infection was obtained from static fluorescent images utilizing Timer-CVB3 thereby showing viral spread among neighboring cells.
Intriguingly, extracellular microvesicles (EMVs) containing infectious virus were readily observed in cultures of differentiated progenitor cells infected with Timer-CVB3. Many of these shed microvesicles were comprised of microtubule-associated protein light chain 3 (LC3), a protein essential for the generation of double membrane autophagosomes in the cytosol of cells [24]. Through the use of this novel recombinant virus, we hope to gain a better understanding of CVB3 tropism, spread, and pathogenesis in our animal model of infection.
The “fluorescent timer” protein slowly turns from green to red fluorescence allowing for temporal discrimination of recently-infected and previously-infected cells. As shown in Figure 1, the gene encoding Timer protein followed by a polyglycine linker and an artificial viral 3Cpro/3CDpro protease cleavage site was inserted into the backbone of our infectious CVB3 plasmid, as described for other recombinant CVB3s [25] [26]. The artificial viral 3Cpro/3CDpro protease cleavage site efficiently results in the processing of foreign proteins such as the enhanced green fluorescent protein, dsRED protein, and “fluorescent timer” protein from the adjacent viral VP4 protein [26] [27] [28]. A high-titered stock of Timer-CVB3 was generated after transfection of the infectious plasmid in to HeLa cells.
Following the infection of HeLa RW cells with Timer-CVB3 virus stock, “fluorescent timer” protein expression was observed by fluorescence microscopy. Similar to previous studies for “fluorescent timer” protein [23], the transition from green to red fluorescence was observed over the course of 48 hours post-infection (PI). After 24 hours, infected cells predominantly expressed high levels of green “fluorescent timer” protein. By 48 hours PI, infected cells fluoresced both colors, although predominantly red. These results demonstrate that Timer-CVB3 can be used as a molecular timer to mark and follow infected cells temporally. The growth kinetics and plaque size of Timer-CVB3 was found to closely match eGFP-CVB3 and dsRED-CVB3, both of which contain foreign inserts of similar size [29]. Generally, we have observed a reduction in recombinant CVB3 growth kinetics dependent upon the size of the gene insert [25] [26]. Nonetheless, recombinant CVB3s maintain their infectivity in vivo causing neuropathology in the neonatal mouse model [18] [19] and pancreatitis in the adult mouse model [30] [31] similar to wild type CVB3.
HeLa cells were infected with Timer-CVB3, dsRED-CVB3, or eGFP-CVB3 and monitored by flow cytometry (Figure 2), a more sensitive and quantitative readout of viral protein expression allowing analyses at the single cell level. As expected, dsRED-CVB3 and eGFP-CVB3-infected HeLa cells expressed high levels of their respective fluorescent reporter proteins as early as 24 hours PI. However, Timer-CVB3-infected HeLa cells slowly changed fluorescence from green to red over 72 hours PI (moi = 0.01). HeLa cells infected with a greater moi (moi = 0.1) showed an increase in the number of green and red fluorescing cells at 24 hours PI, as compared to HeLa cells infected at the lower moi. Based on the kinetic data of live infected cultures observed by fluorescence microscopy shown in Figure 1, we might have anticipated the complete conversion of “fluorescent timer” protein from green to red in Timer-CVB3-infected HeLa cells harvested at early time points (for example, 24 hours PI) due to the delay between harvesting and ensuing analysis by flow cytometry. However, cells were fixed in 4% paraformaldehyde shortly after harvesting and prior to flow cytometry. Subsequently, the gradual change in fluorescence for Timer-CVB3-infected HeLa RW cells could be arrested in the presence of 4% paraformaldehyde, suggesting that the tertiary structure of “fluorescent timer” protein is stabilized by formaldehyde cross-linking reactions.
By standard plaque assay, Timer-CVB3 plaques were expected to reveal a “bull's-eye” pattern by fluorescence microscopy at 44 hours PI, whereby initial infection is represented by red fluorescence and newly-infected cells via cell-to-cell spread is represented by green fluorescence. The progression of infection was followed over time in HeLa cells infected with Timer-CVB3 and overlayed with agar which hindered virion dispersal in the culture medium. As expected, a gradual change in fluorescence was observed for a single viral plaque in a composite of fluorescence images taken from agar-overlayed infected-HeLa cells at 44 hours PI (Figure 3A). Higher magnification of a segment of the viral plaque revealed a progressive shift from red to yellow to green cells away from the plaque center (Figure 3B). Cells showing cytopathic effects were also evident and these cells exhibited reduced or absent levels of “fluorescent timer” protein. In contrast, HeLa cells infected with Timer-CVB3 (moi = 0.1) and grown in the absence of an agar plug showed diffuse infection and virus spread, although individual cells fluoresced green and yellow at 24 hours PI (Figure 3C and Figure 3E). By 48 hours PI, the majority of HeLa cells grown in complete media in the absence of an agar plug fluoresced red (Figure 3D and Figure 3F).
Neural progenitor and stem cells (NPSCs) grown in culture as free-floating clusters of stem cells or “neurospheres” were recently shown to be highly susceptible to CVB3 infection [12] [19] [28]. We anticipated that the outer shell of stem cells first becomes infected, after which infection extends inward by cell-to-cell spread of CVB3. Therefore, NPSCs were infected with Timer-CVB3 (moi = 0.1), and the progression of infection was monitored over time by fluorescence microscopy (Figure 3G, Figure 3H, Figure 3I, and Figure 3J). By 32 hours PI, Timer protein expression outlined the progressive infection of the outer (yellow-fluorescing, white arrow) and inner stem cells (green fluorescing) within the neurosphere (Figure 3K). By 72 hours PI, red and yellow fluorescing cells were seen within disrupted neurospheres, and signs of cytopathic effects were readily evident. Intriguingly, pretreatment of NPSCs with poly IC, a synthetic analogue of double-stranded RNA known to interact with toll-like receptor 3, reduced the progression of infection as determined by “fluorescent timer” protein expression (Figure 3L, Figure 3M, and Figure 3N). These results suggest that NPSCs respond to immunostimulatory molecules and mount a protective antiviral response following CVB3 infection. No “fluorescent timer” protein signal was observed in mock-infected NPSCs by fluorescence microscopy (Figure 3O).
Our recent results suggest that CVB3 established a carrier-state infection in NPSCs [12] [29]. Carrier-state infection in NPSCs may be the result of continuous virus replication, or alternatively, sporadic reactivation of virus. We expect that the utilization of Timer-CVB3 will be useful in distinguishing these two models of carrier-state infection by examining “fluorescent timer” protein expression. However, timer protein expression was below detection limits despite the presence of infectious virus (as determined by plaque assay) in Timer-CVB3-infected NPSC carrier state cultures (Figure 3P).
Host cells infected with RNA viruses induce protective molecules such as MAVS through RIG-I and MDA5 activation which act as pattern recognition receptors [32]. The host cell undergoes an antiviral state and produces Type I interferons such as interferon-α (IFN-α) and IFN-β which act to protect neighboring cells. The Type I IFN response enables neighboring cells to express protective molecules such as RNase L which limits the spread of viral infection.
We inspected the progression of Timer-CVB3 infection in HeLa cells activated by the Type I IFN response to determine the pattern of virus spread within neighboring cells in the context of an active host antiviral response. Untreated HeLa cells, or HeLa cells treated with IFN-β or Poly IC were infected with Timer-CVB3 at a low (moi = 0.01) or greater moi (moi = 0.1), and the progression of infection was monitored by fluorescence microscopy (Figure 4). Untreated HeLa cells infected with Timer-CVB3 at a low moi initially fluoresced green at 24 hours PI (Figure 4A). By 32 hours PI, yellow and green cells appeared, and the majority of HeLa cells fluoresced yellow or red by 48 hours PI. In both IFN-β and poly IC-treated HeLa cells infected at a low moi, infection appeared delayed at 24 hours PI. Also, fewer infected cells were observed at 32 hours PI for IFN-β -treated HeLa cells, and these cells predominantly fluoresced green. By 48 hours PI, both IFN-β and poly IC-treated HeLa cells fluoresced yellow and red similar to untreated HeLa cells. These results suggest that the Type I IFN response partially protected HeLa cells from a low inoculum of Timer-CVB3 and initially controlled the progression of infection. In contrast, little protection was observed in HeLa cells infected with a higher inoculum of Timer-CVB3 (Figure 4B). This observable yet modest protection against CVB3 may be partly due to the relatively weak Type I IFN response induced by HeLa cells, and also the ability of CVB3 to inactivate key antiviral molecules. For example both MAVS and TRIF, known to be critical components of the innate immune response, have been shown to be cleaved in CVB3-infected cells [33].
The results showing a limited reduction in Timer-CVB3 progression in HeLa cells infected with a low or greater viral inoculum were quantified using ImageJ analysis (Figure 5A and Figure 5B, respectively). Despite the delay of Timer-CVB3 infection as determined by fluorescence microscopy following IFN-β or poly-IC treatment, no significant difference in viral titers was observed over time in HeLa cells infected with a low or high viral inoculum (Figure 5C and Figure 5D, respectively). These results suggest that monitoring the progression of infection using “fluorescent timer” protein may be a more sensitive method of determining the efficacy of an antiviral response within infected cells.
Separately, HeLa cells were infected with Timer-CVB3 at moi = 0.01 or 0.1 and treated with two concentrations of ribavirin (10 µg/mL and 100 µg/mL, respectively; Figure S1). A clear step-wise reduction in cytopathic effects, and “fluorescent timer” protein conversion and expression levels was observed in HeLa cells (moi = 0.01) treated with increasing amounts of ribavirin at 24, 32, and 48 hours PI (Figure S1A). Although less striking, HeLa cells infected at a higher moi (moi = 0.1) and treated with two concentrations of ribavirin also showed delayed cytopathic effects, and reduced levels of “fluorescent timer” protein conversion and expression levels (Figure S1B). These results were quantified using ImageJ analysis to show a substantial difference in the percentage of ribavirin-treated HeLa cells expressing “fluorescent timer” protein over time (Figure S2A and Figure S2B). Also, a significant delay was observed in the conversion of recent viral protein to mature viral protein at 24, 32, and 48 hours PI as compared to untreated HeLa cells. In addition, a step-wise reduction in Timer-CVB3 viral titers was observed in HeLa cells treated with ribavirin, although the reduction was less dramatic in HeLa cells infected with a higher inoculum of virus (Figure S2C and Figure S2D).
We previously described the ability of CVB3 to preferentially replicate in undifferentiated NPSCs [12]. However, a limited amount of viral replication was also observed in differentiated NPSCs comprising of a mixture of progenitor cells, and early neuronal, astrocytic, and oligodendritic cell lineages. Also, the relationship between CVB3 infection and autophagy in differentiated NPSCs proved to be quite distinct and unique as compared to either undifferentiated NPSCs or traditional cell lines [28]. We wished to follow the progression of Timer-CVB3 infection in mixed cell lineages comprising differentiated NPSC cultures. For example, preferential infection of progenitor cells might be anticipated, and delayed infection of cell lineage-committed cells might be revealed by a shift in “fluorescent timer” protein expression pattern whereby progenitor cells fluoresce red and more lineage restricted cells fluoresce green at a particular time point following infection in these mixed cultures. Therefore, NPSCs were differentiated for 5 days and infected with Timer-CVB3.
After 48 hours PI, cells were fixed and stained for progenitor (nestin+), neuronal (neuronal beta-tubulin III+), astrocytic (GFAP+), and oligodendritic (MBP+) cell lineage markers (Figure 6). Four-color fluorescence microscopy identified recent and matured virus protein expression in combination with each cell lineage marker, and DAPI was utilized to identify cell nuclei. The distribution of recent and matured virus protein appeared similar in nestin+ progenitor cells and in the three cell lineages, suggesting that progenitor cells and each cell lineage were equally susceptible to Timer-CVB3 infection upon initial infection.
Previously, we described cellular blebbing in partially differentiated NPSCs infected with eGFP-CVB3 by Hoffman modulation contrast microscopy [12]. Interestingly, cell-associated and cell free microvesicles comprising recent and matured virus protein were readily observed at higher magnification in differentiated NPSCs infected with Timer-CVB3 (data not shown). We hypothesized that these cell-associated and cell-free microvesicles were reminiscent of the cellular blebbing phenomena observed in infected differentiated NPSCs described previously.
In order to better observe cell-associated microvesicles following CVB3 infection, time-lapse videos were constructed from merged fluorescence images (recent and matured virus protein with HMC) of differentiated NPSCs infected with Timer-CVB3. Merged images were taken every 15 minutes for 7 hours to construct time-lapse videos (Video S1). The conversion of “fluorescent timer” protein was monitored over time in Timer-CVB3-infected cells which showed in parallel extensive intracellular membrane remodeling during infection. We hypothesized that the appearance of intracellular microvesicles expressing “fluorescent timer” protein represented the emergence of virus replication organelles recently described by others for CVB3 [34]. Intriguingly, the presence of numerous extracellular microvesicles (EMVs) was also identified in supernatants of the cell culture (Figure 7), although some of these particles may represent cellular debris. Many EMVs contained matured virus protein rather than recent virus protein, as evident by the detection of red “fluorescing timer” protein. Furthermore, some shed microvesicles containing viral protein appeared to attach and remain stationary on uninfected neighboring cells. Images from the time-lapse video of differentiated NPSCs infected with Timer-CVB3, or separately, differentiated C2C12 myoblast cells infected with Timer-CVB3, are shown in Figure 7. The progression of infection in differentiated NPSCs as determined by “fluorescent timer” protein was captured in revealing images taken at 47, 48, and 53 hours PI (Figure 7A, Figure 7B, and Figure 7C, respectively). Higher magnification showed the compartmentalization of distinct intracellular viral replication organelles of varying fluorescence color following Timer-CVB3 infection (Figure 7D), and the accumulation of EMVs (Figure 7E, white arrows) some of which fluoresced red (Figure 7E, pink arrow). We suspect that many additional shed microvesicles contain virus protein albeit at low or undetectable levels as judged by fluorescence microscopy.
Of note, differentiated C2C12 myoblast cells infected with Timer-CVB3 also gave rise to red fluorescing EMVs, beginning at 3 days, and also at 5, 7, and 9 days PI (Figure 7F, Figure 7G, Figure 7H, and Figure 7I, respectively). Mouse C2C12 cells are a myoblast cell line capable of differentiation into myocyte cells. Higher magnification revealed the presence of red fluorescing microvesicles near a green fluorescing cell, suggesting a new infection event had taken place (Figure 7J, pink arrow). By 9 day PI, high magnification showed red fluorescing microvesicles near more defined myocyte cells with signs of cytopathic effects (Figure 7K, red arrow). The presence of numerous red EMVs in two contrasting models of infected differentiating progenitor cells suggests that the differentiation process plays a role in their formation and release from the target cell.
High magnified image frames from the first 90 minutes of the time-lapse video shown in Video S1 revealed the progression of intracellular membrane remodeling and “fluorescent timer” protein expression reminiscent of CVB3 replication organelles described previously by others [34] (Figure 8). As shown in frame 9, the loss of a red fluorescing intracellular membrane complex with matured virus protein represented potential microvesicle egress from the infected cell (Figure 8B–J; red arrow). A high magnification time-lapse video of this area (Video S2) and other areas (Video S3 and Video S4) also revealed the possible ejection of additional microvesicles from the infected cell.
We inspected the location of ‘fluorescent timer” protein with an authentic CVB3 protein within differentiated NPSCs (Figure 9; two representative images shown). An antibody against viral 3A protein was utilized for colocalization experiments since this viral protein has previously been shown to reorganize the host secretory trafficking pathway and facilitate the recruitment of host proteins necessary to form specialized organelles critical for plus-strand RNA virus replication [34]. Partially differentiated NPSCs were infected with Timer-CVB3 for 4 days and fixed with 2% paraformaldehyde. Optical sections of fluorescently stained cells using the viral 3A antibody revealed a close colocalization with both recent and matured “fluorescent timer” protein in differentiated NPSCs utilizing a Zeiss Axio Observer with ApoTome Imaging System (Figure 9A and Figure 9B; white arrows). Intriguingly, the viral 3A protein was preferentially detected in cell-associated vesicles. Both recent and matured “fluorescent” protein was also detected in cell-associated vesicles, although the signal was more diffuse as compared to viral 3A protein. Closer inspection of a cell-associated microvesicle (high magnification) revealed the presence of recent (green), matured (red), and viral 3A protein (blue) (Figure 9B; cyan arrow). These results suggest that “fluorescent timer” protein was preferentially found in regions of the infected cell where active viral replication organelle complexes might be expected.
Previous studies with poliovirus described the possible contribution of the autophagic pathway to virus egress from the host cell [35]. Therefore, we transduced differentiated NPSCs with an adenovirus expressing LC3-GFP as described previously [28], and followed autophagosome formation following infection with dsRED-CVB3 (Figure 10). By day 1 PI, cells expressed LC3-GFP and detectable levels of virus protein (red) (Figure 10A). Cellular blebbing associated with LC3-GFP signal was seen in infected cells as early as day 2 PI (Figure 10B; white arrow). By day 3 PI, abundant EMVs were observed in culture containing both viral material (red signal) and LC3-GFP protein (green signal) (Figure 10C and Figure 10D; white arrows). Higher magnification further revealed the size and structure of representative shed microvesicles which contained both viral material and LC3-GFP (Figure 10E and Figure 10F; white arrows). These results suggest that the observed EMVs following CVB3 infection arise from the autophagy pathway shown previously to be activated following infection [36] [37] [30] [38] [39] [40] [28].
Differentiated C2C12 were infected with eGFP-CVB3 for three days, and virus particles within supernatants were resolved in isopycnic iodixanol gradient fractions (Figure 11A). Fractions #22-24 contained infectious virus particles at a density expected for picornavirus virions (1.22 g cm−3). However, significant levels of infectious virus were also observed in low density fractions (Fractions #6-21) consistent with membrane association (1.04–1.10 g cm−3). The amount of infectious virus associated with low density fractions (Fractions #6-21) represented a significant percentage (approximately 21.5%) of the total infectious virus identified for all fractions. Also, the relatively wider range of low density fractions containing infectious virus suggested a broader density spectrum of membrane-associated virus, as compared to enveloped hepatitis A viruses resembling exosomes of a more restricted range of low density (1.06–1.10 g cm−3).
EMVs were also isolated from infected C2C12 supernatants utilizing Exoquick-TC polymer-based exosome precipitation kit. Purified EMVs were resuspended in 1× PBS and inspected by fluorescence microscopy for viral protein (eGFP) and size distribution using the Length Interactive Measurement feature of AxioVision software (Figure 11B). Numerous microvesicles expressing high levels of eGFP were identified, and particles were quite diverse in size ranging between 0.51–5.53 µm in diameter (Ave = 1.31 µm; Median = 0.92 µm; SD ±0.96), although microvesicles smaller than 0.5 µm were also shown to be present by transmission electron microscopy. The diversity in EMV size and shape might also explain the relatively wide range of low density iodixanol fractions containing infectious virus, as shown in Figure 11A. We tested the ability of purified EMVs to infect HeLa cells following a one hour incubation (Figure 11C). eGFP+ EMVs added to HeLa cells were visualized in the cultures shortly before (0 Hour PI) and after (1 Hour PI) the incubation period (Figure 11C; High Mag). By 3 hours PI, eGFP+ EMVs were no longer observed presumably due to fusion of the EMVs with HeLa cells and dilution of viral protein signal. By 24 hours PI, high levels of viral protein expression and cytopathic effects were readily seen in HeLa cells. These results suggest that the kinetics of EMV-associated CVB3 infection in HeLa appear similar to free infectious virus described previously for eGFP-CVB3 [26].
Differentiated C2C12 or NPSCs were infected with eGFP-CVB3 for three days, and EMVs were isolated utilizing Exoquick-TC polymer-based exosome precipitation kit. EMV precipitates and supernatant fractions from differentiated C2C12 and NPSC cultures were also examined for levels of infectious virus (Figure 12A). Purified EMVs resuspended with 1× DMEM at an equal volume compared to supernatant fractions were inspected by standard plaque assay. Intriguingly, EMV precipitate fractions for both differentiated C2C12 and NPSC cultures (Figure 12A; green bars) comprised a greater concentration of infectious virus compared to the supernatant fraction (Figure 12A; pink bars) utilizing the ExoQuick-TC isolation procedure. Freeze/thaw treatment of EMVs substantially reduced the amount of infectious virus (Figure 12A; green hatched bar). In contrast, freeze/thaw treatment of the supernatant fraction showed no reduction in viral titers (Figure 12A; pink hatched bar). These results suggest that infectious virus was preferentially associated with shed microvesicles released from differentiated C2C12 and NPSC cultures. Also, a portion of virus associated with EMVs may reflect immature particles requiring EMVs for infectivity [41], or may indicate the presence of EMV-associated viral RNA which remains infectious following microvesicle-assisted fusion of uninfected target cells. We suspect that the relative percentage of infectious virus associated with low density iodixanol gradient fractions shown in Figure 11A might be under-represented due to a reduction in viral titers following freeze-thaw of EMVs within the collected fractions. Nevertheless, the two independent methodologies utilized to purify EMVs indicate that a substantial amount of infectious CVB3 may comprise a membrane-associated form.
Purified EMVs were also inspected by western analysis for the presence of CVB3 viral protein 1 capsid protein (VP1), LC3 I and II, and flotillin-1 (Figure 12B). LCII (lipidated form of LC3) studs the inner and outer autophagosome membrane and reflects autophagic activity in cells [24]. Flotillin-1 is a caveolae-associated membrane protein identified as a common marker on exosomes [42], and at compartments of the endocytic and autophagosomal pathways. As shown in Figure 12B, high levels of VP1 were identified in purified EMVs isolated from both differentiated NPSC and C2C12 infected with eGFP-CVB3. Also, both NPSC and C2C12 EMVs were comprised mainly of LC3 II protein, and high levels of flotillin-1 protein were also observed by western analysis. The preferential detection of LC3 II by western blotting might be expected if EMVs originate as double membrane autophagosomes, fuse with the plasma membrane, and are released as single membrane vesicles [43] [35]. Taken together, these results verified the presence of viral proteins and infectious virus in EMVs, and demonstrate possible autophagosome-mediated exit of shed infectious vesicles reminiscent of the “autophagosome-mediated exit without lysis” pathway (AWOL) described previously by Jackson and colleagues for poliovirus infection [35].
We inspected the molecular structure of EMVs In order to clarify their association with infectious virions. Purified EMVs from Timer-CVB3 or mock-infected differentiated C2C12 cells were processed for transmission electron microscopy (TEM). For infected C2C212 cells, virus-like particles enclosed in membrane structures were readily observed by TEM (Figure 13). Single virions were seen within small microvesicles (Figure 13A; green arrow), although free virions were also observed in sections (Figure 13A; pink arrow). We suggest that free virions may represent EMV-associated virions which were disrupted during extensive processing for TEM. Also, virus-like particles were observed in EMVs of various sizes (Figure 13B; green arrows), perhaps reflecting the broad range of densities identified for infectious EMVs following iodixanol gradient purification. Higher digital magnification revealed the icosahedral shape of these virus-like structures (Figure 13C; dashed green polygon), and the diameter of virions was close to 31 nm - similar to previous studies describing the crystal structure of CV [44]. In some cases, EMVs contained three or four virions (Figure 13D and Figure 13E, respectively; green arrows). In contrast, microvesicles isolated from mock-infected C2C12 cells failed to harbor virus-like structures (Figure 13F). These microvesicles were more uniform in size (approximately 100–150 nm) and may represent traditional exosomes released by C2C12 cells [45].
Tracking viral infection may be critical in understanding viral dissemination and pathogenesis. For example, vaccinia virus has been shown to undergo repulsion of superinfection by the early expression of A33 and A36 protein which allows the virus to increase viral spread and maximize the replication rate substantially [46]. Also, human T-lymphotropic virus-1 (HTLV-1) utilizes the virological synapse, a specialized area of cell-cell contact promoting transmission of virus [47]. Much remains to be determined with regards to cell-to-cell spread of CVB3 and virus dissemination within the host [27]. The pancreas is a primary target organ for CVB3 infection in mice, and early viral replication here may seed other organs such as the heart and CNS. For example, transgenic mice expressing the antiviral molecule interferon-γ within the pancreas showed reduced viral titers and substantially less myocarditis [22]. Mice are considered an informative model to evaluate the mechanisms of coxsackievirus pathogenesis, and many of diseases caused by coxsackieviruses have been recapitulated in a mouse model [48]. Importantly, understanding factors contributing to early pancreatic infection may help to reduce viral replication and virus dissemination within the host. Therefore, we engineered the “fluorescent timer” protein into our infectious clone of CVB3 in order to monitor virus dissemination both in culture and in vivo. “Fluorescent timer” protein containing a nuclear localization signal has been previously utilized to track the cytoplasmic accumulation of nuclear proteins within infected cells [49]. However, this unique molecular marker has not been employed to follow the progression of virus infection directly.
By tracking the progression of virus infection utilizing Timer-CVB3, we observed intriguing aspects of virus replication in partially differentiated progenitor cells which we did not anticipate. First, we identified extensive intracellular membrane remodeling in partially differentiated NPSCs which reflecting the formation of viral replication organelles described by others [34]. By time-lapse video, the conversion of “fluorescent timer” protein from green to red as monitored by fluorescence microscopy revealed distinct regional viral replication organelles fluorescing in an asynchronous fashion. Previous studies have described the ability of viral 2B and 3A proteins of CVB3 to ablate protein trafficking and secretion by disrupting the Golgi apparatus [50] [51] [52]. Also, CVB3 has been shown to modulate cell host factors such as PI4KIIIβ, GBF1 and ARF1 in order to remodel intracellular membranes for efficient viral replication [53] [34] [54]. We suggest that ongoing translation during virus replication and the lack of virus protein transit due to compartmentalization and formation of CVB3 replication organelles within the host cell reflects the localized conversion of “fluorescent timer” protein from green to red as shown by time-lapse video.
The “fluorescence timer” protein contains a polyglycine linker and a 3Cpro/3CDpro cleavage site immediately downstream of the viral polyprotein start codon. Hence, “fluorescence timer” protein served as a marker for infected cells similar to other fluorescence proteins, such as eGFP and dsRED, described by our previous studies [26] [27]. Nevertheless, “fluorescence timer” protein may eventually diffuse away from sites of ongoing viral RNA translation and replication within the cell. Therefore, we determined colocalization of mature and recent viral “fluorescence timer” protein with viral 3A protein, an authentic viral protein previously shown to be closely associated with coxsackievirus replication complexes. Colocalization of “fluorescent timer” protein and viral 3A protein was readily observed in partially differentiated NPSCs infected with Timer-CVB3. Viral 3A protein has been shown previously to play a critical role in the reorganization of the secretory pathway and the generation of CVB3 replication organelles [34]. Therefore, following “fluorescent timer” protein signal during live imaging of infected cells may reveal the dynamics of intracellular membrane rearrangements critical for viral replication.
Based on these new data, we suggest that viral replication organelles physically retain proteins following viral RNA translation, and that potential diffusion of “fluorescence timer” protein away from these replication complexes may be restricted following the assembly of virus-modified intracellular membranes. Subsequently, the unique features of Timer-CVB3 have enabled the direct visualization of intracellular membrane remodeling of the infected host cell in real time. The induction and visualization of Timer-CVB3 replication organelles utilizing “fluorescent timer” protein also reflects the ability of enteroviruses to hijack the autophagy pathway and maximize viral replication [31] [30] [37] [55].
Utilizing Timer-CVB3, we monitored the generation of extracellular microvesicles (EMVs) containing viral material which may represent a novel strategy of virus shedding and transmission to neighboring cells. A large numbers of EMVs were released from partially differentiated NPSCs and in C2C12 myoblast stem cells following infection with Timer-CVB3. The extensive intracellular membrane remodeling observed in differentiated NPSCs and C2C12 cells following Timer-CVB3 infection may eventually contribute to microvesicle shedding similar to the formation of exosomes in cell lines, tumor cells and other cell lineages [45]. We initially considered the possibility that shed EMVs represented “cellular debris” released from dying cells. However, we feel that shed EMVs represent an ordered process commandeered by the virus to escape the host cell for at least five reasons: 1) First, we have previously shown little to no cytopathic effects in five and sixteen-day differentiated NPSCs infected with eGFP-CVB3 [12]. 2) Time-lapse photography revealed continuous shedding of intact EMVs from infected cells which remained alive and undamaged (Figure 7). 3) Shed EMVs characteristically contained the mature form of “fluorescence timer” protein, as opposed to recent, or recent plus mature “fluorescence protein” which might be expected if virus-induced cytopathology contributed to the unregulated release of cellular debris (Figure 7). 4) EMVs retained considerable stability in culture (Figure 10) and during substantial and lengthy procedures which included isopycnic gradient centrifugation and Exoquick-TC precipitation (Figure 11). 5) Finally, inspection of EMV-associated proteins (including LC3-II) and infectious virus indicated unique cellular protein signature patterns for EMVs, and the preferential association of virions with shed EMVs (Figure 12). We suggest that the release of EMVs harboring infectious virus represents an active and controlled process involving the autophagy pathway.
We examined the presence of autophagic proteins within EMVs released from differentiated NPSCs following CVB3 infection. Differentiated cells were transduced with LC3-GFP and infected with dsRED-CVB3. Shed microvesicles expressing both LC3-GFP and virus protein were readily observed. EMVs isolated from infected NPSCs and C2C12 cells were found to comprise high levels of infectious virus, and virus-like particles were readily observed in EMVs utilizing transmission electron microscopy (TEM). A broader range of densities in membrane-associated fractions were identified for infectious EMVs utilizing iodixanol gradient purification, as compared to enveloped hepatitis A viruses (eHAVs) resembling exosomes [41] [56]. These results are consistent with the broad size distribution range observed for EMVs (Figure 11B), and TEM data showing virion-comprising EMVs of various shapes and complex membrane morphology (Figure 13). Furthermore, western analysis of purified EMVs indicated the presence of viral protein 1 (VP1), flotillin-1 (a protein previous shown to be found in exosomes) and the preferential detection of the lipidated form of LC3 (LC3 II). These results suggest that the autophagic pathway contributes to CVB3 shedding similar to the previously described autophagosome-mediated exit without lysis (AWOL) model for poliovirus release [35]. Although proposed single-membrane vesicles derived from the autophagosome pathway and containing poliovirus have been suggested to be short-lived, our results with CVB3 suggest that these structures may be considerably more stable than anticipated [57]. Also, we recently described a reduction in the level of intracellular autophagosomes in differentiated NPSCs following CVB3 infection [28], an observation which contrasts with our results observed for HL-1 cells (cardiomyocyte cell line) or undifferentiated NPSCs infected with CVB3. Based on our new data, we speculate that the observed reduction of intracellular autophagosomes in differentiated NPSCs infected with CVB3 in our previous study is the result of the ejection of autophagosomes specifically within progenitor cells undergoing differentiation [43].
Previous publications have suggested the possible contribution of EMVs or exosomes to virus dissemination [42] [58]. A recent study has shown that hepatitis C viral RNA transfer via exosomes are sufficient to activate plasmacytoid dendritic cells and induce an innate immune response in the host [59]. Also, hepatitis A virus, previously considered to be a non-enveloped virus, was recently revealed to hijack cellular membranes and escape the host cell as enveloped viruses resembling exosomes [41] [56]. Shed microvesicles share assembly pathways with retrotransposon elements and viruses, suggesting that viruses exploit cellular microvesicle pathways to maximize dissemination [60]. Also, double-membrane autophagic vesicles induced following enterovirus infection may fuse with the cell membrane leading to possible extracellular release of either single-membrane-enclosed or free virions from the cytosolic lumen [61]. Shed microvesicles produced in cells may ultimately assist in CVB3 egress from infected cells in the absence of cytolysis [35], and may provide a successful route of transmission to uninfected cells despite the presence of host neutralizing antibodies [62]. Our model of virus dissemination during the differentiation of progenitor cells by the production of EMVs, and the potential use of Timer-CVB3 to characterize this model is shown in Figure 14.
Although we have shown here that EMVs carry infectious virions, viral material shuttled by shed microvesicles during later stages of infection may be limited to infectious viral RNA – perhaps representing a novel strategy adapted by CVB3 to infect neighboring cells during persistence in the absence of virus capsid formation or cellular cytopathicity [20]. Supporting this idea, other groups have shown the ability of exosomes to transport RNAs or micro RNAs from cell to cell [42] [63]. The migratory nature of progenitor cells provides CVB3 with a unique vehicle to disseminate within the host [19], and that cellular differentiation may be a signal for the virus to escape by extracellular microvesicle-mediated discharge. Does CVB3 induce novel autophagosome ejection from infected cells, or does the virus commandeer a natural process for its own selfish purposes? We suggest that CVB3 may upregulate and exploit a natural process which may occur to a limited degree within progenitor cells during differentiation. EMV release may be significantly enhanced in differentiating cells either due to their inherent role in the differentiation process [64] [65] [66], or due to virus-induced enhancement and release during cellular differentiation. Also, we suggest that virus dissemination by EMVs may expand tropism through fusion with cells which fail to express the coxsackie and adenovirus receptor.
Our observation of red fluorescing particles nearby green fluorescing cells (shown in (Figure 7J, red arrow) provided the early inspiration for examining the presence of infectious shed microvesicles in Timer-CVB3-infected C2C12 cells. The unique qualities of Timer-CVB3 contributed to the formulation of our original hypothesis of shed infectious microvesicles by providing kinetic data from a static fluorescence image which would not normally be revealing. For example, if we had simply determined viral protein expression utilizing a single channel color for Figure 7J, we might have simply concluded the presence of cellular debris near an infected cell. Furthermore, the use of Timer-CVB3 in evaluating antiviral compounds of unknown function may provide additional mechanistic information for mode of action for a given compound. For example, an antiviral compound which may hypothetically act to inhibit virus release, although not initial entry, may lead to the presence synchronized green cells, followed later by red cells and the absence of newly-infected green cells. In contrast, an antiviral compound which decreases viral replication or translation may be reflected by the early appearance of dim green cells, followed later by both dim green and dim red-infected cells.
A possible limitation of observing “fluorescent timer” protein expression in infected tissue is that cells are required in this model to be observed directly, and some methodologies (such as flow cytometry or histology) necessitate significant processing time. However, the results of flow cytometry data using HeLa RW cells infected with Timer-CVB3 at various time points over 72 hours PI and fixed with 4% paraformaldehyde (Figure 2) demonstrate that the conversion of “fluorescent timer” protein can be stopped for subsequent analysis at a later time point. Timer-CVB3 will be of particular value when tracking infection in vivo, whereby initial time of infection is not clear, or in mixed cell cultures where one cell type may be first infected followed sequentially by less susceptible cell types. Timer-CVB3 will also assist in characterizing persistent infection in the heart or CNS, whereby the nature of persisting viral material is unclear. For example, new viral protein expression during sporadic infection or virus reactivation might be expected to produce green fluorescence in cells harboring Timer-CVB3 infection, while a carrier-state infection would produce red and green fluorescence simultaneously.
Future studies will focus on isolating infectious EMVs from the sera of mice in the presence of an ongoing neutralizing antibody response. Also, we will determine the ability of shed microvesicles to expand the tropism of CVB3, and identify inhibitors of microvesicle fusion which might act as novel antiviral compounds.
This study was carried out in strict accordance with the requirements pertaining to animal subjects protections within the Public Health Service Policy and USDA Animal Welfare Regulations. All experimental procedures with mice were approved by the San Diego State University Institutional Animal Care and Use Committee (Animal Protocol Form #10-05-013F), and all efforts were made to minimize suffering.
The Timer gene was amplified from the pTimer plasmid (Clontech Inc.) with primers containing SfiI restriction site sequences. Following amplification, the PCR product was cut with Sfi1 and cloned into our parental CVB3 vector (pMKS1) linearized with SfiI [25]. Also, a CMV promoter was cloned into the Timer-CVB3 plasmid. The plasmid was precipitated by ethanol incubation at −20°C for 12 hours. After centrifugation and washing in 70% ethanol, the DNA pellet was air-dried at room temperature for 20 minutes. The pellet was then resuspended in sterile, nano-filtered water at a concentration of 1 microgram/microliter.
The generation of recombinant CVB3s has been described previously [25] [26]. Timer-CVB3 was produced as follows: HeLa cells were transfected with 2.8 microgram of the sterile Timer-CVB3 plasmid using Lipofectamine 2000 (Invitrogen) and incubated at 37°C, 5% CO2 for 3 hours. The transfection solution was removed and the wells were washed in 3 mL of 1×DMEM with 10% FBS. Wash was removed and 2 mL of 1× DMEM with 10% FBS was added. The transfected cells were incubated at 37°C, 5% CO2 for 1day. After 1 day incubation, an additional 2 mL of 1× DMEM with 10% FBS was added to each well of transfected cells on the plate. The plate was incubated at 37°C, 5% CO2 for 2 days.
Transfected cells were observed under fluorescence microscope until the cells were observed to first fluoresce green followed by fluorescing red. Supernatants were transferred to a 50 mL conical, screw-cap tube, and 2 mL of 1× DMEM was placed on top of the cells. The 50 mL screw-cap tube was centrifuged at 2000 RPM for 2 minutes. The supernatant was transferred to a new 50 mL conical screw-cap tube and cells were removed by cell-scraping. The cell solution was transferred to the pellet in the first 50 mL conical, screw-cap tube, and the pellet was resuspended in the solution. The cell solution was then lysed by freeze thaw 3 times: 5 minutes of freezing in 95% ethanol with dry-ice, 30 minutes in ice water, and 5 minutes at room temperature with rocking. The freeze-thawed solution was then centrifuged at 2000 RPM for 2 minutes. The supernatant from this tube was transferred to the original supernatant from the first centrifugation. This solution was then vacuum-filtered through a filter with 0.2 µm pore. This virus solution was labeled passage 1 and used to infect HeLa cells in a 150 mL flask.
HeLa cells grown to 80% confluency were washed two times in 10 mL of 1× PBS. Then, 8 mL of Timer-CVB3 (Passage 1) was added to HeLa cells and incubated at 37°C, 5% CO2 for 1 hour, and the flask was rocked every 15 minutes to fully cover the cells. Then, 18 mL of 1× DMEM with 10% FBS was added and the flask was incubated at 37°C, 5% CO2 for an additional 48 hours. Passage 2 supernatant and cells were separated, and cells were scraped into 10 mL of 1× DMEM, lysed by freeze thaw as described for Passage 1. This virus solution was titrated by plaque assay using HeLa cells at 37°C, 5% CO2 for 48 hours and stored at −80°C.
Virus titrations were carried out as described previously [26]. Briefly, HeLa cells were diluted to a concentration of 1×105 cells/mL, and 3 mL of the cell solution was added to each well in 6-well plates. The plate was incubated at 37°C, 5% CO2 overnight. The next day, serial dilutions were made for each viral sample. After the 1 hour infection, each well was filled with 4 mL of 0.6% Agar, 1× DMEM + 2.5% FBS and 1× Penicillin/Streptomycin (P/S). The plate was incubated at 37°C, 5% CO2 for approximately 44 hrs before fixing. After fixing, 1 mL of 0.25% crystal violet was added to each well, and the plate was incubated at room temperature for 1 hour. After staining, plaques were counted.
HeLa RW cultures were added to 6-well plates as described above. After ∼24 hours, cells were mock-infected, infected with eGFP-CVB3 or dsRED-CVB3 at an MOI of 0.1. Alternatively, cells were infected with Timer-CVB3 at an MOI of 0.01 or 0.1. After 12, 24, 36, 48, or 72 hours PI, cells were fixed in 4% para-formaldehyde and washed three times in 1× PBS. Cells were stored in 1× PBS with 1% BSA at 4°C. Cells were analyzed using a BD FACSAria located in the SDSU Flow Cytometry Core Facility.
Neurospheres were isolated and cultured as previously described [12]. Neurospheres were vigorously dissociated and resuspended in complete NPSC culture medium to a concentration of 105 cells/mL in a 24 well plate and infected with Timer-CVB3 at an MOI of 0.1. Cultures were inspected daily using a Zeiss Axio Observer D.1 inverted fluorescent microscope. NPSCs infected with Timer-CVB3 could establish carrier-state infection, as described previously for eGFP-CVB3 [12]. These carrier-state infections were inspected for “fluorescent timer” protein at the indicated time. Parallel infections were also performed on NPSCs continuously grown and passaged in NPSCs media containing 10 ug/mL Poly IC. Alternatively, 105 cells/ml of NPSCs were differentiated in a 4-well chamber slide (1 ml per well) as previously described [12] and in a u-dish (iBidi, Inc., cat#81156). After differentiation for 5 days, differentiated NPSCs were infected with Timer-CVB3 at an MOI of 0.1. Following infection, differentiated cells in the u-dish were monitored daily for the expression of Timer protein. Once the expression of “fluorescent timer” protein was observed, the culture was imaged at 10 minute intervals, holding the field of image, magnification, channel exposure times, and focus consistent. The focus was adjusted occasionally as needed.
NPSCs were plated onto gelatin/fibronectin-coated micro dishes containing differentiation media at a cell concentration of 105 cells/ml as described previously [12] [28]. The cells were differentiated for 5 days and then infected with Timer-CVB3 (moi = 0.1). Alternatively, differentiated NPSCs were infected with dsRed-CVB3 (moi of 0.1) and transduced with a recombinant adenovirus expressing GFP-LC3 (Adeno-GFP-LC3). Following infection, the cells were observed by fluorescent microscopy at magnification of 320× at the indicated time points.
At 3 or 4 days PI, differentiated cultures in the chamber slides were fixed in 4% para-formaldehyde and washed three times in PBS. Fixed cells were permeabilized with 0.25% TritonX-100 in PBS, washed three times, blocked with 10% normal goat serum (NGS) and immunostained using the following antibodies: Nestin (Covance Inc.; Cat# PRB-315C) at 1∶1000, neuronal class III β-tubulin (Covance Inc.; Cat# PRB-435P) at 1∶1000, GFAP (Sigma-Aldrich Co.; Cat# G 9269) at 1∶1000, MBP (Chemicon Inc.; Cat# AB980) at 1∶1000, Viral 3A protein [51] at 1∶100. Primary antibodies were diluted in 2% Normal Goat Serum (NGS) in PBS (150–200 µl per slide) in humidified chamber and incubated overnight. Slides were washed with PBS for 5 min (3×). Secondary antibodies (at 1∶1000) conjugated to Alexa-Fluor-647 (Invitrogen; Cat# A21245), were diluted with 2% NGS in PBS (150–200 µl per slide) and incubated overnight. Following incubation with the secondary antibody, slides were washed with PBS for 5 min (3×), and mounted in Vectashield with DAPI. Three to five representative images of the cultures were taken for each sampling time point at multiple magnifications.
HeLa RW cultures were plated in a 6-well plate at a concentration of 105 cells/ml (3 ml per well), treated with either 10 µg/ml Poly IC, 50 µ/ml IFN-β, or media alone and allowed to adhere to the plate overnight. Timer-CVB3 was then added at moi = 0.01 (14 µl of 106 pfu/mL virus stock) or moi = 0.1 (14 µl of 106 pfu/mL virus stock). Eight images of each culture were taken at 100× total magnification per condition, per day. Samples for viral titrations were taken immediately following imaging. For each image, the number of red, yellow, green, or round colorless cells were counted using the ImageJ Cell Counter plug-in.
EMVs were purified from differentiated C2C12 cells. C2C12 cells were seeded onto T-25 flasks in DMEM supplemented with 10% fetal bovine serum + antibiotics and infected with eGFP-CVB3 (moi = 100; 2.9×108 pfu/mL virus stock concentration). A gradient maker was used establish a continuous 8–20% iodixanol gradient (Opti-Prep; Sigma-Aldrich Co.). 1 ml of day 5 PI supernatant isolated from infected C2C12 cells was used for the continuous 8–20% iodixanol gradient, and the gradient was centrifuged at 141,000 g (28,700 rpm) in an SW.41 Ti rotor for 48 h at 4C in a Beckman L8-60 Ultracentrifuge. Twenty-four fractions were collected from the top of the gradient and the density for each fraction was determined using a Bausch & Lomb refractometer (Bausch & Lomb, Inc.). Fractions were frozen at −70°C prior to virus titration by plaque assay.
EMVs were purified from differentiated C2C12 cells and NPSCs. C2C12 cells were seeded onto 6-well culture plates at a concentration of 105 cells/well in DMEM supplemented with 10% fetal bovine serum + antibiotics and infected with eGFP-CVB3 (moi = 100; 14 µl of 109 pfu/mL virus stock). NPSCs were plated onto gelatin/fibronectin-coated micro dishes at a cell concentration of 105 cells/ml. NPSCs were differentiated for 5 days, and media was changed every two days. NPSCs were infected with eGFP-CVB3 (moi = 0.1; 14 µl of 106 pfu/mL virus stock). On day 3 PI, supernatants from T-75 flasks were isolated and centrifuged at 3000 g (rcf) for 15 min to spin down cell debris and resuspended in 1× DMEM. Supernatants were transferred to a new conical tube, and 2 ml Exoquick-TC (System Biosciences Inc) was added for every 10 ml of supernatant. The mixture was incubated overnight at 4°C, and EMVs were centrifuged at 1500 g (rcf) for 30 min. Purified EMVs were resuspended in 100 µl 1× PBS. 20 µl of this mixture was placed on a slide and visualized by fluorescence microscopy at 320× magnification. Also, HeLa cells were infected with 50 µl of the purified EMVs and observed by fluorescence microscopy at various time points post-incubation. EMV and non-EMV fractions were freeze/thawed three times, and viral titers were determined by standard plaque. Alternatively, scraped cells or purified EMVs were washed with phosphate buffered saline and then disrupted with chilled RIPA buffer, and incubated on ice for 30 min with vortexing every 10 min. Protein concentrations were measured by bicinchoninic acid assay. Three primary antibodies were used for western blotting: rabbit anti-LC3 A/B (Cell Signaling Technologies, Inc, Cat# 4108), rabbit anti-Flotillin-1 (Cell Signaling Technologies, Inc, Cat# 3253), and mouse anti-enteroviral VP1 (Vector Laboratories Cat# VP-E603). 20 µg of protein derived from purified EMVs or cell lysates were utilized for western blot analysis, as described previously [28].
EMVs were purified from differentiated C2C12 mock-infected or infected with Timer-CVB3. Purified EMVs were resuspended in 1 ml 2.5% glutaraldehyde/PBS and incubated on ice for 1.5 hours. Samples were centrifuged at 4000× g for 20 minutes at 4°C. Pellets were washed without resuspension twice with PBS for 10 minutes. Pellets were then resuspended in 250 µl PBS. The resuspended pellets were centrifuged at 4000× g for 20 minutes at 4°C. Supernatants were removed and pellets were resuspended in 1 ml 1% osmium tetroxide in PBS. The resuspended pellets were incubated on ice in the dark for 1 hour. Samples were centrifuged and washed as previously described. Following the removal of 1% osmium tetroxide, 1% uranyl acetate in PBS was added to the pellets without resuspension and incubated on ice in the dark for 1 hour. Pellets were washed with distilled water three times without resuspension for 10 minutes. Following the removal of distilled water, pellets were then dehydrated with a step-wise ethanol gradient (30%, 50%, 70%, 85%, 95%) for 10 minutes per step without resuspension. Pellets were then washed with 100% acetone three times for 10 minutes per wash. Following acetone removal, 33% EPON/66% acetone solution was then added to pellets without resuspension and placed in spin wheel overnight. Following removal of 33% EPON/66% acetone, 66% EPON/33% acetone solution was added to pellets without resuspension and placed in spin wheel overnight. Following 66% EPON/33% acetone removal, 100% EPON was added to pellet (without resuspension) and incubated at 60°C overnight to harden. Samples were cut into <50 nm sections using a diamond knife and a Leica EM UC6 microtome. Sections were positively stained using lead citrate [67] and imaged using a FEI Tecnai 12 transmission electron microscope.
|
10.1371/journal.pntd.0005786 | A cross-sectional seroepidemiological survey of typhoid fever in Fiji | Fiji, an upper-middle income state in the Pacific Ocean, has experienced an increase in confirmed case notifications of enteric fever caused by Salmonella enterica serovar Typhi (S. Typhi). To characterize the epidemiology of typhoid exposure, we conducted a cross-sectional sero-epidemiological survey measuring IgG against the Vi antigen of S. Typhi to estimate the effect of age, ethnicity, and other variables on seroprevalence. Epidemiologically relevant cut-off titres were established using a mixed model analysis of data from recovering culture-confirmed typhoid cases. We enrolled and assayed plasma of 1787 participants for anti-Vi IgG; 1,531 of these were resident in mainland areas that had not been previously vaccinated against S. Typhi (seropositivity 32.3% (95%CI 28.2 to 36.3%)), 256 were resident on Taveuni island, which had been previously vaccinated (seropositivity 71.5% (95%CI 62.1 to 80.9%)). The seroprevalence on the Fijian mainland is one to two orders of magnitude higher than expected from confirmed case surveillance incidence, suggesting substantial subclinical or otherwise unreported typhoid. We found no significant differences in seropositivity prevalences by ethnicity, which is in contrast to disease surveillance data in which the indigenous iTaukei Fijian population are disproportionately affected. Using multivariable logistic regression, seropositivity was associated with increased age (odds ratio 1.3 (95% CI 1.2 to 1.4) per 10 years), the presence of a pit latrine (OR 1.6, 95%CI 1.1 to 2.3) as opposed to a septic tank or piped sewer, and residence in settlements rather than residential housing or villages (OR 1.6, 95% CI 1.0 to 2.7). Increasing seropositivity with age is suggestive of low-level endemic transmission in Fiji. Improved sanitation where pit latrines are used and addressing potential transmission routes in settlements may reduce exposure to S. Typhi. Widespread unreported infection suggests there may be a role for typhoid vaccination in Fiji, in addition to public health management of cases and outbreaks.
| Fiji has experienced a decade-long increase in typhoid fever cases, a potentially life-threatening systemic bacterial disease caused by Salmonella Typhi. We undertook a representative blood-serum community survey to measure antibodies (IgG) against the Vi antigen of Salmonella Typhi using a rigorous survey design. We found one in three residents of mainland, unvaccinated Fiji had detectable antibody against Vi. This was higher than would be expected from confirmed case notifications received by the national surveillance system. Additionally, similar antibody responses were detected in Fijians of all ethnicities, which contrasts to surveillance cases in which indigenous iTaukei Fijians were disproportionately affected. Serology on a Fijian island where a significant proportion of the population has been vaccinated found that three-quarters of residents were seropositive three years after the Vi-polysaccharide typhoid vaccination campaign. Importantly, in mainland participants, seroprevalence increased with age, suggesting long-standing, low-level, endemic transmission. Pit latrines were associated with seropositivity when compared with septic tanks, and settlements compared with residential housing. Very high antibody titres in a small percentage of participants may suggest carriage of Salmonella Typhi. The seroprevalence findings suggest eliminating typhoid from Fiji by focussing on cases and outbreaks alone will be challenging. Our results support typhoid vaccination and further development of water, sanitation and hygiene infrastructure in Fiji.
| Typhoid fever is a systemic disease resulting from infection by the bacterium Salmonella enterica subspecies enterica serovar Typhi (S. Typhi), a human restricted pathogen transmitted from faeces to water and food or by contact and fomites. [1,2] Infection syndromes range from asymptomatic (including carriage) to severe disease with life-threatening complications, including intestinal perforation, encephalopathy, and haemodynamic shock [3,4]. Many typhoid cases present as non-specific acute febrile illnesses that may be difficult to differentiate from other common tropical infectious diseases such as dengue and leptospirosis. There were an estimated 11.9 million (9.9 to 14.7) cases of typhoid in low and middle income countries in 2010, resulting in 129,000 (75,000 to 208,000) deaths [5].
Fiji is an upper middle income country with an estimated population of 892,000 in 2015 [6] across 100 inhabited Pacific Ocean islands, predominantly residing on Viti Levu and Vanua Levu [7]. Administratively, Viti Levu comprises the Western and Central Divisions, the latter containing the capital Suva. Northern Division comprises of Vanua Levu and Taveuni island, whilst Eastern Division comprises of smaller island groups.
Phylogenetic evidence from genome sequencing suggests that S. Typhi has a long history in Fiji though notified blood-culture confirmed cases numbered fewer than 30 annually up to 2004 [8,9]. Annual blood-culture confirmed cases notified through Divisional hospitals to national surveillance have risen [9,10], from 4.4 cases per 100,000 population in 2004 to 45 cases per 100,000 during 2008–2011 [11]. Northern Division saw 121 cases per 100,000 in 2009 vs 28 per 100,000 in the West and 19 per 100,000 in Central [11]. A proportionate increase in clinically diagnosed typhoid was also reported [12].
Notably, >90% of blood-culture confirmed cases are reported amongst indigenous Fijians (iTaukei, approximately 57% of the population), with few cases reported in Fijians of Indian descent (Indo-Fijians, 38% of the population) or Fijians of Asian or European descent[12]. The causes of this disparity are unknown [12]; health seeking behaviours would be expected to lead to higher relative ascertainment in Indo-Fijians than iTaukei [13].
Historically subsistent through agriculture and fishing, more than half the population now reside in urban areas, including informal settlements close to riverbanks and other flood-prone areas with limited access to water and sanitation infrastructure [14].
Typhoid vaccine is not routinely used in Fiji; however, in 2010, following cyclone Tomas, a Vi-polysaccharide (Vi-PS) vaccination campaign was conducted in the highest incidence areas of Fiji. These were predominantly in the Northern Division, with high coverage on Taveuni island and in the Cakaudrove subdivision on Vanua Levu, with targeted geographical vaccination within subdivisions elsewhere [11] reaching 64,000 people, 8% of the national population. Observed disease incidence rates declined in the targeted areas against increasing or stable rates in other areas[11]. Given the ongoing typhoid transmission and the short duration of Vi-PS protection,[15] a 2012 expert meeting was convened by the Ministry of Health, with Australian Aid support, to “develop, prioritise and implement a comprehensive control and prevention strategy” [12]. Knowledge gap analysis identified that a serological survey could inform vaccination policy [12].
Seroepidemiological surveys can help determine population immunity, pathogen exposure and susceptibility, as well as disease- and exposure-related risk factors [16]. Conducted alongside clinical and/or laboratory surveillance, seroepidemiology can help quantify surveillance under- or over-ascertainment, including for enteric diseases [17–20]. Typhoid transmission dynamics are influenced by setting-specific immunity and carriage [21–23]; however, the seroepidemiological methods to attain these are underexploited [17]. This may be in part due to concerns about the sensitivity and specificity of serology for typhoid, which historically has not demonstrated sufficient discriminatory power for individual-level clinical diagnosis [24], (though recent methods may offer promise [25]) as well as concerns about the specificity of assays for carriage detection [26–28] and the existence of multiple immunological pathways to immunity against typhoid fever [29]. Serosurveys utilising assays based on purified, pharmaceutical-grade Vi polysaccharide, the “virulence” factor expressed by S. Typhi [27], for detection of anti-Vi IgG antibody may offer a more reliable approach by avoiding cross-reactivity that arises when Vi antigen preparations contain other bacterial antigens [30,31]. Furthermore, high anti-Vi titres may indicate prolonged immune stimulus from chronic carriage [30,32–34]. We undertook a joint typhoid and leptospirosis seroepidemiological survey [35], and present typhoid findings here.
The study was approved by the Fiji National Research Ethics Review Committee (2013–03) and the London School of Hygiene & Tropical Medicine observational study research ethics committee (6344). All adult participants provided written informed consent. Parents/guardians provided written informed consent on behalf of all child participants (under the age of 18 years old). Written assent was also provided by children aged 12 years and above.
To characterize the immunoepidemiology of typhoid infection in Fiji, with the aim of informing effective and efficient control measures, we surveyed three groups of people: group 1) unvaccinated areas of the main Fijian islands; 2) residents of Taveuni island, where a high-coverage vaccination campaign was done in 2010 [11]; and group 3) a cohort of convalescing Fijian culture-confirmed typhoid cases, to enable estimation of a threshold for seropositivity.
A sample size of 1,600 was proposed for group 1, giving for 7% seroprevalence confidence intervals (CI) for age band groups of 200, if seroprevalence was 40%, at alpha = 0.05. If age bands had design effect of two due to non-independence within clusters, CI would be sufficiently precise at 10%. Expected seroprevalence levels were informed by prior work on Taveuni (Eric Nilles, data on file).
Enzyme-linked immunosorbent assays (ELISAs) to detect S. Typhi Vi-polysaccharide antigen-specific IgG in human serum samples were performed as described previously [53] (provided by Sclavo Behring Vaccines Institute for Global Health, Siena, Italy). Briefly, ELISA plates were coated overnight with 1μg/ml of Vi polysaccharide antigen. Coated plates were washed and blocked with 5% fat-free milk solution. Following blocking, plates were washed and incubated with serum diluted at 1:200 at room temperature (RT) for 2 hours. Plates were washed and incubated with secondary antibody, alkaline phosphatase-conjugated anti-human IgG at RT for one hour. Finally, p-Nitrophenyl phosphate (SigmaFAST N1891, Sigma-Aldrich, United Kingdom) substrate was added for 30 minutes at RT and absorbance was read at dual wavelengths (405 nm and 490 nm) using an automated microplate reader (Biorad). Optical densities (OD) from blank control wells were subtracted from all sample absorbance values prior to estimation of serum titers from a standard curve. We selected 96 random Fijian plasma samples and subjected them to the anti-Vi ELISA. Twenty samples of these samples (with an OD >2.5) were pooled (standard plasma) and used to generate a standard curve. One ELISA Unit (EU) was defined as the reciprocal of the standard dilution (made by 10 2-fold dilutions of the standard plasma) that gave an absorbance value equal to 1 in this assay. All samples were tested at the Oxford University Clinical Research Unit in Ho Chi Minh City, Vietnam.
A surveillance seropositivity threshold was determined using a mixed effects model of serial titres in the convalescent cases group. Models were fitted by maximum likelihood estimation (ML), using a random-intercept fixed-slope above a threshold value and random intercept time-invariant model below, signifying antibody returning to a baseline level. Data from two patients with titres at the upper limit of detection (25,000 EU) were excluded leaving 70 titres from 28 patients. Thresholds were compared using Akaike’s information criterion (AIC).
Data were entered in EpiData 3.1 [54] and analysed in R 3.3 [55]. Seroprevalences were calculated using intra-cluster correlation coefficients (ICCs) and design effects determined on log titres with clustering at the primary sample unit. Putative risk factors for seropositivity were estimated with Huber-White robust standard errors, clustered on the same, using the “rms” package[56]. A multivariable model was developed from univariable risk factors with p-values of less than 0.25, after-regrouping sparse cells for numerical stability, using a backward stepwise approach fitted by AIC, with deletion of observations with missing data. Potential collinearity was assessed by linear-adjusted generalized variance inflation factors (GVIF) in the “CAR” package [57,58], and variables were removed if GVIF^(1/(2*Df)) was over 2 and not considered epidemiologically important to retain. Data were considered insufficient to examine possible risk factors associated with titres that may indicate typhoid carriage. For comparison to age-based seroprevalence, typhoid fever cumulative incidence expected across the life-course was estimated with binomial confidence intervals from confirmed cases notified in 2008–2014 to the national surveillance system at the Fiji Centre for Communicable Disease Control, Mataika House, Suva.
Group 1: Sixty-four mainland clusters in unvaccinated areas of Viti Levu and Vanua Levu were visited for this sero-survey (Fig 1). Of 1,565 people approached, five declined and 1,560 were enrolled. There were no exclusions on medical grounds. A serum IgG titre against Vi polysaccharide (anti-Vi IgG) could not be attained for 29 participants (median age 23, IQR 6 to 34; 19/29 female; 25/29 iTaukei) but was determined in 1,531 individuals (98%; age range 1 to 85 years, median 29, IQR 16 to 48; 820/1,530 (54%) female; 1,164/1,530 76%) iTaukei; see Table 1; non-responses excluded).
For Group 2, on Taveuni Island, the location for a vaccination campaign, 277 people were approached and 256 participants (127 (49.6%) female) in 11 clusters enrolled, with nil excluded. All provided samples that were successfully assayed for anti-Vi IgG (Table 2 and Fig 1).
Group 3: Thirty-seven patients with recent blood-culture confirmed typhoid provided one or more samples that were assayed for anti-Vi IgG (19 (51.4%) female, median age 30, IQR 14 to 42) (Table 3); 30 provided two or more blood samples; and 19 provided three samples. Median duration from reported fever onset to first sample collection was 187 days (IQR 132 to 272 days).
Mixed model seropositivity threshold estimation in Group 3 (convalescent typhoid cases) exhibited best fit at 64 EU (S1 Fig and S1 Table). The ICC and design effect per Group 1 mainland cluster were 0.09 and 3.03, respectively. Across the five-year age bands (S2 Table), the Group 1 mean ICC and design effect were 0.13 and 1.09, respectively.
At the 64 EU threshold, 32.3% of Group 1 mainland participants (95%CI 28.2 to 36.3%) were seropositive for anti-Vi IgG (S3 Table), compared to 71.5% (95% CI 62.1% to 80.9%) of Group 2 (Taveuni island). For sensitivity analysis, at a threshold of 100 EU, 17·7% of the Group 1 mainland participants (95%CI 14·4 to 21.0%) were seropositive, and 58.6% (95% CI 48.4% to 68.8%) of Group 2 (Taveuni island) (S3 Table). To determine a rough estimate of carriage prevalence within Group 1 (mainland), we examined those with the highest anti-Vi IgG titres; 2.8% (1·4 to 4·2%) of the sampled population had an antibody titre of 500 EU or above and 1.4% (0.4% to 2·4%) of the sampled population had an antibody titre of 1,000 EU or above (S3 Table).
The anti-Vi IgG titre distributions are shown in Fig 2 for each surveyed group. The distribution of antibody titres in group 2 (Taveuni island) was shifted rightward relative to the Group 1 (mainland) titres, as would be predicted with the mass administration of Vi-PS vaccine on Taveuni. Thirty-nine (15%, 11.2 to 20.4%) of the 256 Group 2 (Taveuni) participants had Vi IgG titres at the upper limits of detection. In group 3, the convalescent typhoid cases, titres were bimodal, with the higher peak above the modal titre for the mainland group.
Age trends for unvaccinated iTaukei and non-iTaukei ethnic groups each showed increasing seroprevalence with age at 64 EU threshold (group 1 mainland; Fig 3). These increased from approximately 20% in the youngest age bands to 50% in the oldest. In sensitivity analysis at 100 EU threshold, age and ethnicity trends were comparable, with seroprevalence rising from <10% in younger groups to approximately 30% in the oldest age brackets. Notably, for each ethnic group, seroprevalence by age band was substantially higher than the equivalent cumulative incidence that would arise if considering only confirmed cases, more than ten-fold in iTaukei Fijians and several hundred-fold in non-iTaukei Fijians.
Multivariable analysis of group 1 (mainland) found several factors were significantly associated with seropositivity at a 64 EU anti-Vi IgG threshold after adjusting for potential confounders (Table 4). Western Division residents had an adjusted odds ratio (OR) of 0.6 (95%CI 0.4 to 0.8) for seropositivity in comparison to the Central Division. Age association with seropositivity was retained, with an adjusted OR of 1.3 (95% CI 1.2 to 1.4) for every ten-year increase. Individuals with pit sewage systems had an adjusted OR of 1.6 (95% 1.1 to 2.3, p = 0.01) for seropositivity in comparison to participants with septic tanks. Residence in a settlement rather than residential housing had an adjusted OR 1.6 (95% CI 1.0 to 2.7) for seropositivity.
After adjustment, no significant association with seropositivity was observed at p<0.05 for ethnicity, community type, rural residence, self-reported typhoid vaccination, or self-reported diagnosis of typhoid fever. “Home toilet type” was excluded from consideration for multivariable analysis: pour-flush (water seal) toilets were found to be associated with seropositivity on univariable analysis, however these are installed in response to disease outbreaks and so are confounded by indication (Fiji National Taskforce on Control of Outbreak-Prone Diseases, personal communication 2015). Other candidate risk factors identified on univariable screening (S4 Table) were not retained in the final model, including sex, drinking water sources, kava consumption, bathing or washing in rivers and typhoid cases within the household or social network.
This seroepidemiological survey found seroprevalence of IgG against the Vi antigen of S. Typhi of 32.3% (95%CI 28.2 to 36.3%) in mainland Fiji: one to two orders of magnitude higher than would be predicted from case notifications. Seroprevalence increased with age, suggesting established endemic transmission. Both iTaukei and non-iTaukei ethnic groups exhibit similar seroprevalences across age groups, in contrast to notified disease, disproportionately reported from iTaukei Fijians. Some very high titres suggest that carriage occurs. The higher anti-Vi seroprevalence from Vi-PS vaccinated settings also informs the use of anti-Vi IgG as a surveillance marker in unvaccinated populations.
Our population-representative survey design strengthens external validity of seroprevalence estimates over convenience sampling, particularly for age-based inference, as children are rarely blood donors or inpatients. A limitation is the use of a single antigen due to resource availability, mitigated by mixed model determination of a surveillance titre threshold.
The proportion non-iTaukei (24%, specifically Indo-Fijian (22%)), was lower than expected from census data [13]. Potentially a greater proportion of Indo-Fijians reside in larger communities within nursing zones than documented in the sampling frame, or in vaccinated areas; higher emigration and lower fertility rates may also contribute [13]. Post-stratification weighting was not considered appropriate given demographic trend uncertainties since the 2007 census and sparse sub-nursing zone population records. Representativeness was addressed through survey design, and clustering through design-effects (which were modest) and cluster-robust regression. The slight excess of females in the survey may be due to different residency patterns, such as male overnight residency in agricultural shelters.
Vi IgG titres can be compared across studies, despite incomplete international assay standardisation [59]. The Fiji results contrast to two Vi ELISA serosurveys from Kathmandu, Nepal. The first found rising serum bactericidal activity with age, suggesting a similar acquisition of exposure with age, but found no age trend in anti-Vi IgG. [60] The second, using an assay similar to that applied in Fiji, reported high anti-Vi IgG in all age groups, suggesting hyperendemicity [61]. In Cape Town, South Africa, where typhoid was considered endemic, 40% of sampled unvaccinated 9 year olds were found to have anti-Vi IgG titres believed to be protective [62]. In contrast, we found mean seroprevalence in Fijian 5 to 9 and 10 to 14 year olds was not more than 20% (Fig 3) suggesting a lower force of infection, if antibody thresholds are comparable. This would also be consistent with lower confirmed case incidence [12,62].
Incomplete seroconversion (as seen in other settings[63]) and waning anti-Vi IgG titres observed amongst convalescent cases in our study suggest that multiple infections, symptomatic or otherwise, may be required for the establishment of sustained immunity to typhoid fever and the corresponding seroprevalence patterns observed in the Fijian mainland group. This concurs with prevailing conceptualisation[64,65], and experimental study [66] as well as recent models used to estimate vaccine impact [21,67] and elucidate transmission determinants in Malawi [23]. Papua New Guinea saw a similar upturn in typhoid notifications in the 1980-90s, also from a low-level, sporadic baseline, [68] with rise in population O antigen also observed [69], suggesting an overall increase in transmission; longitudinal seroepidemiology may likewise be informative in Fiji.
Ingestion of a large dose of S. Typhi can overwhelm naturally-acquired (and vaccine-derived) immunity [64,66,70,71], and so age- and ethnically-differential exposure to high and low dose inocula is one mechanism by which divergence between serological and confirmed-case data may be explained, if for example, iTaukei adolescents and young adults ingest larger inocula through exposure to particular foods. Such patterns might also be compatible with genetic differences in typhoid susceptibility, potentially mediated by HLA-type [72], with reduced susceptibility in Fijians of Indian descent, whose South-Asian ancestors may have experienced many millennia longer exposure to S. Typhi than iTaukei Fijians’ forebears [73–75].
Our multivariable analysis suggest that settlement residency and pit latrines use may be risk factors for S. Typhi infection. Whilst infrastructure upgrades may have multiple public health benefits, typhoid prevention should be planned with consideration for findings emerging from case-control and environmental health research [50,76]. Widespread subclinical infection, both transient and chronic, as suggested by this serosurvey, suggests that whilst systematic public health management of cases and outbreaks and early diagnosis and treatment of patients remain of vital importance to reduce morbidity and mortality from typhoid fever in Fiji, a focus on these alone may be insufficient to eliminate transmission. Alongside continued socio-economic development and expanded access to infrastructure for sanitation, water supplies and handwashing with soap, programmatic vaccination may be amongst interventions necessary to bring about effective typhoid control in Fiji.
|
10.1371/journal.pcbi.1004170 | Detailed Contact Data and the Dissemination of Staphylococcus aureus in Hospitals | Close proximity interactions (CPIs) measured by wireless electronic devices are increasingly used in epidemiological models. However, no evidence supports that electronically collected CPIs inform on the contacts leading to transmission. Here, we analyzed Staphylococcus aureus carriage and CPIs recorded simultaneously in a long-term care facility for 4 months in 329 patients and 261 healthcare workers to test this hypothesis. In the broad diversity of isolated S. aureus strains, 173 transmission events were observed between participants. The joint analysis of carriage and CPIs showed that CPI paths linking incident cases to other individuals carrying the same strain (i.e. possible infectors) had fewer intermediaries than predicted by chance (P < 0.001), a feature that simulations showed to be the signature of transmission along CPIs. Additional analyses revealed a higher dissemination risk between patients via healthcare workers than via other patients. In conclusion, S. aureus transmission was consistent with contacts defined by electronically collected CPIs, illustrating their potential as a tool to control hospital-acquired infections and help direct surveillance.
| Recent advances in communication technologies allow monitoring high-resolution contact networks. Close proximity interactions (CPIs) measured by wireless sensors are increasingly used to inform contact networks for the dissemination of pathogens in computational models, although empirical justification is lacking. Here, we conducted a longitudinal prospective study for four months in a hospital, including both patients and healthcare workers (HCWs). High-resolution CPIs were recorded continuously, and participants undertook weekly nasal swabs to detect S. aureus carriage. We set out to test whether the contact network measured by CPIs supported observed transmission episodes. A simulation study was first conducted to choose a test statistic for the association of CPI paths with transmission, showing that CPI path length from transmitter to incident case was the most powerful. Then, we selected patients presenting incident S. aureus colonization in the data. We showed that CPI paths existed to carriers of the same strain, with path lengths significantly shorter than between random pairs of participants, in agreement with the transmission hypothesis. In-hospital contact networks measured by CPIs inform on opportunities for pathogen transmission. These could be used in surveillance systems to help prevent the spread of nosocomial pathogens.
| Chains of transmission in communicable diseases are often identified by ad hoc strategies, combining retrospective information on locations attended and pathogen genetics to identify time-consistent transmission paths.[1] In contrast with such undertakings, “digital epidemiology” propose to use new technologies to prospectively measure contacts and understand transmission[2,3]. Close-proximity interactions (CPIs) between persons recorded by wireless sensors[4] in real-life settings like schools or hospitals[5,6] have been used as indicators of contact in this respect. However, there is no evidence yet that such CPIs actually capture contacts explaining transmission.
To test this hypothesis, we designed a study where both Staphylococcus aureus carriage and CPIs were measured in a 200-bed long-term care facility with 5 wards. This setting has several advantages for our purpose: S. aureus is commonly found in healthcare facilities, colonizing patients and healthcare workers (HCWs); S. aureus carriage in the nares is usually prolonged as the nares are the most consistent area from which it can be isolated,[7] allowing its detection by repeated routine screenings; identical genetic and antibiotic-resistance profiles show that S. aureus strains spread among patients and HCWs[8,9]; long-term care facilities harbor a stable population, with patients staying for extended periods under the care of dedicated staff. The control of S. aureus transmission is also relevant for hospital hygiene, because its carriage increases the risk of healthcare-associated infections.[10]
In our study, S. aureus carriage was identified in patients and, importantly, in HCWs every week by repeated nasal swabbing. During the same period, all participants wore small wireless sensors that recorded their CPIs with each other in real time (every 30 s). To make the best use of this new data and account for the difference in temporal granularity, we first assessed the ability of several statistics to test the correlation of CPI records and S. aureus carriage. These characteristics are first presented based on the analysis of simulations where a pathogen spread according to the CPI network edges, then applied to the original data.
Authorizations were obtained in accordance with French regulations regarding medical research and information processing. All French IRB-equivalent agencies accorded the i-Bird program official approval (CPP 08061; Afssaps 2008-A01284-51; CCTIRS 08.533; CNIL AT/YPA/SV/SN/GDP/AR091118 N°909036). Signed consent by patients and staff was not required according to the French Ethics Committee to which the project was submitted.
The I-Bird (individual-based investigation of resistance dissemination) study was conducted in a 200-bed long-term and rehabilitation hospital in northern France. The hospital is organized in 5 wards corresponding to medical specialties (geriatrics, neurology, nutrition, orthopedics, post-operative care). During the study period, 329 distinct patients stayed in the facility. Hospital staff (HCWs and other administrative personnel) numbered 261. In the text, all hospital staff is referred to as healthcare workers (HCWs). More details about the I-Bird investigation are provided in S1 Text. When relevant, patients, nurses, nurses’ aides and physicians were analyzed according to the ward in which they stayed or worked; night-shift staff, reeducation therapists and administrative personnel were excluded from these analyses as they were not assigned to a particular ward.
During the study period, all individuals (patients and HCWs) wore a small wireless sensor that recorded, every 30 s, the identity of other sensors that were in close proximity (typically < 1.5m, front-facing). The deployment of such sensors did not rely on any stationary infrastructure to record CPIs, as each sensor directly stored timestamped CPIs on its on-board flash memory. Further details on CPI collection and network reconstruction, along with descriptive characteristics of contact patterns, are available in S2 Text. In the following, analyses are conducted on a dynamic CPI network aggregated at a daily scale by pair of individuals (ie network edges). We defined an individual’s k-hop neighborhood as all other individuals in the network who were found within k steps from him. For example, the 1-hop neighborhood of an individual contained all his direct neighbors, while his 2-hop neighborhood contained both his direct neighbors and these neighbors’ neighbors.
All participants underwent weekly nasal swabs to monitor S. aureus carriage. Upon detection of S. aureus colonization, the isolated strains were spa-typed[11,12] and their resistance profiles to 20 antibiotics were determined (see also S1 Text for detailed protocol). The screening procedure had an expected sensitivity of 61.5% and specificity of 98.8%,[13] although higher sensitivity value has been reported (∼80%).[14] The anterior nares were preferred to other body areas because they harbor the most stable S. aureus colonization and also reflect on overall body carriage.[7,15] Furthermore, eradication of nasal carriage is also associated with eradication of skin carriage.[16,17] S. aureus strains were considered identical when they had the same spa type and antibiotic-resistance profile, in accordance with studies comparing spa typing to other molecular techniques.[11,12,18,19] Transmission events were identified by the isolation of a new S. aureus strain from a patient’s swabs, defining “incident colonization episodes”. Because HCWs may be transiently colonized,[8,20,21] which would mostly be missed with weekly swabs, we only considered incidence in patients. To account for imperfect S. aureus detection in case of multiple carriage, we also required that the new S. aureus strain had not been detected in the patient’s previous 2 swabs had he been colonized with another strain in the preceding week.
Each incident colonization episode was investigated to identify time-consistent CPI paths linking the incident case to a previous carrier of the same strain. Recent CPI paths were favored over others by applying the following algorithm:
- All individuals carrying the same strain in the three preceding weeks were defined as “candidate transmitters”, regardless of their CPI connections.
- CPI paths to all candidate transmitters were looked for. In case of existence, the candidate transmitter became a “CPI-supported candidate transmitter”. In this case, the CPI path length (in hops) was computed.
- We sorted all CPI-supported candidate transmitters according to distance in time, then distance in hops. The first CPI-supported candidate transmitter in this list was the “CPI supported transmitter”. In other words, it was the least remote in number of hops among all candidate transmitters arising the least remote in time. In case of ex aequo, one of the candidate transmitters was chosen at random if required.
We investigated 3 weeks before incidence as it allowed finding a CPI-supported transmitter for all incident episodes but 4, which were not CPI-supported by exploring further back in time.
Testing for CPI supported transmission: test statistics & simulations. We identified three observable quantities that would provide evidence for the correlation between CPIs and observed transmission: S1—The proportion of incident colonization episodes with least one CPI-supported transmitter; S2—The proportion of CPI-supported transmitters in direct CPI with an incident case; and S3—the length of the shortest CPI-supported transmission path (a good proxy to the actual transmission path[22]). Each of these characteristics can be used to build a test, where, classically, observations would be compared to the expected values under the null hypothesis of independence between CPIs and transmission. These expected values can be computed by a Monte Carlo approach: carriage information was first randomly permuted between participants. To keep autocorrelation between successive swabs in the same individuals, we permuted carriage information over the 3 preceding weeks simultaneously. As S. aureus prevalence was similar between patients and HCWs (Table 1), we did not take occupation into account for permutations. For each incident colonization episode, 100 replicates of permuted carriage statuses were generated to simulate the distribution of statistics of interest for investigated strategies.
For S1 and S2, we compared the observed percentages of CPI supported paths to that expected after random permutations of carriage (e.g. in S1, the observed percentage of CPI-supported episodes was compared to the average proportion of CPI-supported episodes among all permutations). For S3, the shortest CPI path length was averaged across all replicates for each incident colonization episodes, thus providing the expected distribution under the null. The observed CPI path length distribution was then compared to that expected under the null using the Wilcoxon signed rank paired test.
To first study the characteristics of the three approaches and choose the most powerful, we used a simulation study based on a Susceptible—Colonized—Susceptible transmission model on the CPI network. Stochastic simulations were performed to mimic observations of incident colonization episodes in our study. First, the dynamic CPI network between all participants in a random 3-week long period was selected. All individuals in this network were assumed initially noncolonized (i.e. susceptible), except one randomly chosen to be the initial carrier in the first week. For each day d in the following 3 weeks, a noncolonized individual (say i) could become colonized with probability
Pi(d)=1−[(1−PPA)nPA(i,d−1)*(1−PHCW)nHCW(i,d−1)]
where nPA(i, d-1) is the number of carrier-patient neighbors on day d-1 and nHCW(i, d-1) the number of carrier-HCW neighbors of i, PPA the probability of transmission per contact with a carrier patient and PHCW with a HCW. Colonized individuals cleared colonization at a constant rate qPA = 0.1 days−1 for patients and a gHCW = 0.45 days−1 for HCWs in agreement with other studies.[23,24] The observation model closely imitated those of our investigation: from the wholly simulated transmission chain, we only selected data determining carriage once a week in each participant. As in practice, the status of participants in the same ward was determined the same day. Finally, an incident colonization episode was selected from new carriers in week 3 of the simulation, as in the original data. Simulations were run to produce observations of a variable number of incident cases. The power of each statistical approach, S1 to S3, was determined as a function of the number of incident episodes.
CPIs were recorded among 590 individuals (329 patients and 261 HCWs) during the 4-month period (Table 1), and yielded 85,025 daily CPIs. Each day, a CPI network was defined with study participants as nodes and CPIs as edges (Fig. 1). The collection of daily CPI networks defined the “dynamic CPI network”. While the numbers of CPIs were within the same range for patients and HCWs, the daily-cumulative durations of the CPIs were much longer for patients than HCWs, respectively: 12.2 (± 11.3) h (mean ± SD) vs. 3.7 (± 2.4) h (Fig. 2). Further description of the dynamic CPI network is provided in S2 Text.
During the same time, 4,175 swabs were collected: 37.2% of them (1550 swabs from 363 participants) were positive for S. aureus carriage. In all, 148 different spa types were isolated during the study (Fig. 1, and Fig. 1 in S1 Text for incidence per week and spa types). Notably, 114 strains were isolated more than once, each in three participants on average, suggesting that transmission had occurred among these individuals.
We assessed the power of all proposed strategies with increasing numbers of incident colonization episodes (10, 20, 30, 50, 100 and 153). For each of these amounts, 500 replicates of 100 permutations were performed. Each strategy was performed on every replicate. Table 2 shows the power of the tests to reject the null hypothesis of independence between CPIs and transmission. Strategy S1, based on the existence of CPI supported transmitters, yielded very poor results. Indeed, in almost all situations, a CPI-supported transmitter existed in the original data as well as in the permuted data, so that no difference from random was seen in this characteristic. The percentage of incident cases in direct CPI with CPI-supported transmitters and the shortest CPI-supported path length yielded more useful procedures. The length of the shortest CPI path from transmitter to incident case (S3) was slightly more powerful than the percentage of transmitters in direct CPI with the incident case (S2), although both approaches had large power for rejecting the null for samples of size 153.
The time to first S. aureus colonization was analyzed for 201 patients who were not colonized at admission: 73 experienced incident colonization. The cumulative incidence of S. aureus colonization was 33% (95% confidence interval (C.I.) [25–41%]) 1 month after admission, with almost equal incidence of methicillin-resistant S. aureus (MRSA) (23.2% (95% C.I. [15.1–30.6%])) and methicillin-sensitive S. aureus (MSSA) (16.5% (95% C.I. [9.3–23.2%])). The risk of colonization did not change with the number of distinct direct neighbors during the preceding week (ie weekly degree; hazard ratio (HR) = 1.05 (95% C.I. [0.95–1.21]) for a 5-neighbor increase, P = 0.4), using either the raw number of CPIs (ie the sum of daily degrees; HR = 1 (95% C.I. [0.95–1.10]) for a 5-CPI increase, P = 0.4) or the cumulative duration of CPIs (ie the weight of network edges; HR = 1 (95% C.I. [0.99–1.01]) for a 1-h increase, P = 0.6). The same conclusions were drawn for MSSA or MRSA colonizations.
Overall, 237 incident-colonization episodes were documented in 111 patients (144 MRSA, 93 MSSA). For each incident episode, we identified “candidate transmitters”, i.e. people who had carried the same strain at any time in the preceding three weeks. Among the 237 incident-colonization episodes, 173 (73%) had 307 candidate transmitters, with no difference between MRSA and MSSA episodes (76% (110/144) vs. 68% (63/93), P = 0.16). Episodes without a candidate transmitter did not occur earlier post-admission than others (8.1 vs. 8.3 weeks, P = 0.8), or preferentially in some wards (P = 0.13). Twenty of the 173 episodes with candidate transmitters were discarded as the incident case had a missing CPI record due to sensor failure in the preceding weeks.
We investigated all 153 incident colonization episodes detected from longitudinal swab data. As expected from the simulations, a CPI path existed between the candidate transmitter and the incident case in almost all instances (97%, Table 3). The characteristics of the shortest CPI paths lengths from candidate transmitter to incident case were in favor of transmission along CPIs, with shorter paths in the original data than after random permutations (Fig. 3, Strategy S3: P < 0.001). This was therefore the sign that S. aureus transmissions detected from the I-Bird swabs were driven by CPIs.
A direct CPI contact existed between the candidate transmitter and incident case (i.e. a CPI path of length 1) in 48% of the cases vs. only 30% expected by chance (Strategy S2: P < 0.0001). A CPI path of length 2 was observed in 38% (vs. 53%) of the episodes and of length larger than 2 in the rest of the cases.
In most cases (64%), a CPI supported transmitter was found in the preceding week. The remaining were found in the preceding two (23%) or three weeks (13%).
We next investigated transmission differences according to occupation, focusing on dissemination events for which the path from candidate transmitter to incident case had exactly one (noncolonized) intermediary. For these 2-hop CPI paths, S. aureus spread was more frequently observed when the intermediary was a HCW than a patient (2.1% vs. 1.8%, P = 0.0004). The relative risk (RR) of transmission by a HCW was therefore 1.2 (95% C.I. [1.1–1.3]). This increased risk was more pronounced when the initial carrier was a patient (1.9% vs. 1.5%, P < 0.0001; RR = 1.3 (95% C.I. [1.1–1.4])) rather than an HCW (3.2% vs. 3.0%, P = 0.48; RR = 1.1 (95% C.I. [0.9–1.2])).
All analyses were repeated using a thinned dynamic CPI network obtained by excluding short individual interactions lasting < 5 min prior to daily aggregation. The thinned CPI network was still dense, including 8.1% of all potential interactions. As expected, this decreased density increased the number of intermediaries between 2 persons (mean = 11 ± SD = 6) compared to the full network and decreased the number (mean = 4 ± SD = 1) and duration of daily CPIs. The distribution of within/outside-ward CPIs was almost the same as before, with 75% of CPIs occurring within the ward. Again, no risk factor that could increase the risk of colonization was identified.
The CPI support of transmission was even larger than before, with shorter path lengths in the original network than expected by chance (P < 0.0001). In the thinned network, 64% of candidate transmitters were in the incident case’s 2-hop neighborhood (i.e. direct or with one intermediary), compared to 46% with the permuted carriage data.
Keeping all candidate transmitters with a path to the incident case, rather than only the closest ones did not change the conclusions. Excluding repeated incident episodes of the same strain in a given patient, leading to consider 129 transmission episodes only, did not affect the conclusions drawn regarding CPI support. Finally, keeping only episodes that were CPI-supported in the first week before incidence led to the same conclusion (96 CPI-supported episodes, CPI path length significantly shorter (P = 0.01)).
To account for imperfect sensitivity and possible false negatives in swabs, we discarded all incident episodes where the carriage had been positive/negative/positive with the same strain at both ends, as those may be false negatives. This led to retain 129 incident colonization episodes with candidate transmitters and CPI records, among which 126 were CPI supported (ie 98%). The shortest CPI path length was significantly shorter than chance predicted (P < 0.0001).
The contribution of the contact network between patients and HCWs in explaining hospital-associated infections is widely accepted[25,26], although it has never been tested empirically. Here, using electronic wireless devices to record close proximity interactions among persons in a hospital, we find evidence that these interactions are indeed informative for S. aureus transmission. To date, such high-resolution contact networks were used to structure contacts in computerized models and study how characteristics of individuals[27,28] and network topology[2,29] could influence the course of outbreaks. However, these findings relied on the underlying hypothesis that CPI paths actually captured dissemination paths. Our results provide evidence that CPIs recorded by electronic sensors are indeed relevant to explain transmission. This validates using such networks in future epidemiological studies, and should provide a powerful tool to better characterize risk and plan control measures for pathogens transmission in specific settings.
Recording interactions among individuals is increasingly easy using remote sensors[30]. In the study facility, wearing sensors was well accepted by patients and HCWs. Because the recording was limited (typically < 1.5 m), these signals may be a good proxy for real-life contacts between people. Although contact surveys were shown to document contacts hardly overlapping with those recorded by electronic sensors,[31] direct observations by dedicated investigators provided more congruent data.[32] Individual surveys tend to omit contacts of short duration,[31] and therefore electronic sensors might capture more complete data regarding contact patterns. In our study, although most CPIs occurred within the wards (leading to clustered communities, as seen in Fig. 1), the CPI network was rather dense, covering up to 20% of all possible interactions among participants. The shortest path between any two individuals had few intermediaries, a typical “small-world” feature[33]. HCWs spent approximately 20% of their work shifts in direct contact with patients (1h50 out of 8h), in the same range as that reported in an emergency unit (∼30%)[34] and the cumulated CPIs duration in a HCW compared with that reported for another ICU (∼2.2 h)[32]. However, in sharp contrast with an earlier study conducted in a pediatric setting[30], where almost no contacts existed between patients, CPIs between patients herein were frequent and prolonged, as expected in a long-term care and rehabilitation facility where patients can initiate social interaction with others more easily than in acute-care hospital. Finally, ward organization was important in structuring contacts between patients and HCWs. This contact-clustering suggests that some interaction patterns between patients and HCWs could be rather constant from one hospital to the next, e.g., the numbers and durations of interactions, but that the full contact network may depend on type of care and ward organization. Other features of interest were the quick encounter of most of direct contacts during the first week after admission and that an incoming patient’s 3-hop neighborhood quickly encompassed almost all individuals in the hospital. This small-world feature may profoundly affect the potential spread of pathogens, increasing the size and speed of outbreaks[33,35].
S. aureus carriage was common among patients and HCWs, with a significant percentage of patients already colonized at admission (33.7% for S. aureus; 17% for MSSA; 18.9% for MRSA). Approximately one-third of noncolonized, newly admitted patients became colonized with S. aureus within the first month post-admission, as frequently observed in long-term care facilities[36,37]. The cumulative MRSA and MSSA incidence rate were similar after 1 month, as were their mean carriage prevalence, suggesting little, if any, difference in transmission between resistant and non-resistant strains. Patients admitted to long-term care facilities come from other hospitals and prevalence of carriage at admission is large. We analyzed incident carriage episodes only to focus on transmission occurring within the long-term care facility. The numbers and durations of contacts, although pointing towards increased risk for participants with more and longer contacts, were not by themselves strongly associated with S. aureus colonization: additional information on whether the contacts were carriers is likely to be required in this respect. Finally, because only weekly nasal swabs were conducted, neither hand carriage nor transient colonization episodes (<1 week) were considered, although hand carriage by HCWs has been described[8]. In our analysis, this may lead to imperfect observation of the carriage status, as participants may have cleared colonization between successive swabs.
In our study, S. aureus carriage was determined by nasal swabs. As previously stated, skin colonization was not detected. Yet, previous studies have shown that while S. aureus can be isolated from many anatomic location, nasal swabs were consistent with carriage isolated from other area of the body in 82% of the case[38]. In HCWs, where transient colonization is more likely to occur, this might lead to underestimate prevalence and therefore overestimate incidence, as an (unobserved) skin carriage could lead to a longer, more stable, colonization of the nares. For this reason, we chose not to include incident colonization episodes occurring in HCWs. Although the main route of transmission for S. aureus remains physical contact with carrier individuals, transmission through the environment is also possible, for example in the form of fomites.[26] Our procedure cannot distinguish between routes of transmission leading to short CPI paths, for example, if contact with fomites occurred only when people were at CPI range of a known carrier. Yet, our results suggest that CPIs, as defined in our study setring, correlated with S. aureus transmission routes and are therefore a good proxy[35] for interactions leading to S. aureus dissemination.
The choice of a test statistic to test for transmission along CPIs required giving some consideration to the setup of this study. First, the density of contacts between participants was large, as in occupational networks measured in health care structures.[24,30] This leads to percolation,[39] with typical short distance between participants. Therefore, a CPI path between a S. aureus carrier and an incident case was not specific enough of transmission: it was the rule rather than the exception, explaining why strategy S1 could not evidence association. Second, while CPIs could be recorded continuously, it is obvious that S. aureus carriage cannot be determined as frequently. Some participants could therefore clear carriage between two successive swabs, hiding their role in transmission. In this case, direct CPI would not be seen in the recorded CPIs. However, a short CPI path may be found to a more distant carrier through non-colonized intermediaries and would still be supportive of transmission. The shortest CPI path between CPI-supported transmitters and incident cases allowed to account for carriage gaps in observed transmission chains and was more informative than mere existence of a connecting path. The power comparison between strategies S1, S2 and S3 showed that this was indeed the case. It also evidenced that the proposed tests actually discriminated between transmission along the CPI paths and random transmission with no relationship to the CPI network.
In contrast to weekly swabbing, wireless sensors recorded interactions permanently and made it unlikely that the network of interactions was imperfectly observed. S. aureus transmission is thought to occur mainly through physical contacts and these should be present in the CPI recordings. However, the CPI network may capture additional interactions that are unlikely to lead to transmission. To focus on interactions that were the most likely to lead to transmission, for example nursing care, we discarded all CPIs lasting less than 5 min. In this thinned network, we found again evidence that paths defined by CPIs supported transmission. Finally, the increased likelihood of S. aureus spread through a—seemingly noncolonized—HCW intermediary was in good agreement with their importance in transmission and the occurrence of transient colonization among them.
The hypothesis that CPIs are a good proxy for contacts leading to transmission of S. aureus is highly plausible a priori. Indeed, the main route of transmission for S. aureus is physical contact; those led to CPI records as sensors recorded all physical proximity (<1.5m). The mechanism of transmission was therefore captured by CPIs. With the evidence we provide on the correlation between CPI paths and transmission events, this strengthens the interest of this proxy measure as a cheap, feasible and informative method for studying S. aureus transmission.
The joint analysis of S. aureus carriage and CPI data collected during 4 months provided evidence that CPIs capture contacts associated with transmission. This supports using CPI information to improve the realism of transmission models. This also suggests that a more systematic in-depth study of CPI networks could provide new directions for controlling S. aureus transmission in hospitals.
|
10.1371/journal.ppat.1006813 | Inner tegument proteins of Herpes Simplex Virus are sufficient for intracellular capsid motility in neurons but not for axonal targeting | Upon reactivation from latency and during lytic infections in neurons, alphaherpesviruses assemble cytosolic capsids, capsids associated with enveloping membranes, and transport vesicles harboring fully enveloped capsids. It is debated whether capsid envelopment of herpes simplex virus (HSV) is completed in the soma prior to axonal targeting or later, and whether the mechanisms are the same in neurons derived from embryos or from adult hosts. We used HSV mutants impaired in capsid envelopment to test whether the inner tegument proteins pUL36 or pUL37 necessary for microtubule-mediated capsid transport were sufficient for axonal capsid targeting in neurons derived from the dorsal root ganglia of adult mice. Such neurons were infected with HSV1-ΔUL20 whose capsids recruited pUL36 and pUL37, with HSV1-ΔUL37 whose capsids associate only with pUL36, or with HSV1-ΔUL36 that assembles capsids lacking both proteins. While capsids of HSV1-ΔUL20 were actively transported along microtubules in epithelial cells and in the somata of neurons, those of HSV1-ΔUL36 and -ΔUL37 could only diffuse in the cytoplasm. Employing a novel image analysis algorithm to quantify capsid targeting to axons, we show that only a few capsids of HSV1-ΔUL20 entered axons, while vesicles transporting gD utilized axonal transport efficiently and independently of pUL36, pUL37, or pUL20. Our data indicate that capsid motility in the somata of neurons mediated by pUL36 and pUL37 does not suffice for targeting capsids to axons, and suggest that capsid envelopment needs to be completed in the soma prior to targeting of herpes simplex virus to the axons, and to spreading from neurons to neighboring cells.
| Human and animal alphaherpesviruses establish lifelong latent infections in neurons of the peripheral nervous system and cause many diseases upon primary infection as well as following reactivation from latency. The highly prevalent human herpes simplex viruses HSV-1 and HSV-2 are responsible for facial and genital herpes, potentially blinding eye infections, and life-threatening encephalitis and meningitis. Here, we asked how these viruses master the bottleneck of being targeted from the neuronal somata to the axons. Our data suggest that only transport vesicles harboring fully matured virus particles can enter axons for spreading infection to the brain or the peripheral organs. Our data imply that the limiting membrane of the transport vesicles must expose viral or host receptors to recruit the microtubule motors required for axonal transport. Inhibiting such viral factors on the surface of the transport vesicles might provide novel therapeutic approaches to prevent the spread of alphaherpesviruses in the nervous system.
| Many pathogens spread within the nervous system by fast intracellular axonal transport mediated by microtubule motors, and progeny particles are transmitted via synapses from neuron to neuron [1,2]. Such life style allows pathogens to hide from the extracellular complement system, from antibodies, and from immune cells. But it comes with the challenge of finding the tiny exit gates from the neuronal somata, and of specific targeting to dendrites or to axons. It has been estimated that the cross section of the axonal outlet occupies less than 2 thousandth of the surface area of the soma [3,4]. The molecular post codes achieving this remarkable intracellular trafficking control are not well understood; neither for host nor for pathogen cargoes.
Alphaherpesviruses such as herpes simplex viruses (HSV) replicate in the skin, the eyes, and the oral, genital, or nasal mucosa, and then enter local axon endings of sensory or autonomic neurons in which they establish latency with little gene expression. Latent HSV-1 genomes have been detected in human cranial ganglia, especially in the trigeminal ganglia, while dorsal root ganglia (DRG) can harbor both, HSV-1 and HSV-2 [5–7]. Viral gene expression can be reactivated upon stress and a decline in the local immune responses, and progeny viruses are transported back to the periphery, where they cause lytic infections. Efficient targeting of alphaherpesviruses to the axons is essential for axonal spread and thus disease. Severe outcomes are potentially blinding herpes keratitis upon spread within the eyes or life-threatening herpes encephalitis upon spread within the brain [reviewed in 8,9].
The HSV-1 virion contains a large double-stranded DNA genome encased in a viral capsid, numerous tegument proteins and an envelope with many viral membrane proteins. Furthermore several host proteins and mRNAs have been detected in highly purified inocula [10–12]. HSV-1 assembly begins in the nucleus with genome packaging into capsids, which then traverse the nuclear envelope [reviewed in 13,14]. Cytosolic capsids associate with the inner tegument proteins pUL36 and pUL37 for their intracellular transport along microtubules to cytoplasmic membranes, where they meet other tegument and viral membrane proteins for secondary envelopment and virion formation [15–24]. In addition to pUL36 and pUL37, other structural proteins are required for efficient capsid envelopment. HSV-1 mutants lacking either pUL36 or pUL37, or the membrane proteins pUL20 or glycoprotein K (gK) are severely impaired in cytoplasmic envelopment, and accumulate cytosolic capsids instead [15,16,18,20,25–32]. HSV1-pUL20 and gK form functional complexes that connect with the capsids via pUL37 which in turn binds to pUL36 and the small capsid protein VP26 [33–36]. Another prominent tegument linker is VP16, with binding sites for pUL36, VP11/12, VP22 and gH [22,34,37–39]. VP22 in turn can bind ICP0, pUL16, gD, gE and gM [reviewed in 21,22]. The resulting vesicles are transported to the cell periphery and fuse with the plasma membrane to release infectious virions [reviewed in 13,14,22].
HSV-1 cytosolic capsids and complete virions within transport vesicles are targeted from the neuronal somata to axons to a varying extent; this has led to different hypotheses on the mode of neuronal alphaherpesvirus assembly [reviewed in 23,40,41]. According to the married model, capsid envelopment occurs exclusively in the soma, and only transport vesicles harboring complete virions enter axons. The separate model refers to different cargo structures being targeted independently of each other to the axons; namely capsids with associated tegument proteins as well as vesicles harboring viral envelope proteins and tegument proteins associated to their cytosolic tails.
Many structural proteins contribute to the neuronal spread of alphaherpesviruses, but the molecular determinants that are required for microtubule motor recruitment and for targeting from the soma to the axon gate have not been fully dissected. Purified HSV-1 capsids with inner tegument proteins can recruit the microtubule motors kinesin-1, kinesin-2, dynein and its cofactor dynactin to their surface, are translocated along microtubules in vitro, and can dock to nuclear pores [11,42–44]. pUL36 of HSV-1 and of the porcine pseudorabies virus (PRV) harbor potential binding motives for kinesin-1 light chains, and PRV-pUL36 has been shown to bind dynein and dynactin [45–47]. For these reasons and based on many additional functional studies, pUL36 and pUL37 are considered the most likely candidates for microtubule motor recruitment [reviewed in 23,45,46,48].
We therefore hypothesized that pUL36 and pUL37 might be sufficient for axonal capsid targeting and axonal capsid transport. As capsid envelopment may be a fast process, and since therefore the steady-state concentration of cytosolic capsids might be low, we used HSV-1 mutants to enrich for assembly intermediates. We infected neurons with HSV1-ΔUL20, HSV1-ΔUL37 or HSV1-ΔUL36 that accumulated cytosolic capsids decorated with pUL36 and pUL37, only with pUL36, or lacking both. Our previous data indicate that acquisition of both pUL36 and pUL37 is essential to enable capsids to enlist microtubule transport in epithelial cells [20]. However, while these inner tegument proteins enabled intracellular capsid motility of HSV1-ΔUL20 in epithelial cells and the somata of DRG neurons, they did not suffice to target cytosolic capsids to the axons. Our data indicate that HSV-1 particles have to acquire additional tegument and envelope proteins, and suggest that cytoplasmic envelopment needs to be completed prior to axonal transport which is in accordance with the married model for assembly of alphaherpesviruses in neurons.
To investigate HSV-1 axonal targeting, we cultured primary neurons derived from dissociated DRG of adult mice until they had developed mature neurites. Within 3 to 5 days of culture in vitro (div), the neurons expressed the axonal microtubule-associated protein tau (not shown), phosphorylated neurofilament, un-phosphorylated neurofilament, and ankyrinG. In the somata, there were short β-III-tubulin microtubules and careful analysis often revealed a perinuclear microtubule-organizing center (S1Aii Fig, arrow), but individual microtubules could not be discerned in the neurites (S1A Fig). There were less phosphorylated neurofilament H and M in the somata than in the neurites (S1B Fig), while non-phosphorylated neurofilament H epitopes were distributed more evenly (S1C Fig), as reported for rat nervous tissue [49]. Likewise ankyrinG, another axonal marker was targeted to the neurites (S1D Fig). In situ, peripheral sensory axons also contain ankyrinG along the entire axons up to the dermal-epidermal junction [50].
To evaluate the microtubule polarity in these neurites, we transduced the neurons at 1 div with a lentiviral vector expressing end-binding protein 3 (EB3), which associates with dynamically growing microtubule plus-ends [51]. At 4 div, live-cell imaging as well as time-projections showed that all EB3-GFP comets moved away from the soma, suggesting that most, if not all, microtubule plus-ends faced towards the distal ends of the neurites (not shown). These results indicate that the DRG neurons had regrown neurites with axonal features and with a unipolar microtubule polarity. We have shown previously that such neurons are productively infected, and that they transmit HSV-1 to co-cultured epithelial cells upon infection with the HSV1(17+)Lox-Che (c.f. Table 1), which had been cloned into a bacterial artificial chromosome (BAC), and which expresses mCherry as a reporter [52,53].
To generate cytosolic capsids with different tegument protein composition, and thus to determine whether the requirements for intracellular capsid motility parallel those for axonal targeting, we constructed HSV1(17+)Lox-CheVP26-ΔUL36 [20], -CheVP26-ΔUL37 [20], and -CheVP26-ΔUL20 (this study) in the same genetic background (c.f. Table 1). We deleted the ATG start codon and a second ATG, and inserted three stop codons into the UL20 gene, since it includes the promotor of UL19 that codes for the major capsid protein VP5. Restriction analyses of pHSV1(17+)Lox-ΔUL20 and -CheVP26-ΔUL20 indicated the addition of mCherry to VP26 (AscI and BamHI, C in S2A Fig), and that a resistance gene inserted with the UL20 mutations had been removed (XhoI, asterisks in S2A Fig). Further HindIII (S2A Fig), EcoRI, or EcoRV digestions resulted in the expected fragment sizes (not shown). HSV1(17+)Lox-ΔUL20 and -CheVP26-ΔUL20 were recovered by transfecting the corresponding BACs into Flp-In-CV-1-cells that express pUL20 in trans [54]. Sequencing of the mutated region confirmed the introduction of the intended changes (not shown). The lack of the ATGs and the introduced stop codons prevented the expression of pUL20, whereas pUL37 expression was unchanged (S2B Fig). The intra- and extracellular titers of HSV1(17+)Lox-ΔUL20 and -CheVP26-ΔUL20 were about 1,000-fold lower than their parental strains in non-complementing Vero cells, but higher in a pUL20 complementing cell line (S2C and S2D Fig). Similar results have been reported for HSV1-ΔUL20 mutants in other genetic backgrounds [25,26,54–56]. There were little differences between the parental Lox and the -CheVP26 strains, indicating that tagging VP26 with mCherry (Che) did not impair HSV-1 replication, as reported before [24,57].
Using conventional electron microscopy, we next analyzed virus morphogenesis. Upon infection of Vero cells with HSV1(17+)Lox (Fig 1A) or -CheVP26 (not shown), viral particle maturation proceeded as expected with the formation of nuclear capsids, the appearance of cytosolic capsids, partially and completely enveloped cytoplasmic capsids, and extracellular virions associated with the plasma membrane. With secondary envelopment, capsids acquired an electron dense tegument layer, and the electron dense genomes were well preserved (Fig 1Ai and 1Aiii; asterisk). Therefore the morphology of capsids and tegument of virions inside authentic transport vesicles was more similar to that of extracellular virions (Fig 1Aii; black arrow) than to that of cytosolic capsids (Fig 2Aiii; white arrowhead) even if they were partially enveloped (Fig 1Ai and 1Aiii, black arrowhead). Nuclear capsid egress of the mutants HSV1(17+)Lox-ΔUL20 (Fig 1Bi and 1Bii) and -CheVP26-ΔUL20 (Fig 1Biii) was unaffected, resulting in many cytosolic capsids, and several of them were closely associated with cytoplasmic membranes (Fig 1Bi and 1Bii; black arrowhead). We detected neither capsids with the characteristic morphology of completed secondary envelopment nor virions bound to the plasma membrane in the absence of pUL20. Furthermore, in contrast to infection with HSV1(17+)Lox-ΔUL36 or -ΔUL37 [20], we did not detect any cytoplasmic or extracellular L-particles, which are viral envelopes with an electron dense tegument but lacking capsids [58–60].
Next, we analyzed DRG cultures infected with HSV1(17+)Lox-CheVP26, -CheVP26-ΔUL36, -CheVP26-ΔUL37, -CheVP26-ΔUL20 (Fig 2) or their respective untagged strains (not shown) by electron microscopy. Neurons identified by the characteristic morphology of their nuclei and infected with the parental HSV-1 strains contained nuclear capsids, primary envelopment intermediates (Fig 2A, white star), cytosolic capsids (Fig 2A, white arrowhead), and enveloped virions (Fig 2A, black star). Neurons infected with HSV1(17+)Lox-CheVP26-ΔUL36 (Fig 2B), -CheVP26-ΔUL37 (Fig 2C), or -CheVP26-ΔUL20 (Fig 2D) contained non-enveloped cytosolic capsids (Fig 2, white arrowhead), but did not reveal any extracellular virions bound to their plasma membranes. HSV1(17+)Lox-CheVP26-ΔUL20 infected neurons contained in addition wrapping intermediates (Fig 2D, black arrowhead). A quantification of the number of cytoplasmic HSV-1 assembly intermediates revealed that cells infected with HSV1(17+)Lox-ΔUL20 or -CheVP26-ΔUL20 had a similar ratio of cytosolic capsids to membrane-associated capsids as upon infection with the respective parental strains HSV1(17+)Lox or -CheVP26 (Table 2).
To characterize the surface protein composition of the cytosolic capsids, Vero cells were infected with HSV1(17+)Lox or -ΔUL20 for 16 h, fixed, and ultrathin cryosections were prepared for quantitative immunoelectron microscopy. We have reported similar experiments for -ΔUL36 and -ΔUL37 before [20], and show their results here again for a direct comparison (Fig 3Ai and 3Aii). Cytosolic capsids of both, Lox [20] and -ΔUL20 were labeled with anti-pUL36 (Fig 3Aiii) and anti-pUL37 (Fig 3Aiv). There were on average 2.2 gold particles per capsid for HSV1(17+)Lox, 2 for -ΔUL37 and -ΔUL20, but only 0.5 for -ΔUL36 (Fig 3Bi) using anti-pUL36 antibodies. Upon labeling with anti-pUL37, there were 1.7 or 1.5 gold particles per capsid for HSV1(17+)Lox or -ΔUL20, but only 0.5 for -ΔUL36 and -ΔUL37 (Fig 3Bii). These results indicate that cytoplasmic capsids of Lox-ΔUL20 had recruited pUL36 and pUL37 to a similar extent as those of the parental HSV1(17+)Lox, while capsids of -ΔUL37 lacked pUL37 but still recruited pUL36, and capsids of -ΔUL36 lacked both, pUL36 and pUL37. Thus, the HSV1(17+)Lox-ΔUL36, -ΔUL37, and -ΔUL20 as well as HSV1(17+)Lox-CheVP26-ΔUL36, -ΔUL37, and -ΔUL20 mutants generated cytosolic capsids exposing different proteins on their surface, namely no inner tegument, only pUL36, or both pUL36 and pUL37.
To characterize the intracellular capsid motility in the presence of different inner tegument proteins, we infected Vero cells with HSV1(17+)Lox-CheVP26, -CheVP26-ΔUL36, -CheVP26-ΔUL37, or -CheVP26-ΔUL20 and acquired confocal fluorescence time-lapse movies with a temporal resolution of 5 images per second. For a direct comparison, we analyzed in parallel movies with -CheVP26 and -CheVP26-ΔUL20 recorded in this study as well as movies that we had recorded with -CheVP26-ΔUL36 and -CheVP26-ΔUL37 previously and published in Sandbaumhüter et al. [20]. The time-projections of the tracks show that capsids of Lox-CheVP26 (Fig 4Ai) and -CheVP26-ΔUL20 (Fig 4Aiv) exerted short and long range transport towards the nucleus and towards the cell periphery (S1 and S4 Movies). In contrast, the tracks of -CheVP26-ΔUL36 (Fig 4Aii) and -CheVP26-ΔUL37 (Fig 4Aiii) revealed only random, undirected motility (S2 and S3 Movies).
While the average track length (Fig 4Bi) and the maximum step velocity (Fig 4Bii) were similar for the capsids of -CheVP26, -CheVP26-ΔUL36, -CheVP26-ΔUL37, and -CheVP26-ΔUL20, there were more tracks with a length of more than 5 μm and with a velocity exceeding 1 μm/s for the capsids of the parental Lox-CheVP26 and the -CheVP26-ΔUL20 than for the capsids of -CheVP26-ΔUL36 and -CheVP26-ΔUL37. We furthermore determined the mean square displacement exponent (MSDex) for each track (Fig 4Biii). The MSDex is a characteristic feature for the displacement mode of a single motile particle from its starting position over time. The MSD is defined as the square of the travelled distance from the starting point of the track, which is calculated for each time point, and plotted against the time. The slope of such a curve in a log-log plot defines the MSDex and indicates whether a particle has undergone active transport (MSDex > 1), free diffusion (MSDex = 1), or confined diffusion [MSDex<1; 61,62,63]. While the MSDex of the majority of the tracks denoted free or confined diffusion as expected for any intracellular cargo, the MSDex was greater than 1 for more tracks of Lox-CheVP26 and -CheVP26-ΔUL20 than for -CheVP26-ΔUL36 or -CheVP26-ΔUL37.
We furthermore infected DRG neurons with HSV1(17+)Lox-CheVP26, -CheVP26-ΔUL36, -CheVP26-ΔUL37, or -CheVP26-ΔUL20 and acquired spinning disk microscopy time lapse movies of the cell bodies with a temporal resolution of 20 images per second. Similar as in the epithelial cells, capsids of the parental (Fig 5Ai; S5 Movie) and -CheVP26-ΔUL20 (Fig 5Aiv; S8 Movie) were transported over short and longer distances as indicated by the time projection lines. In contrast, the capsids of -CheVP26-ΔUL36 (Fig 5Aii; S6 Movie) and -CheVP26-ΔUL37 (Fig 5Aiii, S7 Movie) only moved in an undirected fashion. Quantification revealed that the transport characteristics in DRG neuronal cell bodies were the same as in epithelial cells (Fig 5Bi to 5iii). Thus, the motility features of the cytosolic capsids of -CheVP26-ΔUL36 or -CheVP26-ΔUL37 in epithelial cells and sensory neurons were indicative for intracellular diffusion which is also the mode of motility in the absence of microtubules after nocodazole treatment [62]. In contrast, the capsids of -CheVP26 and -CheVP26-ΔUL20 had transport characteristics typical for directed, active microtubule-dependent transport.
After we had characterized the intracellular motility of capsids in the presence of pUL36 and pUL37 (-ΔUL20), of only pUL36 (-ΔUL37), or lacking both (-ΔUL36), we asked which capsid types could be targeted to the axonal outlet. We infected primary DRG neurons with HSV1(17+)Lox, -ΔUL36, -ΔUL37, or ΔUL20 (Fig 6) or the respective HSV1(17+)Lox-CheVP26 strains (S3 Fig, Fig 7). At 24 to 26 hpi, the cells were fixed, permeabilized and labeled with antibodies directed against the capsid protein VP26 (Fig 6i and 6iv panels, green in 6iii), and the tegument proteins VP22 (Fig 6A–6D, ii and v panels, red in iii and vi), the tegument protein VP13/14 (Fig 7, ii panels, green in iii), the envelope protein gD (Fig 6E–6H, ii and iv panels, red iii and vi), the envelope protein gB (Fig 7, vi panels, green in viii), or the neuron specific beta-3-tubulin (Fig 7vii). Representative images from at least 3 independent experiments for the different strains are shown in Figs 6, S3 and 7.
After infection with HSV1(17+)Lox (Fig 6A and 6E) or HSV1(17+)Lox-CheVP26 (Figs S3A and 7A), the nucleoplasm (Figs 6Ai, 6Ei and S3Ai), the cytoplasm (Figs 6Ai, 6Ei and S3Ai) and the axons (Fig 6Aiv, 6Eiv and 7A) were filled with capsids; in the cytoplasm, the capsids were often clustered. Infection with -ΔUL36 (Figs 6B, 6F, S3B and 7B) resulted in a dispersed and more random distribution of capsids in the cytoplasm of the somata (S3B Fig), with fewer capsids being targeted to the axons (Figs 6Biv, 6Fi, 6Fv and 7B). After infection with -ΔUL37 (Figs 6C, 6G, S3C and 7C), capsids clustered within the cytoplasm and the cell morphology was often altered; again very few capsids had been targeted to the axons (Figs 6Civ, 6Giv, 7Ci and 7Cv). The stronger cytopathic effects as indicated by membrane blebbing upon infection upon infection with the HSV1-ΔUL37 strains (S3C Fig) might be due to the lack of pUL37 targeting RIG-I and blocking RNA-induced activation [64], but we did not investigate this further.
Unexpectedly, there were also only very few axonal capsids upon infection with -ΔUL20 (Figs 6Hiv and 7D). The capsid distribution of -ΔUL20 in the somata was similar to that of the parental HSV1(17+)Lox, but with reduced accumulation at the plasma membrane (Fig 6Div, S3Di). After infection with -ΔUL36, -ΔUL37, or -ΔUL20, the outer tegument proteins VP22 (Fig 6Bv, 6Cv and 6Dv) and VP13/14 (Fig 7Bii, 7Cii and 7Dii) as well as gD (Fig 6Fv, 6Gv and 6Hv) and the envelope proteins gB (Fig 7Bvi, 7Cvi and 7Dvi) had still been targeted to the axons as for the parental strain (Figs 6A, 6E and 7A), indicating that axonal transport per se had not been impaired. Nevertheless, the deletion mutants -ΔUL36, -ΔUL37, and -ΔUL20 were incapable of efficient axonal capsid transport.
For quantitation, we scaled up two experiments to image random axonal regions and determined the number of different viral structures per axon length using a novel automated image analysis algorithm (S4 Fig). We determined the imaged axon length, the total number of capsids as detected by anti-VP26, and the fraction of these capsids colocalizing with gD or VP22 (c.f. S1 Table). After infection with the parental strain HSV1(17+)Lox, we counted 20 to 30 capsids per 100 μm axon length (Fig 8A; total VP26, black bars) of which 75% colocalized with gD (Fig 8B; VP26 + gD). Furthermore, there was a similar number of membrane structures containing gD and colocalizing with VP26 or not (Fig 8C, gD); the size of these structures varied considerably (Fig 8D). After infection with Lox-ΔUL36 (Fig 8; light grey columns), or -ΔUL20 (Fig 8; dark grey columns), there were only few capsids targeted to the axons (Fig 8A), which again mostly co-localized with gD (Fig 8B). However, the number of gD-containing membrane structures was only moderately reduced for the HSV-1 mutants suggesting that axonal transport per se had not been inhibited (Fig 8C). In uninfected neurons, there was also some anti-gD background signal (Fig 8C, mock), but these cross-reacting structures were much smaller (Fig 8D). We also quantified the number of capsids detected by anti-VP26 (Fig 8E), and the fraction of these capsids colocalizing with the outer tegument protein VP22 (Fig 8F) of a parallel set of samples. After infection with the parental strain HSV1(17+)Lox, we counted again 20 to 30 capsids per 100 μm axon length (Fig 8E; total VP26, black bars), of which 60% to 70% colocalized with VP22 (Fig 8F; VP26 + VP22). Furthermore, there was a similar number of structures containing VP22 colocalizing with VP26 or not (Fig 8G; VP22) which also varied considerably in size (Fig 8H). After infection with Lox-ΔUL36, or -ΔUL20, again fewer capsids had been targeted to the axons (Fig 8E), of which almost 80% co-localized with VP22 (Fig 8F). The number of VP22 containing structures was moderately reduced for the HSV-1 mutants supporting the notion that axonal transport per se had not been inhibited (Fig 8G). In uninfected neurons, there was almost no anti-VP22 background signal (Fig 8G, mock), and again these cross-reacting structures were much smaller (Fig 8H).
These experiments show that the inner tegument proteins pUL36 and pUL37 on the capsids of -ΔUL20 were not sufficient for axonal targeting. Instead, association with membranes and outer tegument proteins, and thus possibly completion of secondary envelopment seem to be a prerequisite for efficient capsid targeting into the axons.
Animal models as well as cultured neurons have contributed tremendously to elucidating the mechanisms of viral neuroinvasion, spread within the nervous system, and peripheral recurrent diseases [1,2,23]. Various aspects of human diseases are mimicked in murine HSV-1 infection models [65]. For example, upon infection of the cornea, HSV-1 reaches the trigeminal ganglia and causes herpetic keratitis; infection of the flank skin establishes latency in the DRG, and results in zosteriform skin lesions upon reactivation [52,66–70]. To further characterize alphaherpesvirus axonal targeting, we investigated here HSV infection of primary neurons derived from the DRGs of adult mice, and developed novel image processing algorithms to determine the amount of different axonal assembly intermediates in an unbiased fashion. We infected such neurons with HSV-1 mutants that assemble cytosolic capsids lacking pUL36 and pUL37, capsids associated only with pUL36, or capsids that recruited both, pUL36 and pUL37 to determine whether these inner tegument proteins implicated in microtubule-mediated transport were sufficient for intracellular capsid motility in epithelial cells and the somata of neurons as well as for capsid targeting to axons. Our data indicate that in addition to acquiring pUL36 and pUL37, capsids needed to complete secondary envelopment to be efficiently targeted to axons, to spread from neurons to neighboring cells, and thus to cause recurrent diseases.
Embryonic and neonatal neurons dissociated from trigeminal ganglia, DRG, or superior cervical ganglia of chicken, mouse, rat, or human have been used to study axonal trafficking and egress of HSV particles [71–78]. However, primary cultures from embryonic tissue do not present optimal models for age-related changes in differentiation, physiology, or late-onset disease [79–81]. We and others [82] therefore used neurons from adult mice to complement results obtained by infecting adult animals with HSV-1 [74,83] and to obtain functional insights into the process of axonal targeting of HSV-1 viral structures. The neurons from the DRG of adult mice formed neurites with axonal features which contained ankyrinG and microtubules of uniform polarity with the plus-ends pointing towards the axon endings. Furthermore, we have shown previously that HSV1(17+)Lox strains that have been cloned into a bacterial artificial chromosome and therefore lack the OriL [84,85] infect such neurons and spread the infection to neighboring epithelial cells [52,53]. Hence murine adult DRG neurons provide a versatile system to study productive HSV-1 infection and transport in axons with unipolar microtubules in vivo and in vitro.
While there is a growing consensus that the swine alphaherpesvirus PRV relies on the married model for axonal transport in all neuron types, the picture is less clear for HSV-1. According to the married model, capsids are enveloped exclusively in the somata, and thus it would be sufficient to expose a host or a viral postal code on the cytosolic membrane surface to target vesicles harboring fully assembled virions to axons [reviewed in 23,40,41]. But for HSV-1 two types of axonal cargo, free cytosolic capsids as well as capsids surrounded by an envelope and a transport vesicle membrane, have been detected to a varying degree [71,73,76,78,86–88].
Complementary methods have been used to image viral assembly intermediates. Antibodies have less access to capsids associated with membranes than to cytosolic capsids [89,90]. While fluorescent protein tags circumvent this challenge, they may shift the relative abundance of different assembly intermediates [57,72,73,84,91–94]. Furthermore, the resolution limit of confocal microscopy is larger than a capsid diameter of 125 nm leading to an overestimation of capsid colocalization with membrane markers. On the other hand, conventional electron micrographs provide sufficient contrast to distinguish hexagonal capsids from membrane vesicles or fully enveloped capsids within transport vesicles but only upon sufficient heavy metal deposition and in very thin sections of 50 nm, which results in an underestimation of membrane-associated capsids [47,71,87]. In embryonic rat hippocampal neurons, many cytosolic HSV-1 capsids are indeed not associated with cytoplasmic membranes as shown by 3-dimensional cryoelectron tomography [88]. The latter is in fact the best imaging method for this question, but limited to specialized centers and not amenable to medium-throughput [95]. For these reasons, we developed novel image quantitation algorithms and novel mutants to characterize HSV-1 axonal targeting, and used both, fluorescent tagging and antibodies to detect the capsid protein VP26, in combination with antibodies against the tegument proteins VP22 and VP13/14, and the envelope proteins gD and gB.
The separate model of alphaherpesvirus egress postulates that different subassemblies, namely cytosolic capsids as well as membrane vesicles conveying envelope proteins, are targeted to axons to be assembled in the axons or just prior to trans-synaptic spread or even during budding at the axonal plasma membrane [reviewed in 23,40,41]. Neuronal cargoes enlist the so-called smart motors that utilize stabilized microtubules which lead to the tiny outlets to dendrites or to axons [96,97]. One such smart motor is kinesin-1 that preferentially binds to de-tyrosinated and acetylated microtubules, modifications that are typical for axonal microtubules while others, such as kinesin-5 or kinesin-13 prefer tyrosinated microtubules [97,98]. Based on extensive genetic, biochemical and cell biology experiments, pUL36 and pUL37 are considered the most likely candidates for motor recruitment to cytosolic capsids [reviewed in 45,48]. Alphaherpesviruses lacking pUL36 or pUL37 are impaired in microtubule-mediated intracellular transport [15,16,18–20,30–32,99–101]. HSV-1 capsids exposing the inner tegument proteins pUL36 and pUL37 but not naked capsids translocate along microtubules in vitro and recruit kinesin-1 and kinesin-2 from a brain cytosolic extract [11,43]. Finally, pUL36 of HSV-1 and PRV contains binding sites for dynein and dynactin, and possibly the light chains of kinesin-1 [46,47].
We therefore generated HSV-1 mutants that assembled un-enveloped cytosolic capsids with different tegument proteins. Consistent with their proposed function to recruit microtubule motors, the capsids of HSV1-ΔUL20 harboring pUL36 and pUL37 but not the ones of HSV1-ΔUL36 or HSV1-ΔUL37 were capable of active transport in epithelial cells or the somata of DRG neurons. The capsids of HSV1-ΔUL20 therefore resemble capsids which recruit the smart motor kinesin-1 from a cytosolic brain extract in vitro [11]. Our data indicate that capsid-associated pUL36 and pUL37 were sufficient for microtubule transport, but not for efficient axonal targeting.
It is possible that the capsids of HSV1-ΔUL20 had been associated with a larger amount of outer tegument proteins covering potential motor-binding sites on the inner tegument. Capsids with lower amounts of outer tegument recruit more kinesin-1 from a cytosolic brain extract than capsids with more outer tegument [11]. Furthermore these outer tegument proteins might lead to an association with cytoplasmic membranes, although our quantitative electron microscopy data indicate that while HSV1-ΔUL20 was impaired in completing secondary envelopment, there was no higher association with wrapping membranes when compared to the respective parental strains. Those capsids associated with wrapping membranes might have been connected to larger membrane systems in neurons, which might have reduced their motility, and thus their chances to find the axonal exit. Axonal transport in general had not been impaired upon infection with either HSV-1 mutant, since the envelope proteins gB and gD as well as the outer tegument proteins VP13/14 and VP22 had been targeted to axons. Our data are consistent with the notion that cytosolic capsids might rely on the inner tegument for kinesin-1 mediated transport in epithelial cells and the somata, but that transport vesicles harboring complete virions are more efficiently targeted to axons than such cytosolic capsids.
Our data indicate that in addition to pUL36 and pUL37, other HSV-1 components and processes depending on pUL20 are required for axonal targeting in adult neurons. The simplest interpretation of our data is that HSV-1 relies also on the married model for axonal targeting and trafficking, and that the membrane of the transport vesicles harboring complete virions harbors a viral or a host receptor for smart axonal microtubule motors on its cytosolic surface. Indeed, kinesin-1 can co-traffic with PRV and HSV-1 tegument and envelope proteins [102–104]. In the absence of pUL20, secondary envelopment was not completed, and apparently also the tegument adopted a different conformation or composition, as our electron microscopy analysis revealed a different tegument contrast for capsids associated with wrapping membranes than for complete virions. Furthermore without pUL20, the envelope proteins gD and gH/gL, and possibly also their interaction partners VP22 and VP16 [38], are not properly targeted to cytoplasmic virus assembly sites [105].
Of the alphaherpesvirus envelope proteins, gE/gI and pUS9 are also necessary for efficient axonal spread but not for cytoplasmic capsid envelopment [23]. pUS9 of PRV associates with the neuron-specific kinesin-3 subunit KIF1A, and contributes to initial axonal sorting of PRV particles [106–108]. Similarly, pUS9 of HSV-1 is required for efficient axonal targeting, and its large cytosolic domain interacts with the KIF5B subunit of conventional kinesin-1 [74,109]. If the intracellular trafficking and targeting of gE/gI or pUS9 had been impaired in the absence of pUL20, this would explain why even capsids associated with wrapping membranes could not be targeted to the axons upon infection with HSV1-ΔUL20. According to the loading hypothesis, capsids require in the somata in addition to the inner tegument proteins pUL36 and pUL37 also the membrane proteins gE/gI and pUS9 to be loaded onto kinesin-1 [13,109,110].
Vesicles harboring gD and possibly VP22 did neither require pUL36, pUL37 nor pUL20 for axonal targeting. Future work elucidating which assembly intermediates of wild-type HSV-1 as well as -ΔUL36, -ΔUL37, -ΔUL20 and other deletion mutants recruit which smart microtubule motors will reveal specific viral and specific host components that need to be assembled onto an alphaherpesvirus cargo for efficient axonal targeting. Our data are consistent with the notion that cytosolic capsids rely on the inner tegument for dynein and kinesin-1 mediated transport prior to secondary envelopment in epithelial cells and in the neuronal somata, but that transport vesicles harboring complete virions are more efficiently targeted to axons than such cytosolic capsids.
Cell lines were cultured in a humidified incubator at 37°C and 5% CO2 and passaged twice per week. BHK-21 cells (ATCC CCL-10) were maintained in MEM (Cytogen, Wetzlar, Germany) supplemented with 10% (v/v) FBS (fetal bovine serum; PAA Laboratories GmbH, Cölbe, Germany), Vero cells (ATCC CCL-81) in MEM supplemented with 7.5% (v/v) FBS, and HEK-293T cells (ATCC CRL-11268) and Flp-In-CV-1 cells in DMEM (Invitrogen, Karlsruhe, Germany) supplemented with 2 mM L-glutamine and 10% FBS (v/v). The pUL36 trans-complementing Vero-derived HS30 cell line [31] was provided by Prashant Desai (Johns Hopkins University, Baltimore, USA) and the pUL37 trans-complementing rabbit skin 80C02 cell line [15] by Frazer J. Rixon (University of Glasgow, Scotland, UK). These complementing cells were maintained in MEM containing 1% non-essential amino acids (Cytogen) and 7.5% or 10% (v/v) FBS, respectively. Every 5th passage was cultured in the presence of G418 (500 μg/ml; PAA Laboratories GmbH). The trans-complementing Flp-In-CV-1-derived cell line expressing pUL20 under the control of the HSV-1 gD promoter [54] was provided by Konstantin G. Kousoulas (Louisiana State University, Louisiana, USA) and maintained in DMEM (Invitrogen) supplemented with 10% (v/v) FBS and 2 mM L-glutamine and 125 μg/ml hygromycin B (Invitrogen).
Primary neurons from dorsal root ganglia (DRG) of adult C57Bl/6JHanZtm mice were cultured using established protocols [111,112]. Briefly, mice were sacrificed, and the DRG from the cervical, thoracic and lumbar level were dissected and collected in 1x HBSS-complete buffer (Hank’s balanced salt solution, pH 7.4 with 5 mM HEPES and 10 mM D-Glucose). The DRG of three mice were pooled and treated with 20 mg/ml papain (Sigma-Aldrich, Schnelldorf, Germany; in 0.4 mg/ml L-Cysteine, 0.5 mM EDTA, 1.5 mM CaCl2, pH 7.4) for 20 min at 37°C, and with 10 mg/ml collagenase IV (Invitrogen) and 12 mg/ml dispase II (Sigma-Aldrich) in 1x HBSS-complete buffer for 20 min at 37°C. DRG and cells were sedimented and re-suspended in 1 ml 1xHBSS-complete buffer and triturated using Pasteur pipettes with narrowed ends. The suspensions were spun for 8 min at 381 x g through 20% (v/v) Percoll (Sigma-Aldrich) cushions in CO2-independent medium (Life Technologies Gibco, Carlsbad, CA, USA) containing 10 mM D-glucose, 5 mM HEPES, 10% FBS, 100 U/ml penicillin and 0.1 mg/ml streptomycin. The cells were washed with 2 ml CO2-independent medium, sedimented 2 min at 1,000 x g, suspended in Ham’s F-12 nutrient mix medium with 10% FBS, 50 ng/ml 2.5S nerve growth factor (Promega Corporation, Fitchburg, WI, USA), 100 U/ml penicillin and 0.1 mg/ml streptomycin, and seeded onto cover slips of 20 mm diameter in 24-well plates or glass bottom dishes (Nunc LabTek II Chambered cover glass 4-chamber #155382, #1.5 borosilicate glass, Thermo Scientific). The cover slips and glass bottom dishes had been pre-coated with 0.01% (w/v) poly-L-lysine (150,000–300,000 g/mol, Sigma-Aldrich) and 7 ng/μl murine laminin (Invitrogen). The cells were cultured in a humidified incubator at 37°C and 5% CO2, and the media were replaced twice a week. 1-β-D-arabinofuranosylcytosine (Sigma-Aldrich) was added at 1 to 2 div to a final concentration of 2μM to suppress proliferation of dividing, non-neuronal cells, but removed prior to HSV-1 infection.
We used rabbit polyclonal antibodies (pAb) to detect VP26 [VP26aa95-112; 113], pUL36 [#147; 43], pUL37 [32], pUL20 [114], VP13/14 [R22, 115], gB [R69, 116] or VP16 (BD 3844–1, Becton-Dickinson, Franklin Lakes, NJ, USA). Mock infected neurons showed little binding to antisera raised against HSV-1 proteins that had been cleared by pre-adsorption [89] on uninfected neurons. Mouse monoclonal antibodies (mAb) were used to detect VP22 [mAb22-3; 117] or gD [mAb DL6; 118]. To detect host antigens, we used mAb 1501 for actin (Millipore, Billerica, MA, USA), mAb 5564 for β-III-tubulin (Millipore), mAb 106/36 for ankyrinG (E9PE32; UC Davis/NIH NeuroMab Facility), SMI-310 for phosphorylated 200 kDa and 160 kDa neurofilament (ab24570, Abcam, Cambridge, UK), and SMI-320 for non-phosphorylated 200 kDa neurofilament (ab28029, Abcam).
We used the clinical isolate HSV1(17+) [119] and its derivatives HSV1(17+)Lox [20], HSV1(17+)Lox-CheVP26 [20], HSV1(17+)Lox-ΔUL36 [20,47], HSV1(17+)Lox-CheVP26-ΔUL36 [20,47], HSV1(17+)Lox-ΔUL37 [20], HSV1(17+)Lox-CheVP26-ΔUL37 [20], HSV1(17+)Lox-ΔUL20 (see below) and HSV1(17+)Lox-CheVP26-ΔUL20 (see below). Our pHSV1(17+)Lox BAC plasmids contain a BAC cassette with a chloramphenicol resistance gene, a Cre recombinase gene with an intron under the control of a eukaryotic promoter, a single flippase recognition target site (FRT) and LoxP sites at both ends inserted between the genes UL22 and UL23, and an almost complete HSV1(17+) genome lacking only the OriL [84,85]. The Cre recombinase excises the BAC cassette upon transfection into eukaryotic cells. HSV-1 stocks were prepared in BHK-21, HS30 (for -ΔUL36), 80C02 (for -ΔUL37) or Flp-In-CV-1-pUL20 (for -ΔUL20), and extracellular virus was harvested by sedimentation from the supernatant of infected cells as described previously [84,89,90]. After DNase treatment, the inocula of HSV1(17+)Lox-ΔUL37 and Lox-CheVP26-ΔUL37 had genome to PFU ratios below 2,300 while that of the other strains had genome to PFU ratios of below 108 [120].
To express EB3 in neurons, we used lentiviral transduction. The spleen focus forming virus promoter was replaced by the human cytomegalovirus immediate early promoter in the plasmid pRRL.PPT.SF.GFPpre [121, provided by Axel Schambach, Hannover Medical School, Hannover, Germany] via PCR by generating 5’PstI and 3’BamHI restriction sites adjacent to the HCMV immediate early promoter of pCMV-Tag 2B (Agilent Technologies, Santa Clara, California, USA) using the primers 5’-GAACCTGCAGCGTATTACCGCCATGCATTAGT-3’ and 5’-GAACGGATCCCCAGCTTTTGTTCCCTTTAGTG-3’. Furthermore, with the restriction sites 5’NdeI and 3’AgeI, parts of the human cytomegalovirus immediate early promoter and EB3 were cloned from pEGFP-N-EB3 [51, provided by Marco van Ham, Helmholtz Centre for Infection Research, Braunschweig, Germany] to generate pRRL.PPT.HCMV.GFPEB3pre. HEK 293T cells (5 x 106 per 10 cm dish) were transfected with 5 μg pRSV_Rev (provided by Axel Schambach), 2 μg pMD2.g (Addgene Inc., Cambridge, MA, USA, Cat. No. 12259, deposited by D. Trono, provided by Axel Schambach), 10 μg pCDNA3.GP.CCCC (provided by Axel Schambach), and 10 μg transfer plasmid as described previously [122]. The supernatants were harvested and spun in a Beckman SW 28 rotor at 23,000 rpm or a SW32.Ti rotor at 24,000 rpm for 90 min at 4°C (Beckman Coulter, Krefeld, Germany). The resuspended lentiviral particles were snap frozen in liquid N2 and stored at -80°C. For lentiviral transduction, DRG neurons were prepared and seeded as described and at 1 div neuronal growth medium was replaced by one containing 20 mM HEPES and the lentiviral particles.
HSV-1 inocula titers were determined by plaque assays [89,90,120]. Vero, Flp-In-CV-1 or the derivative Flp-In-CV-1-UL20-expressing cells were cultured in 6-well dishes for 16 to 20 h to almost confluency. The respective inocula were diluted in CO2-independent medium (Life Technologies Gibco) containing 0.1% (w/v) cell-culture grade bovine serum albumin (BSA; Capricorn Scientific, Ebersdorfergrund, Germany), added to the cells for 1 h on a rocking platform at room temperature, and then replaced by growth medium containing 20 μg/ml pooled human IgGs (Sigma-Aldrich) to neutralize HSV-1 in the culture medium. At 3 dpi, the cells were fixed with ice-cold, water-free methanol and air-dried prior to staining with 0.1% (w/v) crystal violet in 2% (v/v) ethanol for 1 min. After removing the excess of crystal violet, the cells were air-dried, and plaques were counted while using a binocular loupe (Nikon, Tokyo, Japan) to calculate the virus titer as plaque forming units (pfu) per ml.
We used the BACs pHSV1(17+)Lox and pHSV1(17+)Lox-CheVP26 to construct HSV1(17+)Lox-ΔUL20 and HSV1(17+)Lox-CheVP26-ΔUL20. To prevent expression of pUL20, we mutated the ATG-start codon and a second ATG codon of UL20 to CTG followed by an immediate insertion of 3 stop codons. Recombinant PCR fragments of pEPkan-S2 [provided by B. Karsten Tischer and Nikolaus Osterrieder, Freie Universität Berlin, Germany, 123] were amplified to mutate the 5’ region of UL20 using traceless Red recombination, and transformed for homologous recombination into E. coli GS1783 (provided by G. Smith, Northwestern University, Chicago, IL, USA) harboring pHSV1(17+)Lox or pHSV1(17+)Lox-CheVP26 [85,124]. We used the forward primer 5’-CCTTGCGGTTTCGGTCTCCCCACCTCCACCGCACACCCCCTGACCCTGTAGTAATAGCGGGATGACCTTCCTCTGGTTAGGGAAACAGGTAATCGATTT-3’, the reverse primer 5’-TCGTCGACCAGATCTCGATCACCAGAGGAAGGTCATCCCGCTATTACTACAGGGTCAGGGGGTGTGCGGTGGCAGGTGGTGCCAGTGTTACAACCAATTAACC-3’, the forward sequencing primer 5’-AAAGACCGGCTGGGTATG-3’, and the reverse sequencing primer 5’-GGGCGTAGGCGTAAATTC-3’. The mutated start codons are underlined and the inserted stop codons are shown in bold. The BAC plasmids were digested with 15 U/μg DNA of AscI, BamHI, XhoI or HindIII for 3.5 h, and analyzed on 0.6% (w/v) agarose gels in 0.5x TBE buffer (0.44 M Tris-HCl, 0.44 mM boric acid, 10 mM EDTA, pH 8) run at 66 mA for 17 h (Peqlab system, Erlangen, Germany). To reconstitute viruses, sub-confluent Vero or Flp-In-CV-1-UL20-expressing cells were transfected with 10 μg BAC-DNA (MBS mammalian transfection kit; Stratagene, La Jolla, CA, USA) per 6 cm dish and cultured until cytopathic effects had developed. Cells and medium were pooled, virus was released by three cycles of freeze-thawing and the resulting cell pellet was used to further propagate the virus according to standard protocols. The genomes of the novel HSV1(17+)Lox strains were sequenced around the pUL20 start site that had been targeted, and the engineered mutations were confirmed. For the growth curves, sub-confluent Vero, Flp-In-CV-1 or the derivative Flp-In-CV-1-UL20-expressing cells were inoculated at 5 pfu/cell, the cells and the supernatants were harvested at the indicated time points, and virus titers were determined by plaque assay.
Vero cells were seeded on cover slips, 3.5 cm cell culture dishes or glass-bottom chambers (Nunc LabTek II Chambered cover glass 4-chamber #155382, #1.5 borosilicate glass, Thermo Scientific) and infected 16 to 20 hours after the seeding as described before [16,20,29,84,89,120,125]. For synchronous infections, the cells were pre-cooled for 20 min on ice, and inoculated with 10 pfu per cell or mock treated as a control in CO2-independent medium containing 0.1% (w/v) BSA for 2 h on ice while rocking. The cells were then shifted to regular growth medium at 37°C and 5% CO2 for 1 h. Non-internalized virus was inactivated at 4°C by a 3 min acid wash (40 mM citrate, 135 mM NaCl, 10 mM KCl, pH 3). Neurons cultured on cover slips in 24-well plates were pre-incubated at room temperature with CO2-independent medium for 20 min, and inoculated with 1 to 5 × 106 PFU in 200 μl per well in CO2-independent medium. After 30 min, the virus-suspension was replaced by 500 μl F-12 medium and cells were incubated again at 37°C and 5% CO2.
Cells were lysed with hot sample buffer (50 mM Tris-HCl, pH 6.8, 1% [w/v] SDS, 1% [v/v] β-mercaptoethanol, 5% [v/v] glycerol, 0.001% [w/v] bromophenol blue) containing protease inhibitors AEL (aprotinin, E-64, leupeptin, Sigma), ABP (antipain, bestatin, pepstatin, Sigma) and PMSF (Roth, Karlsruhe, Germany). The proteins were separated by SDS-PAGE in 12.5% gels and transferred in 48 mM Tris, 380 mM glycine, 0.1% [w/v] SDS and 10% [v/v] methanol to nitrocellulose membranes (Pall Corporation, Pensacola, FL, USA). After blocking with 5% (w/v) low-fat milk powder in PBS containing 0.1% (v/v) Tween 20, the membranes were incubated with primary antibodies and secondary antibodies coupled to alkaline phosphatase (Dianova, Hamburg, Germany), transferred to 100 mM Tris-HCl, pH 9.5, 100 mM NaCl, 5 mM MgCl2, and stained with 0.2 mM nitroblue tetrazolium chloride and 0.8 mM 5-bromo-4-chloro-indolyl-3-phosphate. For documentation, the membranes were imaged with a digital scanner (ScanJet 6300, Hewlett Packard, Wilmington, DE, USA).
Cells were fixed at room temperature with 3% (w/v) paraformaldehyde (PFA) in PBS for 20 min, followed by 50 mM NH4Cl for 10 min and permeabilization with 0.1% Triton-X-100 for 5 min, or at 37°C with PHEMO fix [3.7% (w/v) PFA, 0.05% [w/v] glutaraldehyde, 0.5% [v/v] Triton-X-100 in PHEMO buffer with 68 mM PIPES, 25 mM HEPES, pH 6.9, 15 mM EGTA, 3 mM MgCl2, 10% [v/v] dimethyl sulfoxide], washed two times with PHEMO buffer followed by the NH4Cl treatment. The HSV-1 Fc-receptor and unspecific protein binding sites were blocked with 10% (v/v) human serum of HSV-1-seronegative healthy volunteers and 0.5% (w/v) BSA. Samples were labeled with primary and pre-adsorbed secondary antibodies to prevent cross reactivity to antibodies of other species, namely goat-anti-mouse coupled with rhodamine X, carbocyanine 5 or AlexaFluor488, and goat-anti-rabbit coupled with AlexaFluor488 (Invitrogen). DNA was stained using To-Pro-3-iodide or DAPI dyes (Invitrogen), and the cover slips were mounted in Mowiol 4–88 containing 10% (w/v) 1,4-diazabicyclo-[2.2.2]octane. The specimens were analyzed by confocal fluorescence microscopy (LSM 510 Meta, software LSM 510 version 4, ZEISS, Göttingen, Germany). Pseudo-coloring, brightness and contrast were adjusted identically across each set of images using Adobe Photoshop CS4.
To analyze systematically the degree of co-localization of tegument (anti-VP22, red channel) or envelope proteins (anti-gD, red channel) with capsids (anti-VP26, green channel), and the targeting efficiency of the different viral particles to the axons, we developed a novel image processing pipeline. The semi-automated approach consists of three steps: (i) manual annotation of regions of interest (ROIs) to identify the axons (cf. S4 Fig), (ii) automated particle detection of the capsids, and (iii) automated detection of the co-localizing tegument or envelope protein signal using pixel-wise image classification.
(i) ROI selection: To distinguish the axon from other structures, the regions of interest were manually selected. To simplify the process of region annotation, a ROI is defined by positioning a few control points in the image and adjusting the region width at the control points, as shown in S4 Fig. Usually, three to four control points were sufficient for accurate annotation.
(ii) Capsid particle detection: The capsids have a spherical shape with a characteristic diameter. Due to blurring caused by the image acquisition process, the fluorescence signal of a single capsid was modeled by an isotropic Gaussian distribution. Template matching was applied to discriminate between well-separated viral particles and other signals such as background noise or agglomerations of multiple, non-separable capsids. In detail, a detector response map Rσ(x, y) was computed for each pixel (x, y) of the image using normalized cross-correlation of the image signal Icapsid(x,y) with a Gaussian shaped template model Tσ(u,v)=exp[−u2+v22σ2] with variance σ2 and filter size of 1 + 2 ⋅ ⌈σ ⋅ s⌉. Candidate particles were extracted from the detector response map Rσ(x, y) by searching for local maxima in Rσ(x, y) using an 8-neighborhood. In case of a plateau, a morphological shrinking operator was used to only consider the central position of the plateau as a detection candidate. To discriminate from noise, the candidate points pi = (xi, yi) were only accepted as capsids if the detector response Rσ(xi,yi) was greater than a threshold τcorr and the fluorescence signal Icapsid(xi,yi) was greater than a threshold τintens. The parameters σ, s, τcorr, and τintens were trained on manually annotated test images using grid search. The final parameter values used throughout the experiments were σ = 0.7, s = 1.5, τcorr = 0.4, and τintens = 0.5.
(iii) Tegument or envelope protein signal detection: We used pixel-wise image classification in order to discriminate between signal and background noise. The classifier used various features derived from the gD or VP22 fluorescence signals IgD(x,y): Gaussian blurred intensity signal IσgD=IgD*Tσ, non-linearly distorted intensity Es;σgD=exp(s⋅IσgD) normalized cross correlation (NCC) of IgD to a Gaussian kernel RσgD=NCC(IgD,Tσ), and a high-pass filter clamped to values between 0 and 1. Finally, the feature vector is
F=[R0.5gD,R1gD,R2gD,R4gD,IgD,E10;0gD,I2gD,H2gD,I5gD,H5gD,E10;5gD,I10gD,H10gD,E20;10gD,I20gD,H20gD,E50;20gD].
We used a logistic regression classifier which was trained on manually annotated test images. The gD or VP22 regions were extracted from the resulting pixel-wise classification images using a connected components analysis with an 8-neighborhood. A capsid was said to co-localize with gD or VP22 if the capsid center position coincided with a gD or VP22 region dilated by 1 pixel. The dilation operation accounted for uncertainties in both the capsid position estimation and the pixel-wise gD or VP22 classification.
Several statistical parameter were derived from the automatically detected capsids and gD or VP22 regions. The number of capsids or gD and VP22 regions per 100 μm was derived by projecting the particle locations orthogonally to the axial center line of the ROI as illustrated in S4 Fig.
For differential interference contrast imaging and to avoid evaporation at 37°C, a glass lid was attached to the chambers with vacuum grease (Dow Corning, Midland, Michigan, USA). Movies of infected Vero cells were recorded at 37° with a high temporal resolution of 5 images per second and a pixel size of 79 nm using a confocal laser scanning microscope equipped with a heating unit (PeCon, Erbach, Germany) as described before [20]. Movies of infected DRG neurons were recorded at 37°C (Incubator from PeCon, Erbach, Germany) with high temporal resolution 20 images per second on a Nikon Ti microscope equipped with a Yokogawa (Ratingen, Germany) CSU-X1 spinning disk and an Andor iXion Ultra 897 EMCCD camera (Belfast, UK).
Automated tracking of cytoplasmic capsid motility was performed using the ImageJ plugin MOSAIC [63,126]. First, the nuclei were identified by the confined nuclear capsids mobility and the nuclear capsid fluorescence was removed. Second, the Gaussian blur filter of ImageJ was applied (Sigma radius 2 pixel). No further image processing was performed. Third, automated tracking was performed using the Particle Tracking plugin with the following settings: Kernel radius 3, Cutoff radius 3, percentile 0.5, displacement 20, and link range 1. Of the tracks identified, we only considered further tracks with a length of 10 frames or longer. Track length and maximum step velocity were calculated from tracks with a MSD exponent of 1.2 or larger, representing non diffusive transport events.
Cells were infected as described above and fixed at the indicated time points. For conventional electron microscopy cells were fixed with 2% glutaraldehyde in cacodylate buffer (130 mM (CH3)2AsO2H, pH 7.4, 2 mM CaCl2, 10 mM MgCl2) for 1 h at room temperature. Cells were washed and contrasted with 1% (w/v) OsO4 in cacodylate buffer (165 mM (CH3)2AsO2H, pH 7.4, 1.5% (w/v) K3[Fe(CH)6] followed by 0.5% (w/v) uranyl acetate in 50% (v/v) ethanol overnight. The cells were embedded in Epon plastic (Serva, Heidelberg, Germany) and 50 nm ultrathin sections were cut parallel to the substrate. Cytoplasmic capsids were counted and the respective cellular areas that had been sampled were measured using an Image J plugin.
For immunoelectron microscopy, infected cells were fixed at the indicated time points with 2% (w/v) PFA and 0.2% [w/v] glutaraldehyde, in PHEM-buffer (60 mM PIPES, 25 mM HEPES, pH 6.9, 10 mM EGTA, 2 mM MgCl2), embedded, frozen and sectioned. Sections were labeled using specific antibodies and protein-A gold of 10 nm (Cell Microscopy Centre, Utrecht School of Medicine, The Netherlands). Sections were contrasted using 0.5% (w/v) uranyl acetate in 2% methylcellulose (Merck) [127]. Images were acquired with an electron microscope at 200 kV equipped with an Eagle 4k camera (Tecnai G2; FEI, Eindhoven, The Netherlands). Immunogold labeling was quantified by counting each gold particle within a radius of 100 nm around the center of a cytoplasmic capsid [20]. Data were analyzed by using Kruskal-Wallis followed by Dunn’s post testing, and the p values were adjusted for multiple testing (software Prism, version 6; Graphpad, San Diego, CA, USA).
The mice (strain C57Bl/6JHanZtm, not genetically modified) were bred and maintained without any perturbation. On the day of the experiment they were picked up from the animal facility and within 3 h sedated with CO2-inhalation prior to killing by cervical dislocation without any prior experimental perturbation. DRG were dissected afterwards. According to the German Animal Welfare Law §4, killing of animals does not need approval if the removal of organs serves scientific purposes and the mice had not undergone experimental treatment before. The animal care and sacrifice was performed in strict accordance with the German regulations of the Society for Laboratory Animal Science (GV-SOLAS), the European Health Law of the Federation of Laboratory Animal Science Association (FELASA) and the German Animal Welfare Law. According to the German Animal Welfare Law, this study does not contain animal experiments that require pre-approval, but the total number of killed mice was reported at the end of each year to the animal welfare deputy of Hannover Medical School. The number of animals killed according to §4 of the German Animal Welfare Law was registered with the animal welfare application number 2012/20 at the LAVES (Niedersaechsisches Landesamt fuer Verbraucherschutz und Lebensmittelsicherheit, Oldenburg, Germany), and the experiments were performed before 2013. Human sera of adult, healthy, HSV-1 seronegative volunteers were obtained after written informed consent by the blood donors. Permission was granted by the Institution Review Board (Hannover Medical School; Approval Number 893).
|
10.1371/journal.pcbi.1002050 | How Modeling Can Reconcile Apparently Discrepant Experimental Results: The Case of Pacemaking in Dopaminergic Neurons | Midbrain dopaminergic neurons are endowed with endogenous slow pacemaking properties. In recent years, many different groups have studied the basis for this phenomenon, often with conflicting conclusions. In particular, the role of a slowly-inactivating L-type calcium channel in the depolarizing phase between spikes is controversial, and the analysis of slow oscillatory potential (SOP) recordings during the blockade of sodium channels has led to conflicting conclusions. Based on a minimal model of a dopaminergic neuron, our analysis suggests that the same experimental protocol may lead to drastically different observations in almost identical neurons. For example, complete L-type calcium channel blockade eliminates spontaneous firing or has almost no effect in two neurons differing by less than 1% in their maximal sodium conductance. The same prediction can be reproduced in a state of the art detailed model of a dopaminergic neuron. Some of these predictions are confirmed experimentally using single-cell recordings in brain slices. Our minimal model exhibits SOPs when sodium channels are blocked, these SOPs being uncorrelated with the spiking activity, as has been shown experimentally. We also show that block of a specific conductance (in this case, the SK conductance) can have a different effect on these two oscillatory behaviors (pacemaking and SOPs), despite the fact that they have the same initiating mechanism. These results highlight the fact that computational approaches, besides their well known confirmatory and predictive interests in neurophysiology, may also be useful to resolve apparent discrepancies between experimental results.
| Dopamine is a neurotransmitter which plays important roles in the control of voluntary movement, motivation and reward, attention, and learning. Dysfunction of midbrain dopaminergic systems is involved in various diseases such as Parkinson's disease, schizophrenia and drug abuse. This underlines the importance of a tight regulation of dopamine levels in the brain. At the cellular level, the release of dopamine is directly correlated to the type of electrical activity (the firing pattern) of nerve cells that produce it, the so-called “dopaminergic neurons”. Therefore, an in depth understanding of the mechanisms underlying the electrical behavior of dopaminergic neurons is of critical importance to find new strategies for the treatment of diseases that result from dysfunction of this system.
| Midbrain dopaminergic (DA) neurons sustain important physiological functions such as control of movement [1] and signalling of positive error in reward prediction [2]. A dysfunction of the DA system is implicated in the pathophysiology of Parkinsons disease, schizophrenia, and drug abuse [3]. Under physiological conditions, DA neurons can switch between three distinct modes: tonic (pacemaker), irregular, and burst firing [4], [5].
The nature of the channels involved in the low frequency pacemaking of DA neurons is still strongly discussed. Indeed, whereas many studies have shown that L-type calcium channels are critical for this spontaneous activity, others, including ours, have observed little effect of a blockade of these channels on this firing pattern (see Table 1). Therefore, the respective contribution of calcium and sodium channels in pacemaking remains unclear. On the other hand, it is commonly accepted that low-frequency spontaneous firing requires oscillations in the cytoplasmic free calcium concentration [6], [7].
In the presence of the sodium channel blocker tetrodotoxin (TTX), DA neurons also exhibit slow oscillatory potentials (SOPs) [8], [9], which have been shown to be sustained by L-type calcium channels [10]. Guzman et al. recently observed that SOPs and spikes are not correlated in rate and regularity, from what they concluded that pacemaking and SOPs are driven by different mechanisms [10]. This conclusion is used to support the hypothesis that L-type calcium channels would not strongly contribute to pacemaking in DA neurons, which is in opposition with many experimental results [6], [11] (Table 1). An additional controversy results from the fact that the block of SK channels prolongs depolarizing plateaus under sodium channel inhibition, whereas it only slightly affects the firing rate when the neurons fire action potentials.
In this paper, we use a mathematical analysis to extract the mechanisms underlying the spontaneous activity of DA neurons. For this purpose, we develop a minimal model of a DA neuron, in which we include the minimal set of conductances that are able to reproduce the firing patterns exhibited by these cells (see below). This minimal model is able to exhibit pacemaker firing in the absence of synaptic afferents, and to switch from a low frequency single-spike firing to bursting when SK channels are blocked in the presence of excitatory inputs, as reported experimentally [12], [13]. Moreover, SOPs are present during the inhibition of sodium channels, and the effect of a SK channel blockade observed experimentally is reproduced in the model. In order to validate our analyses of the minimal model, we also test its predictions on a published detailed model of these neurons [14], [15].
The main conclusion of our analysis is that pacemaker firing in DA neurons is sustained by the cooperation of sodium and L-type calcium channels (and more modestly N-type or P/Q-type calcium channels [6]), whereas variations of the intracellular calcium concentration play a major role in the rate of this spontaneous firing pattern. On the basis of this mechanism, we identify potential causes for the experimental discrepancies mentioned above, using our minimal model, as well as the detailed model. We observe that neurons only differing by less than 1% in their maximal sodium conductance react oppositely to a blockade of L-type calcium channels. Experiments performed in rat brain slices confirm that L-type calcium and sodium channels cooperate to generate pacemaking in these neurons.
As a secondary conclusion, our model shows that, even though the initiation of SOPs and spikes is sustained by the same mechanism, these oscillatory patterns are not correlated, which is in agreement with experimental results [10]. We show that this absence of correlation is due to different mechanisms of depolarizing phase termination in the two oscillatory behaviors, as well as different kinetics of calcium entry (resp. exit) during depolarizing phases (resp. hyperpolarizing phases). These results show that the lack of correlation between pacemaking and SOPs does not exclude a shared mechanism.
In order to extract the essential mechanisms of pacemaking of DA neurons, we developed a minimal model endowed with the minimal set of conductances which is necessary to reproduce the firing patterns of these cells. The conductances present in the model are shown in Fig. 1 and the corresponding equations are detailed in the Methods section. In parallel with the membrane capacitance and a leak current, the model is composed of sodium channels and delayed-rectifier potassium channels for the generation of action potentials (fast dynamics). Calcium enters through L-type calcium channels endowed with calcium dependent inactivation (CDI) [16], [17], whereas calcium extrusion is carried out by calcium pumps. In addition, small conductance calcium-activated potassium (SK) channels are present. Finally, an excitatory synaptic current is added to test the effect of noise of variable amplitude on the firing of the modeled cell (small amplitude noise to mimic in vitro-like conditions and high amplitude noise to mimic in vivo-like conditions). In the presence of SK channels, the minimal model does not produce high firing frequencies in response to a step of applied depolarizing current but it does so in response to a high amplitude synaptic input. Calcium dynamics is well-known to be essential to classical models of bursting neurons [18], hence the combination of a calcium-regulated inward current (carried by L-type calcium channels) and a calcium-regulated outward current (carried by calcium pumps). In our model, however, the neuron produces only one action potential per oscillation cycle of , which results in low-frequency single-spike firing rather than bursting in in vitro-like conditions (Drion et. al, unpublished).
The proposed minimal model is able to reproduce the firing patterns exhibited by DA neurons, namely pacemaker firing in vitro, irregular single-spike and burst firing in vivo, or be in a hyperpolarized state (Fig. 2). As it has been shown experimentally, the switch between irregular single-spike firing and bursting can be induced by a blockade of small conductance calcium-activated potassium (SK) channels [12], [13]. In the remainder of this paper, we will only discuss the mechanisms of spontaneous firing in vitro (i.e. when the neuron is submitted to small amplitude noise). Also, in order to verify that our simplified model correctly captures the firing mechanisms of DA neurons, we systematically test the predictions of the minimal model against simulations performed with the detailed model (see the equivalent circuit in [14]). This much more complex model exhibits the majority of the channels which have been reported to be present in DA neurons experimentally and takes the general architecture of the cell into account.
The frequency of the spontaneous activity is limited by calcium influx. Indeed, a rise of the cytoplasmic calcium concentration strongly reduces the excitability of the cell. Therefore, to ensure a calcium entry (resp. exit) during the spike generation (resp. between two successive spikes), one or several types of calcium channels (resp. calcium pumps) must be present. Moreover, in order to generate a spontaneous activity even in the absence of calcium-activated potassium channels (as is observed in DA neurons [19]), the calcium dynamics must be regulated by calcium, e.g. through the presence of calcium-inactivated calcium channels. Under these two conditions, the cell is able to generate a low-frequency spontaneous activity as well as fast firing when submitted to excitatory synaptic inputs, with a very different firing rate between these two firing patterns (0.5 to 5 Hz and more than 15 Hz, respectively, see Fig. 2).
In the absence of synaptic inputs, DA neurons fire spontaneously in a very regular manner [19], [20]. The mechanisms underlying this pacemaker activity are still strongly discussed, experimental results being contradictory (Table 1). However, it is commonly accepted that pacemaking of DA neurons requires calcium oscillations and that SK channels are not critical to sustain this firing pattern [19]. Fig. 3 illustrates the mechanisms involved in pacemaker firing in the minimal model. As it has been demonstrated experimentally [10], the spontaneous activity is synchronized with calcium oscillations. Namely, each action potential is generated when the intracellular calcium concentration reaches a constant minimal value (Fig. 3A). This is consistent with experimental results [9], [10].
The mechanisms involved in the pacemaker activity of the minimal model can be fully understood using bifurcation analysis. The bifurcation diagram shown in Fig. 3B illustrates how the spike generation is governed by the intracellular calcium concentration, the latter being the bifurcation parameter. The diagram defines three distinct ranges of intracellular calcium concentration : a low range where the only stable steady-state is depolarization block (); a high range where the only stable steady-state is hyperpolarization (); and an intermediate bistable range where a limit cycle may coexist with the two stable steady-states (). The thresholds separating these zones are mainly dependent on the regulation of calcium channels and calcium pumps by the intracellular calcium concentration. Thus, a rise of the intracellular calcium concentration induces an inactivation of the L-type calcium channels [16], [17], which reduces the amount of inward (i.e. depolarizing) current, and an activation of calcium pumps, which increases the amount of outward (i.e. repolarizing) current [21]. As a consequence, the excitability of the cell decreases with the intracellular calcium concentration. A variation of the intracellular calcium concentration that exceeds the high threshold will induce a switch from firing to a hyperpolarized state. Therefore, the high threshold, which is defined as the value of at which the stable limit cycle disappears, defines the maximal possible intracellular calcium concentration which is compatible with firing.
The low threshold in Fig. 3 defines the value of intracellular calcium concentration at which an action potential is spontaneously generated. As a consequence, the current which initiates the depolarization at this point is the critical one for pacemaking. For instance, for the set of parameter values used in Fig. 3, action potentials are generated by an opening of L-type calcium channels, and a complete blockade of these channels prevents firing. But if we use other sets of values for the conductance of sodium and L-type calcium channels, the mechanism of spontaneous initiation of spikes can vary. This suggests a strong cooperation between sodium and L-type calcium channels to drive the pacemaker activity of DA neurons, as suggested by Guzman et. al [10].
The important consequences of this cooperation are illustrated in Fig. 4. The threshold separating a hyperpolarized state (white) from pacemaking (blue) is defined by an almost linear combination of the L-type calcium conductance and the sodium conductance . Moreover, when the conductance of one channel type is above a threshold value, these channels are sufficient to drive a spontaneous activity during blockade of the others. As a consequence, very similar neurons which would have minimal differences in their sodium or L-type calcium channel density may exhibit very different responses to experimental manipulations that shut off one of the two conductances, as illustrated in the various inserts of Fig. 4.
Panels A to D of Fig. 4 show the behavior of the minimal model in control conditions and during blockade of L-type calcium channels or sodium channels for a particular set of conductances. The blockade is modeled by setting the conductance to zero. Only the sodium and L-type calcium conductances slightly differ in the four represented situations. Note that the electrical behavior of each modeled neuron in control conditions is almost similar: the differences in parameter values are very small, and the mechanisms that underly their spontaneous activity are similar. However, these neurons react very differently to the blockade of one conductance. In the case of neuron D (), L-type calcium current inhibition completely inhibits the spontaneous activity of the cell, whereas an oscillatory behavior remains after a sodium current inhibition. On the basis of these experimental-like scenarii, we would conclude that the pacemaker activity of neuron D is driven by L-type calcium channels. If we examine neuron A (), L-type calcium current inhibition barely affects the firing rate and pattern of the cell, whereas a sodium current inhibition induces a hyperpolarization of the membrane. These observations would therefore lead to the opposite conclusion, namely that the pacemaker activity of neuron A is driven by sodium channels. As a consequence, the results of the two experiments would be contradictory, despite the great similarity of the neurons. More generally, Fig. 4 illustrates the fact that very similar neurons may produce drastically different responses to the same experimental manipulation. Remarkably, the four distinct behaviors exhibited by almost identical neurons in the minimal model can be exactly reproduced in the detailed model (see details in Supplementary Fig. S1).
In order to confirm that the spontaneous initiation of spikes in DA neurons is mainly sustained by the cooperation between sodium and L-type calcium channels, we performed extracellular recordings (additional to those reported in Tab. 1) of these neurons in slices from adult rats containing the substantia nigra pars compacta. A potential advantage of this recording method is that it does not disrupt the contents of the neuron, contrary to conventional patch-clamp recordings. For these experiments, we superfused the slices with blockers of synaptic transmission (10 CNQX, 1 MK801, 10 SR95531, 1 sulpiride and 1 CGP55845, which block AMPA, NMDA, , D2 and receptors, respectively), in order to isolate the neurons from their afferences. Control experiments showed that application of the synaptic blockers alone induced a small increase in firing rate which was stable for at least one hour (n = 4, supplementary Fig. S2).
We next tested the effect of 20 nifedipine (a L-type calcium channel blocker), 30 TTX (a sodium channel blocker), as well as their simultaneous application on the firing rate of DA cells, respectively. This concentration of TTX was used because it had been shown to block a major fraction (about 80) of the somatic sodium conductance [22], but did not abolish action potentials. The precise experimental protocol is shown graphically in Fig. 5B and is described in detail in the Methods section.
These recordings were performed on eleven neurons. In one case, nifedipine completely inhibited the spontaneous activity of the cell. In four other cells, nifedipine produced little effect, whereas TTX completely suppressed the firing (not shown). In the six other neurons, coapplication of nifedipine and TTX inhibited the firing to a greater extent than either agent alone. Indeed, an ANOVA test showed that the firing rate of the neurons was different in the four experimental conditions (synaptic blockers alone, +nifedipine, +TTX, +TTX and nifedipine, F[3], [20] = 21.12, p = 0.000002). The application of nifedipine alone did not significantly affect the firing rate of these cells (from 2.510.26 Hz to 2.110.16 Hz, mean s.e.m., p0.52, Tukey's post hoc test) (Fig. 5). The application of TTX alone significantly but only partially decreased the firing rate of the cells to 1.590.10 Hz (p0.05), and reduced the amplitude of spikes. The latter effect is in agreement with the fact that the maximal sodium conductance is quite reduced. The simultaneous application of TTX and nifedipine reduced the firing rate of the cells to 0.370.25 Hz (p0.001), and this reduction was significantly larger than the effect of either agent alone (p0.001 vs nifedipine, p0.01 vs TTX). Moreover, the simultaneous pharmacological block of sodium (80) and L-type calcium channels almost completely eliminated the spontaneous firing in 5 of these 6 cells (frequency 0.4 Hz). In summary, our experiments confirm that the degree of cooperation between the two currents is highly variable from neuron to neuron, even in a fixed experimental protocol.
The fact that L-type calcium channels and sodium channels can cooperatively drive the pacemaker activity of DA neurons can be explained by comparing the I-V curves of the two currents (Supplementary Fig. S3A). Indeed, Putzier et al. recently showed that the only critical parameter of the L-type calcium current for pacemaking is the value of its half-activation potential (), which should not be too negative [11]. Moreover, they showed that artificial NMDA receptors that would have a similar half-activation potential would also induce sustained firing. The half-activation potentials of sodium and L-type calcium channels are very similar (Supplementary Fig. S3A), which explains why they can cooperate in the pacemaking generation.
This similarity is observed in the detailed model as well (Supplementary Fig. S3B). If we compare the I-V curves of the other calcium currents, it is clear that, in this model, N-type calcium channels have a similar to the one of sodium and L-type calcium channels, whereas of T-type calcium channels is much more negative. On the basis of these observations, the previous analysis suggests that N-type calcium channels should be able to induce pacemaking if their density is high enough whereas T-type channels, which have a significant more negative , should not. We test this hypothesis in Fig. 6 (where we reduce the sodium conductance to reduce the contribution of sodium channels to the pacemaking) using the detailed model. As predicted, a N-type calcium current of sufficient amplitude is able to drive a low-frequency pacemaker activity, whereas a T-type calcium current is not, even with a very high maximal conductance value. Moreover, N-type channels are able to induce SOPs when sodium channels are blocked, as L-type channels do (Fig. 4C and D). These simulations are therefore in agreement with the prediction that whatever the nature of the depolarizing current, the only parameter which is critical is its voltage dependence, and more precisely its half-activation potential [11].
Our minimal model also sheds light on the mechanisms of, and related controversies on, slow oscillatory potentials (SOPs). Indeed, it has experimentally been shown that SNc DA neurons exhibit SOPs during application of TTX in vitro [8], [9], and that these SOPs are driven by L-type calcium channels [10]. Recently, it has been shown that pacemaking and SOPs are uncorrelated [10], which led to the conclusion that these two oscillatory behaviors are driven by different mechanisms, thus rejecting L-type calcium channels for pacemaking.
In addition, it has been reported that block of the SK current on TTX-treated DA neurons strongly affects the shape of SOPs. Namely, this manipulation significantly increases the duration of the depolarizing and hyperpolarizing phases [9]. This observation led to the hypothesis that, in control condition, this inhibition might induce burst firing in DA neurons. However, experiments have invalidated this intuitive suggestion, SK channel blockade only inducing irregularities in the firing of these cells, but not bursting (at least not reproducibly) [19]. These contradictory observations can be explained through an analysis of the behavior of the minimal model in configurations that mimic these experiments.
For a sufficiently high value of (compare Fig. 4C,D and Fig. 4A,B), SOPs are also observed after blockade of the sodium current in the minimal model (Fig. 7C). As in the case of pacemaking, these SOPs are synchronized with the calcium oscillations (supplementary Fig. S4). Fig. 7 shows how sodium channel blockade affects the bifurcation diagram in the presence and in the absence of SK channels. As illustrated in the figure, sodium channel blockade has no effect on the low threshold, which implies that a same mechanism initiates both spikes and SOPs in the minimal model.
In contrast, sodium channel blockade has a critical impact on the high threshold. Indeed, the Hopf bifurcation which defines the high threshold in pacemaking vanishes when sodium channels are blocked. As a consequence, the high threshold of SOPs is defined at the right saddle-node bifurcation, which is masked by the Hopf bifurcation in control conditions. This implies that, whereas the depolarization mechanism is identical in spikes and SOPs, the repolarization mechanisms are different.
This difference has significant consequences on the model behavior, all consistent with experimental data:
Fig. 8A shows that the interevent interval histograms of pacemaking (in light blue) and SOPs (in dark blue) do not match either in the presence (Fig. 8A, left) or in the absence (Fig. 8A, right) of SK channels. Moreover, depending on the parameters of the model, spike rate can be lower or higher than SOP oscillation rate (Fig. 8B), as observed experimentally [10]. In Fig. 8, the parameters that are varied are the conductances of L-type calcium channels and SK channels. The main origin of this lack of correlation is that the quantity and kinetics of calcium entry taking place during an action potential strongly differs from the one taking place during the depolarized phase of SOPs. This affects the rate of their respective oscillations (Supplementary Fig. S5).
In addition, Guzman et. al recently showed that spikes are much more regular than SOPs in DA neurons [10]. In its low-noise configuration (representing the in vitro condition), our model exhibits similar results (Fig. 8C, which can be compared to Fig. 3e of [10]). Indeed, we found that SOPs are much more sensitive to noise than spikes in the presence of SK channels, mainly because SOP oscillations induce less calcium entry than spikes and because the kinetics of calcium entry is slower (supplementary Fig. S5).
These results show that a lack of correlation between spikes and SOPs does not necessarily imply that the generating mechanism of these two oscillatory behaviors is different. Therefore, this experimental observation may not be used to dismiss a role of L-type calcium channels in the generation of spikes in DA neurons.
In spite of many experimental studies, the precise mechanisms underlying the spontaneous initiation of spikes in DA neurons are still largely debated in the literature. Using our minimal model as well as a detailed model of a DA neuron, we extracted two critical parameters for the low frequency spontaneous firing.
Firstly, low-frequency single-spike firing and high-frequency intra-bursts firing have to be sustained by two dynamics operating on different time scales. Fast firing is limited by the refractory period of action potentials, which is fixed by the kinetics of voltage-gated channels. On the other hand, the dynamics that are the most likely to limit the rate of low-frequency firing are the variations of the intracellular calcium concentration. Indeed, an accumulation of calcium in the cytoplasm strongly reduces the excitability of the cells, through the inactivation of depolarizing currents (i.e. L-type calcium currents) and the activation of hyperpolarizing currents (i.e. calcium pumps and SK channels). This is in agreement with experimental data, which show that replacement of calcium with either cobalt [6] or magnesium [7] strongly affects pacemaking.
Secondly, the spontaneous initiation of action potentials in DA neurons is the result of the cooperation between various depolarizing currents. In agreement with the experimental results of Putzier et al. [11], we found in the detailed model that any depolarizing current having a half-activation potential less negative than (voltage-dependant sodium channels, L-type and N-type calcium channels) may play a role in this initiation. Their relative contribution, as well as the robustness of pacemaking, depend on the respective density of each channel type. Therefore, the fact that the selective blockade of a particular channel does not completely disrupt pacemaking does not mean that these channels are not involved in physiological pacemaking. This might be an important note of caution for experimentalists. Moreover, we confirmed experimentally that L-type calcium and sodium channels do indeed cooperate to generate pacemaking with a degree of cooperation that is highly variable. This precise observation has never been made previously.
The most contradictory experimental results obtained on DA neurons are probably those concerning the role of L-type calcium channels in the spontaneous initiation of spikes in vitro (Table 1). Using very similar modeled neurons that differ by less than 1% in one conductance parameter (all remaining parameters being identical), we were able to reproduce these contradictory results both in our minimal model and in a detailed model of DA neurons.
Such subtle differences in conductance parameters are quite likely to occur in various experimental conditions. For example, it has recently been shown that there are quantitative differences between DA neurons from the SNc and the VTA in terms of density of these conductances [6], [7]. It was proposed that sodium channels play a major role in the spike generation of VTA DA cells, whereas calcium channels are predominant in SNc DA neurons. Among other findings, replacement of calcium with cobalt in SNc neurons completely inhibits the firing, whereas replacement of calcium with magnesium in VTA neurons increases the firing rate [6], [7]. Both these effects are also observed in the minimal and detailed models with slightly different maximal sodium conductances (supplementary Fig. S6).
A second source of contradictory experimental results might be the difference between the preparations that are used in different laboratories. For instance, in the case of DA neurons, in which the initial segment is often remote from the soma [23], it is clear that the total sodium current will be smaller in acutely dissociated neurons than in neurons recorded in the slice preparation. In terms of the model that we have developed, this probably means that the small conductance variations illustrated in Fig. 4 could result from minor experimental variations such as dissociated neurons vs neurons recorded in slices, as well as variable developmental stages of the animals.
SOPs exhibited by DA neurons during blockade of sodium channels have been largely studied [8]–[10] and compared to the spontaneous spiking activity. Moreover, it has been recently shown that the two oscillatory patterns are not correlated [10], a phenomenon that is also observed in our model. It is tempting to conclude from this observation that their underlying mechanisms are different. However, our analysis shows that the generation mechanisms are actually the same. Moreover, Fig. 7 clearly illustrates that the different effect of SK channel blockade on pacemaking and SOPs simply arises from the fact that the high intracellular calcium concentration threshold only slightly changes when sodium channels are present, whereas a more than two fold change in this high threshold occurs when sodium channels are blocked.
DA neuron electrophysiology has been modeled by several groups [14], [24]–[29]. Most of these models were elaborated in order to reproduce experimental observations. Although they clearly succeed at making detailed predictions, their complexity may prevent a detailed analysis of the mechanisms underlying pacemaking in these neurons. On the basis of these models, we attempted to develop a minimal model containing only the most critical conductances needed to reproduce firing patterns of DA neurons. This model allowed us to explain some discrepancies found in the literature on the mechanisms underlying pacemaking of these cells. Importantly, we were able to confirm our conclusions on one detailed model [14].
There are some conceptual differences between our minimal model and some earlier models. For example, the calcium current included in the Wilson and Callaway model [26] does not include any calcium-induced inactivation, whereas ours does. We included inactivation because recent experiments have demonstrated that the L-type channels expressed by DA neurons belong to the class [30] and this channel subtype is known to undergo calcium-induced inactivation in expression systems [17]. Clearly, this point has to be carefully investigated in future experiments on DA neurons. It should be pointed out, however, that inactivation of L-type channels is not fundamental for the results that we obtained in our model. Indeed, the major role of this phenomenon is to decrease the excitability of the cell when intracellular calcium rises after the action potential. This decrease could be induced by other mechanisms, such as activation of a potassium ERG-type current [29].
One limitation of our minimal model is that it is essentially qualitative and does not take into account the very specifics of DA neurons. However, because of its generality, our model could be a starting point to analyze similar firing patterns of a number of other cell types. Finally, our analysis demonstrates the value of a simplified model to reconcile apparently contradictory experimental observations.
All procedures were carried out in accordance with guidelines of the European Communities Council Directive of 24 November 1986 (86 609 EEC) and were accepted by the Ethics Committee for Animal Use of the University of Liège (protocol 86).
The model follows the common equation(1)where is the membrane capacitance, the membrane potential of the cell, accounts for all the ionic currents and is any externally applied current. The fast dynamics include a sodium current and a delayed-rectifier potassium current . These currents are responsible for the creation of action potentials. The slow dynamics describe the fluctuations of . Calcium influx is mediated by a L-T-type calcium current , which is both voltage-gated and calcium-regulated, and the calcium is pumped out of the cell by calcium pumps, which generate an outward calcium current . In order to reproduce in vivo conditions, the model can be subjected to excitatory synaptic inputs, which activate a synaptic current . The background of this synaptic activity () was modeled through stochastic waveforms. A calcium-activated potassium current and a leak current are also implemented in the model.
Simulations were performed using the Neuron software, which is freely available for download at http://www.neuron.yale.edu. Analyses were performed with Matlab 7.4.0.
|
10.1371/journal.pntd.0005432 | Selective inhibition of RNA polymerase I transcription as a potential approach to treat African trypanosomiasis | Trypanosoma brucei relies on an essential Variant Surface Glycoprotein (VSG) coat for survival in the mammalian bloodstream. High VSG expression within an expression site body (ESB) is mediated by RNA polymerase I (Pol I), which in other eukaryotes exclusively transcribes ribosomal RNA genes (rDNA). As T. brucei is reliant on Pol I for VSG transcription, we investigated Pol I transcription inhibitors for selective anti-trypanosomal activity. The Pol I inhibitors quarfloxin (CX-3543), CX-5461, and BMH-21 are currently under investigation for treating cancer, as rapidly dividing cancer cells are particularly dependent on high levels of Pol I transcription compared with nontransformed cells. In T. brucei all three Pol I inhibitors have IC50 concentrations for cell proliferation in the nanomolar range: quarfloxin (155 nM), CX-5461 (279 nM) or BMH-21 (134 nM) compared with IC50 concentrations in the MCF10A human breast epithelial cell line (4.44 μM, 6.89 μM or 460 nM, respectively). T. brucei was therefore 29-fold more sensitive to quarfloxin, 25-fold more sensitive to CX-5461 and 3.4-fold more sensitive to BMH-21. Cell death in T. brucei was due to rapid inhibition of Pol I transcription, as within 15 minutes treatment with the inhibitors rRNA precursor transcript was reduced 97-98% and VSG precursor transcript 91-94%. Incubation with Pol I transcription inhibitors also resulted in disintegration of the ESB as well as the nucleolus subnuclear structures, within one hour. Rapid ESB loss following the block in Pol I transcription argues that the ESB is a Pol I transcription nucleated structure, similar to the nucleolus. In addition to providing insight into Pol I transcription and ES control, Pol I transcription inhibitors potentially also provide new approaches to treat trypanosomiasis.
| Trypanosoma brucei is protected by an essential Variant Surface Glycoprotein (VSG) coat in the mammalian bloodstream. The active VSG gene is transcribed by RNA polymerase I (Pol I), which typically only transcribes rDNA. Pol I transcription inhibitors are under clinical trials for cancer chemotherapy. As T. brucei relies on Pol I for VSG transcription, we investigated its susceptibility to these drugs. We show that quarfloxin (CX-3543), CX-5461, and BMH-21 are effective against T. brucei at nanomolar concentrations. T. brucei death was due to rapid and specific inhibition of Pol I transcription. Incubation with Pol I transcription inhibitors also resulted in disappearance of Pol I subnuclear structures like the nucleolus and the VSG expression site body (ESB). Rapid ESB loss followed the Pol I transcription block, arguing that the ESB is nucleated by Pol I transcription. Pol I transcription inhibitors could therefore potentially function as novel drugs against trypanosomiasis.
| Human African Trypanosomiasis (HAT) or African Sleeping Sickness is endemic to sub-Saharan Africa, with distribution restricted by the tsetse fly insect vector [1]. Most of the HAT disease burden (98%) is the chronic form of the disease caused by Trypanosoma brucei gambiense. The remainder of the HAT cases are the acute form of the disease caused by Trypanosoma brucei rhodesiense [2]. Although there has been a progressive decline in the annual number of HAT cases, 1.8 million Africans are still thought to be living in high or very high risk areas, with 11.3 million people at moderate risk of contracting HAT [3]. Vaccines are ineffective against T. brucei, as it uses a highly sophisticated strategy of antigenic variation of a surface Variant Surface Glycoprotein (VSG) coat allowing effective escape from the mammalian immune system [4, 5]. As individual trypanosomes have thousands of different VSG genes and pseudogenes, and new chimeric VSG variants are continuously being generated [6], treatment of HAT has relied on drug treatment rather than immunisation.
Only a limited number of drugs are effective against HAT, of which many are toxic, expensive, or difficult to administer in the field [7]. Pentamidine is used against early stage infection of T. brucei gambiense, with eflornithine (DFMO) or NECT (nifurtimox-eflornithine combination treatment) also effective against later stages of T. brucei gambiense infection [8, 9]. Suramin has been used since 1922 against early stage T. brucei rhodesiense, with the highly toxic arsenical drug melarsoprol effective against the later stages of the disease once parasites have penetrated the blood-brain barrier [8]. A constant concern with this limited number of treatment options is human drug toxicity, as well as the development of parasite resistance. Drug resistant T. brucei strains are easily generated in the laboratory and have been found in the field [10, 11]. While there is a clear need for new treatments for HAT, the decreasing number of HAT cases has reduced the incentive to develop new drugs.
This increases the attractiveness of ‘repurposing’ drugs which have already undergone clinical trials for use against other human diseases [12–14]. The anti-trypanosomal drug eflornithine is a good example of drug “repurposing”. Originally developed as a drug against cancer, it was subsequently repurposed for use against T. brucei gambiense, as well as female hirsutism [15]. Similarly the drug tamoxifen, which is effective against estrogen receptor-positive breast cancer, also has efficacy against Leishmania [16, 17]. In addition to repurposing drugs, target repurposing can also be used to exploit libraries of small molecules developed against human target molecules. For example, an extensive panel of human kinase inhibitors are currently being investigated for their selective potential against essential T. brucei kinases [18].
Here, we investigate the efficacy of RNA polymerase I (Pol I) transcription inhibitors as treatment against T. brucei. Pol I transcribes the ribosomal RNA genes (rDNA), which accounts for up to 60% of transcription in proliferating cells [19]. Tumour cells are highly sensitive to disruption of Pol I transcription [20], while normal cells remain largely unaffected [21]. Tumour cell death upon Pol I transcription inhibition is not due to ribosome depletion, but due to cell “checkpoint” activation [20–22], explaining why in particular tumour cells undergoing uncontrolled cell proliferation are exquisitely sensitive to inhibition of Pol I transcription. Chemical inhibitors of Pol I are therefore currently being investigated for treatment against cancer [20, 23, 24].
African trypanosomes are unicellular eukaryotes which proliferate at high rates within the mammalian bloodstream. They have a particularly high dependency on Pol I transcription, as they transcribe the gene encoding their VSG coat from a Pol I transcribed VSG expression site (ES) transcription unit [25, 26]. VSG is highly essential in T. brucei both in vitro and in vivo, and perturbation of VSG synthesis in vivo results in very rapid trypanosome clearance within hours [27]. We therefore hypothesised that Pol I transcription inhibitors might represent useful drug leads against T. brucei.
A number of compounds have been shown to selectively impact on Pol I transcription (referred to here as Pol I inhibitors), including quarfloxin, CX-5461 and BMH-21. Quarfloxin has anti-cancer activity, and using a mouse xenograft model system, a large range of cancer cell lines were shown to be sensitive to it in vivo [28]. Quarfloxin was subsequently investigated for its therapeutic potential [28–30], and reached Phase II clinical trials before withdrawal due to problems with bioavailability [31]. The Pol I inhibitor CX-5461 was identified in a small molecule screen of compounds inhibiting transcription [32]. Similar to quarfloxin, CX-5461 prevented proliferation of a broad range of cancer cell lines in vitro or solid tumours in vivo [21, 33, 34]. This lead to its investigation as an anti-cancer therapy, and it is currently in Phase I clinical trials against breast cancer (Clincaltrials.gov NCT02719977) and hematologic cancer (ANZCTR ACTRN12613001061729)[23, 35]. The small molecule BMH-21 was identified in a small molecule screen targeting the p53 tumour suppressor pathway, and is also considered to have possible therapeutic potential against cancer [36, 37].
Here we show selective sensitivity of T. brucei compared with human breast epithelial or fibroblast cell lines for the Pol I inhibitors quarfloxin, CX-5461 and BMH-21. Trypanosome sensitivity for these drugs is within the nanomolar range, and at concentrations which are therapeutic against cancer. We show that these Pol I inhibitors specifically target Pol I transcription in T. brucei, as incubation results in very rapid and specific disappearance of Pol I derived RNA precursor transcripts. In addition, incubation with these compounds leads to Pol I subnuclear structures including the VSG expression site body (ESB) and the nucleolus disassembling. These chemical inhibitors demonstrate that the ESB, like the nucleolus, is a Pol I transcription-seeded subnuclear structure. In addition, these data argue that repurposing these potential cancer chemotherapy agents could potentially provide therapeutic potential against African trypanosomes.
No patient material was used. No animal experiments were performed. All experiments were performed in vitro with the established laboratory strain Trypanosoma brucei 427.
The laboratory bloodstream form strain Trypanosoma brucei 427 was cultured in vitro according to [38]. The T. brucei SM or “single marker” cell line [39] or the SM derived S16_221Pur cell line (containing a construct with a puromycin resistance gene inserted behind the active VSG221 ES promoter) were used for all in vitro cytotoxicity and proliferation assays, and were cultured in drug free media for at least 48 hours prior to treatment. The T. brucei S16221PuroGFP cell line was used for RNA precursor transcript analysis [40, 41]. Selection on puromycin maintained active transcription of the VSG221 ES in these cell lines. For the immunofluorescence analyses, the T. brucei TY-YFP-RPA2 cell line was generated through transfection of pEnT5H-Y:NLS:RPA2 (gift of the Gull lab) into T. brucei SM cells [42]. This resulted in a single RPA2 allele (second largest Pol I subunit) endogenously tagged with Yellow Fluorescence Protein (YFP) at the N-terminus. For analysis of VEX1 foci, the T. brucei S16_221Pur cell line was transfected with the pNATVEX1x12myc construct (gift of the Horn lab) [43]. This resulted in the addition of a 12x myc C-terminal epitope tag to VEX1.
RNA polymerase I inhibitors used were CX-5461 (Ross Hannan, ANU, Australia) [32], quarfloxin (CX-3543) (Adooq Bioscience) [28] or BMH-21 (Sigma) [36], with suramin (Sigma) used as a lethality control. Stock solutions of CX-5461 (10mM) were made up in 50 mM NaH2PO4 (pH 7.0). Quarfloxin (1 mM) and BMH-21 (1 mM) were dissolved in DMSO (dimethyl sulfoxide) (Sigma ≥99.9%). A 10 mM solution of suramin was made in nuclease free water immediately prior to use. All compounds were diluted directly in HMI-9 media immediately before use.
For proliferation assays, T. brucei SM was treated with inhibitors, and densities were determined using a Neubauer haemocytometer. Growth curves were repeated minimally in triplicate. For wash-out assays, T. brucei SM was treated with various concentrations of inhibitors. After two hours the cells were washed, resuspended in drug free HMI-9 medium, and cell proliferation was monitored using a haemocytometer. For the Alamar Blue in vitro cytotoxicity assay [44], 200 μl samples of T. brucei SM cells (2 x 103 cells ml-1) were plated in 96 well plates, and incubated for 72 hours with two fold serial dilutions of the inhibitors. At 72 hours, 20 μl of Resazurin (0.125 mg ml-1) (Sigma) was added, and the parasites were incubated for a further 18 hours. Fluorescence was measured using a Tecan infinite plate reader (excitation at 530 nm, emission at 585 nm). The change in fluorescence (minus the chemical only control) was plotted as a function of the concentration of chemical compound using the sigmoidal dose-response (variable slope) algorithm using GraphPad Prism version 5.
For the mammalian cell toxicity assays, either the spontaneously immortalised Michigan Cancer foundation (MCF10A) human breast epithelial cell line was used [45], or the BJ3 cell line, which is a human foreskin fibroblast cell line immortalized with the h-Tert catalytic subunit of telomerase [46]. Cells were seeded in 96-well plates, and after 48 hours each well was treated in quintuplicate using ten different concentrations of CX-5461, quarfloxin, BMH-21 or suramin. After incubation for 48 hours, the cells were imaged and cell confluence was calculated using an IncuCyte ZOOM (Essen Biosciences). Cell confluence was normalised to a drug vehicle only control which was set to 100%. An Alamar Blue assay was also performed on the MCF10A cell line in the presence of the different Pol I transcription inhibitors. Here the Alamar Blue readout was also normalised to the vehicle only control. Experiments were performed as three biological replicates with the exception of the Incucyte data for MCF10A in the presence of BMH-21 and doxycycline which were obtained in duplicate. The dose response curves and IC50s were calculated using GraphPad Prism.
Trypanosomes were fixed in 2% paraformaldehyde, permeabilised with 0.1% NP-40 for 5 minutes at room temperature before incubation for one hour with an antibody against the nucleolar marker L1C6 [47] or anti-myc tag antibody (clone 4A6, EMD Millipore). Slides were subsequently incubated with a goat anti-mouse secondary antibody coupled to Alexa-594 (Molecular Probes), before mounting in Vectashield containing DAPI (Vector Laboratories). Microscopy was performed using a Zeiss Imager.M1 microscope equipped with a Zeiss AxioCam MRm camera and Axio Vision Rel 4.8 software. A Z-stack of images was taken at 200 nm intervals, and the images from the individual channels were processed using Image J software. For quantitation of ESB signal using YFP, VEX1 signal using the anti-myc tag antibody or nucleolar signal using the L1C6 antibody, the exposure time and contrast was uniformly adjusted across all conditions. Quantification of ESB and nucleolar signal was carried out using compressed stacks of all appropriate channels in which the images acquired on the Z-axis were analysed. The ESB and nucleolar status was recorded for a total of approximately 100 cells in G1 (two biological replicates of approximately 50 cells) for each condition.
RNA transcript analysis was performed using quantitative reverse transcription PCR (qRT-PCR). Total RNA was isolated using the RNeasy kit (Qiagen) and DNase treated with TURBO DNA-free kit (Invitrogen). Reverse transcription was performed using 100 ng RNA as a template for cDNA synthesis using random hexamer primers (Promega) and the Omniscript RT kit (Qiagen). qPCR was performed on the 7500 Fast Real-Time PCR system (Life Technologies) using Brilliant II SYBR Green QPCR Low ROX Master Mix (Agilent Technologies). DNase treated RNA without reverse transcriptase was used as a control. The amplification conditions for each primer pair were optimised, and the primer sequences are listed in S1 Table. Transcript levels were normalised against actin mRNA, and plotted as relative change relative to the zero hour time point.
Where indicated, the data was analysed using the Student’s t-test (paired, two-tailed) (GraphPad Prism version 5). Data was considered ‘significant’ where P = 0.01-0.05 (*), ‘very significant’ where P = 0.001-0.01 (**) or ‘extremely significant’ where P = <0.001 (***).
Bloodstream form T. brucei utilises RNA polymerase I (Pol I) to transcribe rRNA within the nucleolus, as well as the active VSG ES in the ESB [48–50]. Blocking VSG synthesis triggers a precytokinesis arrest within one cell division, and very rapid trypanosome clearance in vivo [27]. The vital role of this multi-functional RNA polymerase I in African trypanosomes therefore makes it an appealing drug target. In light of this, we tested the RNA Pol I inhibitors quarfloxin (CX-3543), BMH-21 and CX-5461 for their efficacy against bloodstream form Trypanosoma brucei. We generated dose response curves using an Alamar Blue in vitro cytotoxicity assay, with the anti-trypanosomal agent suramin as a lethality control (Fig 1A–1D) [28, 32, 36, 51]. We found that T. brucei showed susceptibility to each of these Pol I inhibitors in the nanomolar range. T. brucei was most sensitive to BMH-21 with an IC50 for proliferation of 134 ± 8 nM (Table 1). Next most effective were quarfloxin with an IC50 of 155 ± 9 nM and CX-5461 with an IC50 of 279 ± 16 nM. As expected, T. brucei was susceptible to suramin with an IC50 of 63 ± 5 nM, comparable to the value of 53.0 nM found for T. b. rhodesiense using a similar Alamar Blue assay [51].
Toxicity of these compounds was also determined in mammalian cells using an Alamar Blue cytotoxicity assay in a spontaneously immortalised breast epithelial cell line (MCF10A) [45] (Fig 1E–1H). The compound doxycycline was used as a positive control for toxicity. These human epithelial cells were the least susceptible to CX-5461 (IC50 of 6.89 ± 4.83 μM), followed by quarfloxin (4.44 ± 3.29 μM) and BMH-21 (460 ± 80 nM). In parallel, proliferation of these human cells in the presence of Pol I inhibitors was monitored by assessing cell confluence using an IncuCyte ZOOM. This automated system allows monitoring of cells in real-time using live cell imaging, and produced comparable results to the Alamar Blue assay (S1 Fig) (Table 1).
In parallel, the BJ3 human foreskin fibroblast cell line that had been immortalised with hTert (the catalytic subunit of telomerase) was also tested for sensitivity to these compounds [46] (S2 Fig)(S3 Fig). The IC50 values corresponding to either growth arrest or cell death shown in S2 Fig were obtained by observing the cell images (S3 Fig) and determining the point where the specific dose curve plateaus at the higher concentrations. If the cells were morphologically sound, then the dose curve represented growth arrest of the cells. If the cells were dead, then the dose curve represented cell death. The IC50 values for cell death were used to calculate the selectivity index scores. This was done as the mammalian cells used in the dose response assays were intended to approximate normal quiescent human cells, hence the IC50 values for growth arrest are not relevant.
Similar to the breast cancer cells, these human fibroblasts were the least susceptible to CX-5461 (IC50 of 9.78 ± 0.79 μM), followed by quarfloxin (2.72 ± 0.17 μM) and BMH-21 (1.36 ± 0.22 μM) (Table 1). If one therefore compares the relative selectivity of these compounds for T. brucei, CX-5461 is the most effective compound, with T. brucei 25-35 fold more sensitive than human cells. T. brucei is 18-40 fold more sensitive to quarfloxin and 3-10 fold more sensitive to BMH-21. As expected, the trypanocide suramin showed minimal toxicity to mammalian cells (IC50 of 1636 ± 317 μM).
We next investigated the effect of these Pol I inhibitors on T. brucei proliferative growth over 48 hours (Fig 2). The drug concentrations of quarfloxin, BMH-21 or CX-5461 that resulted in suppression of proliferative growth in T. brucei were 300 nM, 300 nM and 1 μM respectively (Fig 2A–2C), with suramin serving as a positive control (Fig 2D). The impact of these Pol I inhibitors on T. brucei cell growth was rapid, and evident within two cell divisions.
Under therapeutic conditions effective drug concentrations in the blood typically fluctuate through time. We therefore tested the reversibility of these Pol I transcription inhibitors on the inhibition of T. brucei proliferation using wash-out assays. T. brucei was incubated with various concentrations of these inhibitors for two hours. Cells were subsequently washed and resuspended in drug-free medium, before cell proliferation was monitored (Fig 3). Incubation of T. brucei with quarfloxin and CX-5461 for two hours still resulted in significantly reduced proliferation at similar concentrations of drug (1 μM) even after the wash-out was performed. This indicates that quarfloxin and CX-5461 were working irreversibly. In contrast, wash-out of the BMH-21 inhibitor at all except for the highest concentration (3 μM) resulted in restored T. brucei growth. This indicates reversibility in how this compound is inhibiting trypanosome growth, which could potentially limit its use as a therapeutic agent.
As these Pol I inhibitors dramatically affected T. brucei proliferation, we next investigated if Pol I transcription of the rDNA or VSG ES was selectively inhibited (Fig 4A and 4B). The rDNA in T. brucei is transcribed as an approximately ten kilobase transcription unit, with the pre-rRNA precursor transcripts subsequently undergoing rapid processing reactions (Fig 4A) [52]. After incubation of T. brucei with the Pol I inhibitors, we determined the abundance of these unstable rRNA precursors using qRT-PCR as an indirect method of rRNA synthesis rates as performed previously [53]. Incubation with 1 μM quarfloxin, resulted in a dramatic reduction in pre-rRNA precursors within 15 minutes (Fig 4C), and pre-rRNA Precursor 1 transcript levels were reduced by 98.5 ± 0.66% (primer a), and rRNA Precursor 2 decreased by 91.3 ± 2.6% (primer b). Comparable degrees of repression of rDNA transcription were seen using primer pairs c and d, which detect both rRNA Precursors 2 and 3, where levels were also drastically reduced within fifteen minutes by 83.4 ± 2.1% or 89.5 ± 2.7% respectively.
The VSG221 ES is a highly active Pol I transcription unit in the bloodstream form of T. brucei. We investigated the repression of transcripts derived from regions of the ES which could be expected to be unstable, as they are encoded by intergenic regions or pseudogenes. We monitored the presence of a VSG221 precursor transcript using a primer pair 520 bp upstream of the VSG221 gene (Fig 4B). This VSG221 precursor transcript was reduced by 94.5 ± 1.9% after 15 minutes incubation with quarfloxin (Fig 4C). A single copy VSG pseudogene (ψ) is located approximately 4.7 kb upstream of the telomeric VSG221 in the approximately 60 kilobase VSG221 expression site [49]. Even when the VSG221 expression site is transcriptionally active, only very low levels of this VSG pseudogene transcript are present, presumably as a consequence of rapid degradation by nonsense mediated decay [54]. After incubation of T. brucei with quarfloxin for 15 minutes, levels of VSG pseudogene transcript were reduced by 92.3 ± 1.6%.
The rapid disappearance of these RNA precursors indicates that both the Pol I derived pre-rRNA as well as the Pol I derived VSG expression site encoded RNA precursor transcripts for VSG221 and the VSG pseudogene have very short half-lives of approximately three or four minutes. Precursor transcripts from the tubulin gene cluster are also highly unstable, and have been estimated to have a half-life of about one minute [55, 56]. However, treatment with quarfloxin for fifteen minutes did not result in a reduction in the levels of Pol II derived tubulin precursor transcript (129 ± 10.3% normal). Actin mRNA has been estimated to have a half-life of about 30 minutes in bloodstream form T. brucei [57]. We did not find that actin mRNA levels were affected by any of the inhibitors used in this study, and therefore the qPCR results were normalised using this transcript. These experiments therefore provide evidence for the selective inhibition of Pol I transcription in T. brucei with quarfloxin, which was highly significant (*** P = <0.001) in all cases.
Similar striking repression of Pol I transcription was observed after incubation of T. brucei with 1 μM BMH-21 for fifteen minutes (*** P = <0.001) (Fig 4D). In addition, incubation with 1 μM CX-5461 inhibitor for 15 minutes produced similar results to the other two inhibitors with specific knockdown of Pol I transcripts (*** P = <0.001 in all cases) (Fig 4E). To confirm that the selective reduction in Pol I derived precursor transcripts was not simply a consequence of cell lethality following treatment with the Pol I inhibitors, cells were treated for 15 minutes with 800 nM suramin (Fig 4F). Despite eventually causing cell death, this suramin treatment did not lead to reduced levels of any of these transcripts. In summary, these data show that incubation of T. brucei with all three Pol I inhibitors results in robust and rapid inhibition of Pol I transcription.
One of the hallmarks of blocking Pol I transcription is disassembly of the nucleolus as its integrity is dependent on rRNA synthesis [58]. We therefore monitored the disappearance of intact nucleoli as well as the ESB using immunofluorescence microscopy after incubation of T. brucei with the different Pol I inhibitors. The T. brucei RNA polymerase I complex contains at least twelve subunits, of which RPA2 is the second largest subunit [59, 60]. RPA2 has previously been endogenously tagged at the N-terminus with Yellow Fluorescence Protein (YFP) in bloodstream form T. brucei, allowing visualisation of the Pol I complex in both the nucleolus and the ESB [42]. In these experiments the ESB was visible in 67% ± 5% of nonmitotic cells [42]. We generated T. brucei expressing YFP tagged RPA2 from the endogenous RPA2 locus in the bloodstream form of the parasite, as described previously by Daniels et al [42]. As expected, there was no growth reduction in these cell lines, and the YFP-RPA2 subunit showed the expected subnuclear location for a Pol I subunit.
Using the DNA stain DAPI, the T. brucei nucleus can be visualised as a large stained focus, with a smaller focus indicating the kinetoplast (mitochondrial) DNA. The Pol I subunit RPA2 was visualised using YFP fluorescence, and the location of the nucleolus was confirmed using the L1C6 nucleolar marker [47]. As expected, YFP-RPA2 was enriched in the nucleolus in all of the cells, and in a small extra-nucleolar ESB structure in 64 ± 17% of the cells (Fig 5, Fig 6) as shown earlier [42]. We next investigated the effect of the Pol I inhibitors on the presence of the Pol I complex. As one of the hallmarks of blocking rRNA transcription is disintegration of the nucleolus [58], an antibody against the nucleolar marker protein L1C6 was used to visualise this. T. brucei cells in the G1 stage of the cell cycle have one nucleolus, and incubation of T. brucei with quarfloxin (3 μM) resulted in disintegration of this nucleolus into multiple L1C6 containing spots. In addition, there was loss of Pol I signal from the ESB, as well as from the nucleolar remnants (Fig 5). Incubation with 3 μM BMH-21 and CX-5461 also resulted in similar ESB loss and nucleolar disintegration (although not complete loss) in T. brucei (Fig 5). In contrast, incubation with lethal concentrations of the trypanocide suramin (800 nM) did not impact on distribution of Pol I or result in changes in ESB or nucleolar morphology.
We next quantitated the effect of different concentrations of quarfloxin on the presence of the ESB. Incubation with 300 nM quarfloxin for one hour resulted in a reduction in ESB positive cells from 64 ± 17% to 30.5 ± 5%, while in 1 μM quarfloxin only 7.5 ± 0.7% of the cells were ESB positive after one hour (Fig 6A). No cells contained an ESB after one hour of incubation with 3 μM quarfloxin. A similar analysis was performed with BMH-21, where very similar results were obtained (Fig 6B). Incubation with a range of concentrations of CX-5461 also resulted in a time and dose dependent decrease in presence of the ESB (Fig 6C). We used suramin in order to establish that the observed loss of the ESB in these experiments was not simply a consequence of subnuclear architecture falling apart in dying cells. Suramin rapidly kills T. brucei at a concentration of 800 nM (Fig 2). After one to three hours incubation with 800 nM suramin, there was only a marginal reduction in ESB positive cells (Fig 6D).
Recently, the VEX1 protein has been discovered to play a role in ES regulation in bloodstream form T. brucei, although how this operates mechanistically remains unclear [43]. In bloodstream form T. brucei VEX1 associates with the ESB, serving as a marker for this subnuclear structure. We investigated the effect of incubation with Pol I transcription inhibitors on VEX1 localisation. We incubated bloodstream form T. brucei, which had an endogenous copy of VEX1 epitope tagged with myc, with 3 μM quarfloxin, BMH-21 or CX-5461 (Fig 7). Similar to as observed with YFP tagged RPA2, VEX1 signal was drastically reduced after incubation of the cells with all three Pol I transcription inhibitors. In contrast, although incubation with the trypanocide suramin resulted in severe reduction in trypanosome growth no specific disintegration of the ESB was observed as visualised using epitope tagged VEX1 (Fig 7). These results support the observation that the ESB (as visualised using VEX1) is a transcription nucleated structure.
Similar to the Pol I ESB body, the structure of the Pol I nucleolar structures rapidly disintegrated (but did not disappear entirely) within one hour in cells incubated with these Pol I inhibitors. Incubation with 3 μM quarfloxin, BMH-21 and CX-5461 universally resulted in nucleoli fragmenting into multiple foci as monitored using the nucleolar marker L1C6 (Fig 8). Again, incubation with suramin did not lead to significant changes in structure of the nucleoli.
We do not know if these Pol I transcription inhibitors operate mechanistically in a similar fashion in T. brucei compared with mammalian cells. In the case of BMH-21, in human cells BMH-21 blocks Pol I transcription, resulting in degradation of the Pol I large subunit [37]. We investigated if incubation of T. brucei with increasing concentrations of BMH-21 resulted in similar degradation of the T. brucei RPA2 Pol I subunit. However even after incubation of T. brucei TY-YFP-RPA2 with 3 μM BMH-21 for one hour, we did not see evidence for degradation of the RPA2 Pol I subunit (S4 Fig). As BMH-21 intercalates with DNA, it is possible that this DNA binding activity disrupts Pol I transcription in T. brucei in a different fashion to mammals.
Collectively, these data show that bloodstream form T. brucei is highly sensitive to Pol I inhibitors compared with the human host. Moreover, Pol I inhibitors rapidly and selectively inhibit Pol I transcription in T. brucei, leading to loss of the ESB and fragmentation of the nucleolus. Together, these experiments provide evidence that the ESB within T. brucei is fundamentally a Pol I transcription nucleated structure.
Rapidly proliferating cells require high levels of transcription of the Pol I transcribed rDNA, and chemical inhibitors of this Pol I transcription are effective against cancer [20, 21, 28, 33, 37]. Pol I not only transcribes the rDNA in bloodstream form T. brucei, but also the highly essential VSG gene. T. brucei is therefore particularly reliant on high levels of Pol I transcription, as even minor perturbations of VSG synthesis would be disastrous for parasite survival in vivo [27].
Here, we show that three established mammalian Pol I transcription inhibitors including CX-5461, quarfloxin and BMH-1 are selectively toxic for T. brucei. The most selective Pol I inhibitors against T. brucei are CX-5461 with an IC50 of 279 nM and quarfloxin with an IC50 of 155 nM. T. brucei is therefore 25-35 fold more susceptible to CX-5461 or 18-40 fold more susceptible to quarfloxin compared with the spontaneously immortalised MCF10A human breast epithelial cell line or the immortalised BJ3 human fibroblast cell line. This differential sensitivity could potentially provide a “therapeutic window” to treat a human trypanosome infection using these compounds. Although in T. brucei BMH-21 has a lower IC50 concentration (134 nM) compared with CX-5461 or quarfloxin, BMH-21 is also more toxic to human cells and there is only 3-10 fold selectivity for T. brucei. The selective toxicity of 25-35 fold for CX-5461 and 18-40 fold for quarfloxin in T. brucei compared with human cells meets DNDI criteria for drug screening for Kinetoplastid diseases where the minimum selectivity index is ≥ 10 (Ioset et al at: www.dndi.org/2009/media-centre/scientific articles). This makes Pol I inhibitors intriguing drug leads. However, further studies including structure-activity relationship (SAR) studies would clearly be required to optimise both their potency and their selectivity against T. brucei.
We show that all three Pol I inhibitors selectively and rapidly block Pol I transcription, with rRNA precursor transcripts in some cases reduced by 98%, and VSG expression site derived precursor transcripts reduced by 94%. In contrast, we saw no reduction in levels of precursor transcript from the Pol II transcribed tubulin transcription unit. Incubation of T. brucei with these three Pol I inhibitors lead to disappearance of Pol I signal within the nuclei within one hour. In addition, within one hour there was disappearance of the ESB (as visualised using RPA2 or VEX1) and rapid disintegration of the nucleolus, which shattered into small dots as visualised using an antibody for the nucleolar protein L1C6.
It has been demonstrated that nucleoli are ‘Pol I transcription-seeded’ structures which require active transcription for maintenance of their structure [58, 61]. This transcription needs to be specifically mediated by Pol I, as nucleoli do not form if rDNA is transcribed by Pol II [62]. The fact that the nucleoli in T. brucei rapidly disassemble after incubation of the cells with specific Pol I inhibitors, argues that nucleoli are also Pol I transcription-nucleated structures in African trypanosomes. As well as disintegration of the nucleoli, incubation with all three Pol I inhibitors also leads to a very striking and rapid disappearance of the ESB. After one hour incubation with 3 μM BMH-21 or CX-5461, only 1 ± 1.4% or 4.5 ± 0.7% of the cells respectively still had an ESB. After a one hour incubation with 3 μM quarfloxin, none of the cells had an ESB. Similar rapid disappearance of the ESB in the presence of Pol I inhibitors was observed if the ESB was visualised using VEX1. These results would therefore suggest that the T. brucei ESB similar to the nucleolus, requires active Pol I transcription for its maintenance.
The active VSG gene is transcribed at a high rate from a single active telomeric VSG expression site locus, allowing the cell to produce vast amounts of VSG transcript (approximately 10% total mRNA) from a single copy VSG gene. This highly efficient VSG transcription, as well as rapid processing of the abundant VSG mRNA is presumably facilitated by the presence of the VSG expression site within the ESB subnuclear structure. The ESB has been proposed to function as a specialised factory assembled on the active VSG expression site, containing high concentrations of Pol I transcription factors as well as the RNA processing machinery necessary for efficient expression of VSG [50].
It has been postulated that the ESB is a coherent architectural structure, rather than simply a consequence of resident Pol I on the active ES DNA. This is based on the observation that removal of the DNA in DNase I treated methanol fixed cells, still resulted in concentrated localisation of Pol I in an ESB [50]. However, a potential complication in interpreting these experiments could be a possible reduction in motility in Pol I and associated transcription complexes after the methanol precipitation needed for fixation, even in the absence of DNA. Is the ESB a ‘Pol I transcription-seeded’ structure as has been proposed for the nucleolus? Our results would agree with this, and suggest that the ESB is a subnuclear structure which is nucleated around sites of active Pol I transcription, as the ESB rapidly disappears after Pol I transcription is blocked.
The ESB shares some, but not all components with nucleoli. Key Pol I transcription factors including CITFA-7 and the architectural chromatin protein TDP are located in both Pol I subnuclear structures [53, 63]. However, other nucleolar components including fibrillarin as well as the L1C6 nucleolar protein are found in the nucleolus but not the ESB [47, 50]. The only factor that has yet been proposed to be ESB specific is VEX1, although its mode of action in VSG control is yet to be determined [43]. Our experiments using Pol I inhibitors show that these two types of Pol I transcription complexes at either the rDNA or the active ES are inhibited in a similar fashion, although there are clearly significant differences in their control.
It is striking that all three of these mammalian Pol I inhibitors appear to specifically inhibit Pol I transcription in T. brucei despite having differences in their modes of action in mammalian cells. BMH-21 is a DNA intercalator, which binds GC-rich sequences present at high frequency in ribosomal DNA genes [37, 64]. In human cells BMH-21 blocks Pol I transcription elongation, leading to proteasome dependent destruction of RPA194 (largest Pol I subunit), which is correlated with cancer cell killing [37]. In T. brucei, although there is also rapid disappearance of Pol I subnuclear structures like the nucleoli and the ESB after one hour incubation with 1 μM BMH-21, we did not see evidence for significant reduction in levels of YFP2-RPA2 as monitored by Western blot analysis (S4 Fig). However, despite the presence of unreduced amounts of YFP-RPA2 protein, intercalation of BMH-21 with DNA could still potentially result in disassociation of Pol I complex from the DNA.
In mammalian cells, CX-5461 inhibits Pol I mediated initiation of transcription at the rDNA through interference of SL1 transcription factor binding to the rDNA promoter [33]. T. brucei does not have an obvious SL1 homologue, and it is still unclear which protein fulfils its function in trypanosomes. It is therefore, not straightforward to test if the mode of action of CX-5461 in T. brucei is indeed similar to that in mammalian cells. Quarfloxin is a fluoroquinolone which intercalates with DNA, and is thought to preferentially bind G-rich stretches forming G-quadruplexes within the rDNA and at telomeres [29, 30, 65]. Quarfloxin accumulates in the nucleoli, where it selectively interferes with rRNA synthesis at the level of Pol I elongation, leading to reduced levels of rRNA precursor transcript [28]. Although T. brucei rDNA has G-rich regions, it is unclear if these indeed form G-quadruplexes. T. brucei telomere repeats have the same sequence (GGGTTA) as those in mammalian cells, and could therefore also be expected to form G-quadruplex structures [65]. As well as interfering with Pol I transcription of the rDNA in T. brucei, it is possible that quarfloxin also interferes with telomere function as well. A similarity between BMH-21 and quarfloxin is that they intercalate with DNA, particularly in G-rich regions, and it is presumably this feature that is behind their efficacy in T. brucei. These Pol I inhibitors could therefore provide a new tool for specifically blocking Pol I transcription in African trypanosomes, which should help us dissect how the regulated Pol I mediated transcription of the VSG variant antigen genes is controlled.
Pol I inhibitors are currently being investigated for their suitability for treating cancer. The inhibitor quarfloxin progressed to Phase II clinical trials, but was withdrawn due to problems regarding bioavailability [31]. However, CX-5461 is still under investigation for cancer treatment, and is currently in Phase I clinical trials against breast cancer (Clincaltrials.gov NCT02719977) and haematologic cancer (ACTRN12613001061729). Could these Pol I inhibitors indeed be used to treat human African trypanosomiasis? Pol I is of particular importance for bloodstream form T. brucei as very high levels of transcription of the single active VSG gene are essential to provide the huge amount of VSG protein required to coat this extracellular parasite. As blocking VSG synthesis results in very rapid parasite clearance in infected mice [27], we would predict that targeting VSG transcription should allow us to target this particularly vulnerable feature of the parasite. It is likely that even minor perturbation of VSG synthesis would result in rapid trypanosome clearance in vivo. This could potentially allow the repurposing of a potential anti-cancer agent for treatment of a tropical parasite.
As clinical trials are the main cost in developing a pharmaceutical treatment, one could envisage that repurposing an existing drug could provide an economical means to increase the number of substances that can be used to treat trypanosomiasis [13, 14]. This is particularly an issue as the current relatively small number of annual human African trypanosomiasis cases make the cost of developing new drugs problematic. However, despite some of the complications that would need to be considered regarding potential toxicity of these compounds, our data indicate that these Pol I inhibitors could potentially function as new chemical weapons against human African trypanosomiasis.
|
10.1371/journal.pmed.1002389 | Self-monitoring of blood pressure in hypertension: A systematic review and individual patient data meta-analysis | Self-monitoring of blood pressure (BP) appears to reduce BP in hypertension but important questions remain regarding effective implementation and which groups may benefit most. This individual patient data (IPD) meta-analysis was performed to better understand the effectiveness of BP self-monitoring to lower BP and control hypertension.
Medline, Embase, and the Cochrane Library were searched for randomised trials comparing self-monitoring to no self-monitoring in hypertensive patients (June 2016). Two reviewers independently assessed articles for eligibility and the authors of eligible trials were approached requesting IPD. Of 2,846 articles in the initial search, 36 were eligible. IPD were provided from 25 trials, including 1 unpublished study. Data for the primary outcomes—change in mean clinic or ambulatory BP and proportion controlled below target at 12 months—were available from 15/19 possible studies (7,138/8,292 [86%] of randomised participants). Overall, self-monitoring was associated with reduced clinic systolic blood pressure (sBP) compared to usual care at 12 months (−3.2 mmHg, [95% CI −4.9, −1.6 mmHg]). However, this effect was strongly influenced by the intensity of co-intervention ranging from no effect with self-monitoring alone (−1.0 mmHg [−3.3, 1.2]), to a 6.1 mmHg (−9.0, −3.2) reduction when monitoring was combined with intensive support. Self-monitoring was most effective in those with fewer antihypertensive medications and higher baseline sBP up to 170 mmHg. No differences in efficacy were seen by sex or by most comorbidities. Ambulatory BP data at 12 months were available from 4 trials (1,478 patients), which assessed self-monitoring with little or no co-intervention. There was no association between self-monitoring and either lower clinic or ambulatory sBP in this group (clinic −0.2 mmHg [−2.2, 1.8]; ambulatory 1.1 mmHg [−0.3, 2.5]). Results for diastolic blood pressure (dBP) were similar. The main limitation of this work was that significant heterogeneity remained. This was at least in part due to different inclusion criteria, self-monitoring regimes, and target BPs in included studies.
Self-monitoring alone is not associated with lower BP or better control, but in conjunction with co-interventions (including systematic medication titration by doctors, pharmacists, or patients; education; or lifestyle counselling) leads to clinically significant BP reduction which persists for at least 12 months. The implementation of self-monitoring in hypertension should be accompanied by such co-interventions.
| Self-monitoring of BP appears to lower BP in people with hypertension, over and above usual care.
Implementation of self-monitoring has been inconsistent, perhaps because important evidence gaps remain regarding how best to use it and for which patient groups.
To better understand the effect of self-monitoring on BP lowering and BP control.
Specifically, to examine the effect of self-monitoring in combination with various co-interventions, and in different groups of patients.
We undertook a systematic literature search to identify all studies that included self-monitoring of BP in people with high BP.
For studies published since the year 2000 with at least 6 months of follow-up data and at least 100 patients, we contacted authors to gain access to the original data collected for each individual patient (15/19 studies with the primary outcome provided data: 7,138/8,292 randomised participants).
We then used these data to perform IPD meta-analysis to evaluate the effect of self-monitoring on BP levels and in the control of hypertension using 1 year of follow-up as our primary end point.
We predefined levels of intensity of co-intervention and subgroups of patients for further analysis.
Self-monitoring worked best when combined with more intensive interventions such as self-management, systematic medication titration, or lifestyle counselling, but had little or no effect on its own.
Self-monitoring was most effective in those with fewer antihypertensive medications and higher baseline sBP up to 170 mmHg. No differences in efficacy were seen by sex or by most comorbidities.
Self-monitoring can be recommended to lower BP when combined with co-interventions involving individually tailored support.
Self-monitoring alone does not seem to lower BP but may be useful for other reasons including engaging with patients or reducing clinician workload.
| Treatment of hypertension results in significant reductions in risk of subsequent cardiovascular disease [1,2]. Despite strong evidence for such treatment, international epidemiological studies suggest that many people remain suboptimally controlled [3]. Self-monitoring of blood pressure (BP), where individuals measure their own blood pressure, usually in a home environment, can improve BP control and is an increasingly common part of hypertension management. Such monitoring can be accompanied by additional support such as from a nurse or pharmacist [4].
Self-monitoring is well tolerated by patients and has been shown to be a better predictor of end organ damage than clinic measurement [5–8]. This is despite potential issues with quality control of self-measurement such as poor technique or withholding of results [9,10]. The latter can be reduced to an extent by the use of telemonitoring [11].
Previous meta-analyses have shown that self-monitoring reduces clinic BP by a small but significant amount compared to conventional care (around 4/1.5 mmHg) [4,12–14]. Analysis by Bray and colleagues suggested that when self-monitoring was accompanied by a co-intervention, participants were more likely to meet target BP, but it remains unclear which interventions are most effective and what specific populations (if any) should be targeted [14].
The aim of this work was therefore to use individual patient data (IPD) from relevant trials to assess the effectiveness of BP self-monitoring on BP reduction and hypertension control, evaluating how best to utilise self-monitoring of BP and to determine which subpopulation is most likely to benefit.
This study systematically reviewed the existing literature to identify randomised trials examining the efficacy of self-monitoring of BP compared to control. Authors of all eligible trials were approached for access to IPD. A protocol with detailed methods has been published previously [15]. The methods used are summarised below.
Medline, Embase, and the Cochrane Library were searched for trials using BP self-monitoring in hypertensive patients (S2 Fig; search date June 2016).
Two reviewers (RM and KT) independently assessed the articles for eligibility and inclusion; disagreements were resolved by discussion. Randomised trials were eligible that recruited patients with hypertension being managed as outpatients using an intervention that included self-measurement of BP. Self-monitoring had to be without medical professional input (i.e., by patient with or without carer support) and using a validated monitor, with or without other co-interventions, and where a comparator group had no organised self-measurement of BP. Included studies were required to have involved at least 100 patients, followed up for at least 24 weeks, and to have been published since 2000. This was to ensure that self-monitoring equipment was likely to be relevant to contemporary medical management (i.e., automated oscillometric monitors). Relevant outcomes were systolic blood pressure (sBP) and/or diastolic blood pressure (dBP) measured in clinic, by researcher or by ambulatory measurement, and achievement of BP control.
Authors whose trials met the inclusion criteria were approached for provision of IPD including demographic details, comorbidities, antihypertensive medications, lifestyle factors, and BP end points (clinic and/or ambulatory). Study-level data were extracted where available from published articles and checked by the original authors. In particular, any co-interventions were carefully documented and prospectively allocated to 1 of 4 levels of interventional support based on a previous classification [4] (S1 Table).
Study quality was assessed in terms of potential bias from randomisation, blinding, outcome assessment, and method of analysis using an adaptation of the Cochrane tool [16].
Original data were kept on a secure server and assembled in a consistent format for all trials. Three researchers (KT, RM, and JS) cross-checked trial details, summary measures, major outcomes, and definitions against published articles. Any apparent inconsistencies were checked with the original trial authors. Overall ethical approval was not required as this study does not include identifiable data; collaborating groups gained individual approval where required for data sharing.
A 2-stage IPD meta-analysis was conducted using linear regression for continuous outcomes and logistic regression for proportions, aggregated across studies by random-effects inverse variance methods. Intention-to-treat comparisons of outcomes between the self-monitoring and comparator arms were summarised with forest plots using the I-squared (I2) statistic for heterogeneity. Regression models were adjusted for age, sex, baseline clinic BP, and diabetic status (the latter due to the lower BP target generally used in a diabetic population).
The primary outcomes were change in sBP and dBP at 12 months and likelihood of uncontrolled BP below target at 12 months (control as defined by each trial). Analyses are reported in subgroups, by pre-specified level of self-monitoring intervention as described in Table 1 and in the published protocol [15].
Subgroup analyses examined the effect of self-monitoring on BP mean and control by age, sex, baseline sBP, the presence and number of antihypertensive medications prescribed, and comorbidities (myocardial infarction [MI], stroke, diabetes mellitus [DM], chronic kidney disease [CKD], and obesity [defined as a body mass index (BMI) ≥ 30 kg/m2]). All subgroup analyses were adjusted for age, sex, baseline clinic BP, level of intervention, and individual study (contributing to each analysis).
Sensitivity analyses included incorporation of aggregate data from studies that did not contribute IPD [17–23], exclusion of individual patients for whom a lower home BP target was not used (due to study design or the presence of comorbidities such as diabetes) [24–27], influence of BP inclusion criteria (clinic or ambulatory) from ambulatory outcome studies, different assumptions regarding BP of patients lost to follow-up (controlled or uncontrolled), and influence of adjusting for medication changes (in those studies which recorded changes in medication). Finally, the influence of each study on the overall results was assessed using an influence analysis. Egger’s test for funnel plot asymmetry was applied to consider possible publication bias (S21 Fig) [28].
There were no deviations from the protocol [15]. Five post-hoc analyses were undertaken: firstly, an additional subgroup analysis was carried out (resistant hypertension [defined as BP > 140/90 mmHg and 3 medications at baseline or any BP level and 4 or more medications at baseline]); secondly, the distribution of baseline antihypertensive medications was compared in patients with and without a history of stroke using Pearson’s chi-squared; thirdly, the effectiveness of self-monitoring in stroke was assessed controlling for the number of baseline medications; fourthly, the influence of blinding was assessed; and finally, sBP was plotted against medication changes.
All analyses were conducted using STATA version 13.1 (MP parallel edition, StataCorp, College Station, Texas, USA), using the ipdmetan package [29]. Data are presented as proportions of the total study population, means with standard deviation or relative risk (RR) with 95% confidence intervals unless otherwise stated.
Of 2,846 unique studies from the combined searches, 132 were assessed in full and 36 studies were deemed potentially eligible (S1 Fig). One study which would otherwise have been eligible was excluded because the comparator group used ambulatory monitoring to guide treatment, a control intervention that had not been anticipated in the protocol but which was not comparable to any other included studies [30]. Of the 36 potentially eligible studies, 19 had published data at 12 months, the primary outcome. Authors from 24 of the potentially eligible studies provided IPD, with 1 group submitting additional data from an unpublished study. These 25 studies were published from 2005–2014, were conducted in North America and Europe (11 United States; 6 United Kingdom; 3 Italy; 1 each from the Netherlands, Australia, Spain, Finland, and Canada), and included a wide range of self-monitoring protocols, co-interventions, and populations (Table 1) [23–27,31–48]. Authors from the remaining 12 studies were either unable to provide IPD (2 studies) or did not respond to the request for data (10 studies). Four studies which followed up patients for 12 months did not provide IPD, so that data for the primary outcome were available from 15/19 studies (7,138/8,292 [86%], of potential participants) (S2 Table) [17,18,22,49]. A total of 838 patients (12%) were lost to follow-up across all included studies, and a further 227 patients from the potentially available studies were lost to follow-up, leaving 6,300/7,227 patients (87%) for inclusion in the final analysis of the primary outcome (12 months follow-up).
Overall, the information from the included trials was judged to be at low risk of bias: most studies used computerised generation of randomisation sequences (23/25, 92%), appropriate allocation concealment (24/25, 96%), and all used an intention-to-treat approach with either multiple imputation for missing data or analysis of complete cases. Most studies (19/25, 76%) followed up more than 80% of participants, but only 12/25 (48%) used blinded assessment of outcome (S3 Table). An influence analysis assessed the impact of each individual study on the overall results. Included studies were predominantly publically funded (S4 Table).
Overall, self-monitoring was associated with reduced clinic sBP between baseline and 12-months follow-up compared to usual care (systolic −3.2 mmHg, 95% CI −4.9 to −1.6 mmHg) (Fig 1). Significant heterogeneity was present between studies: I2 = 76%, P < 0.001. Self-monitoring was also associated with reduced dBP at 12-months follow-up (diastolic −1.5 mmHg, 95% CI −2.2 to −0.8 mmHg) and significant heterogeneity remained (I2 = 62%, P < 0.001) (Fig 2). Similar reductions in BP were seen after 6-months follow-up, but the point estimates after 18-months follow-up were smaller, albeit from only 5 studies (S3, S4, S6 and S7 Figs).
Clinic BP control was improved at 12-months follow-up (RR of uncontrolled BP 0.7 [95% CI 0.56 to 0.86]) again with significant heterogeneity between groups (Fig 3). Similar results were seen at 6 and 18 months (S5 and S8 Figs, respectively).
The reductions in clinic sBP varied with different levels of intervention: level 1 (with no co-intervention) −1.0 mmHg, [95% CI −3.3 to 1.2 mmHg]; level 4 (personal support throughout the trial) −6.1 mmHg, [95% CI −9.0 to −3.2 mmHg] (Fig 1) (heterogeneity in outcome between different levels of intervention P < 0.001). Within predefined categories of intensity of co-intervention, significant heterogeneity remained, apart from within level 2.
A similar pattern of reductions was seen in dBP: level 1 (with no co-intervention) −1.1 mmHg, [95% CI −2.4 to 0.2 mmHg]; level 4 (personal support throughout the trial) −2.3 mmHg, [95% CI −4.0 to −0.6 mmHg] (Fig 2) (heterogeneity in outcome between different levels of intervention P < 0.001). Within predefined categories of intensity of co-intervention, significant heterogeneity remained in levels 1 and 4.
BP control (defined according to individual study targets, Table 1) at 12 months also differed by level of intensity. The RR of having uncontrolled BP with a self-monitoring intervention at 12 months varied from level 1 (RR 1.0, 95% CI 0.7 to 1.4) to level 4 (RR 0.4, 95% CI 0.3 to 0.6) (Fig 3) (heterogeneity between levels of intervention P < 0.001). Heterogeneity within levels of intervention in this analysis was low for levels 2 and 4 of co-intervention, although the I2 remained above 50% for level 1. Similar results were seen at 6-months follow-up (21 studies) and at 18-months follow-up (5 studies) (S5 and S8 Figs, respectively).
Four studies had data at 12 months using ambulatory BP as the outcome (1,478 participants); these were studies with no co-intervention (level 1; n = 3) or automated feedback only (level 2; n = 1). No change was seen in ambulatory sBP associated with self-monitoring (1.1 mmHg [−0.3, 2.5]) (Fig 4) or ambulatory dBP (0.8 mmHg [−0.2, 1.9]), and there was no significant heterogeneity between studies in either case (Fig 5). At 6 months, data were available for 5 studies with no difference seen in ambulatory sBP (−1.0 mmHg [−2.8, 0.9]) or dBP (−0.4 mmHg [−1.6, 0.8]) (S9 and S10 Figs, respectively). The additional study, which used a level 3–intensity intervention, increased heterogeneity as it had a significant outcome.
No ambulatory data were available at 18 months.
Subgroup analyses using data from 12-months follow-up showed little difference in either reduction of systolic or diastolic clinic BP or likelihood of uncontrolled BP depending on history of MI or presence of CKD or diabetes (Figs 6, 7 and 8) (I2 ≤ 20% for all subgroups).
However, a history of stroke was associated with a reduced effectiveness of self-monitoring in terms of clinic sBP lowering (I2 = 77%, P = 0.04), though this difference was not observed for dBP or maintained in the likelihood of control analysis (RR I2 = 42%, P = 0.19). Post-hoc analyses showed that the distribution of number of medications between stroke and non-stroke patients was similar (S5 Table), and adjusting for baseline medication use did not explain the lack of effectiveness in patients with stroke. There was moderate heterogeneity between age groups for the effect of self-monitoring on systolic and diastolic clinic BP (I2 = 31%, P = 0.20 and I2 = 33, P = 0.19, respectively) but not in the likelihood of uncontrolled BP (I2 = 0.0%, P = 0.60). Considering the effect of obesity, there was no difference in the effect on systolic clinic BP reduction (I2 = 0, P = 0.72) but there was some evidence of heterogeneity of effect for dBP (I2 = 63, P = 0.10) and the risk of uncontrolled BP (I2 = 61%, P = 0.11).
Fewer baseline antihypertensive medications were associated with larger reductions of BP and better control (Figs 6–8). Post-hoc analyses, comparing those with resistant hypertension to those without, suggested that self-monitoring was less effective at achieving BP control in the former (RR of uncontrolled BP = 0.62, 95% CI 0.54–0.71 [non-resistant hypertension] versus RR of uncontrolled BP = 0.94, 95% CI 0.65–1.36 [resistant hypertension], I2 = 76%, P = 0.04). Similarly, the post-hoc analysis plotting change in sBP against medication changes was consistent with the hypothesis that self-monitoring interventions resulted in BP decreases via increases in prescribed medication (S22 Fig).
Inclusion of aggregate data for clinic BP at 12 months from the 4 eligible studies that did not contribute IPD (S2 Table) and exclusion of studies that did not use a lower home BP threshold did not materially change the results (S11 and S12 Figs). The exclusion of studies that randomised on the basis of ambulatory BP monitoring (ABPM) or studies that randomised on clinic BP did not change the impact of clinic or ambulatory measurement of sBP at 12 months (S13 and S14 Figs). Assuming patients lost to follow-up had uncontrolled BP attenuated the results, whereas assuming that they had controlled BP accentuated them (S15 and S16 Figs, respectively). Exclusion of patients with controlled BP at baseline also accentuated the results (S17 Fig). A post-hoc comparison of studies with blinded outcome (2,829 patients) versus unblinded (3,257 patients) showed that blinding was associated with a reduced point estimate for the change in sBP at 12 months in those studies examining higher-level interventions, albeit with overlapping confidence intervals (level 1 & 2 intervention studies: −1.51, 95% CI −4.06 to 1.04 [blinded] versus −0.83, 95% CI −2.38 to 0.73 [unblinded]; level 3 & 4 intervention studies: −4.67, 95% CI −7.51 to −1.84 [blinded] versus −6.16, 95% CI −9.36 to −2.95 [unblinded]).
Where studies had measured changes in antihypertensive medication over time, there was evidence of attenuation of the change in sBP when the analysis was adjusted for change in medication (S18 and S19 Figs). The influence analysis did not suggest that any one study was materially influencing the results (S20 Fig and Egger’s test found no evidence of asymmetry in the funnel plot (P = 0.9, S21 Fig).
Using IPD from 25 studies totalling 10,487 patients, this meta-analysis provides strong evidence that the degree of BP lowering is related to the intensity of the co-intervention (i.e., additional support) combined with self-monitoring, with little or no effect from self-monitoring alone.
These results held whether systolic or diastolic clinic BP or clinic BP control were assessed and were consistent at both 6 and 12 months. No data were available from studies with intensive co-interventions which used ambulatory BP monitoring to measure outcomes at 12 months or longer, and those with little or no co-intervention showed similar effects to the clinic BP data (no effect in either case). There was a suspicion of attenuation of the effect of self-monitoring in the few studies to date that have followed up patients for longer than 1 year but data were sparse. Future research might be directed towards longer studies with ambulatory BP measurement (or other measurements to reduce the white coat effect) for outcomes. Self-monitoring appeared most effective at lowering BP in people on fewer BP medications at baseline, and there was a suggestion of a greater effect with higher BP—provided BP was not 170 mmHg or above. Analyses considering those with apparent resistant hypertension at baseline suggested that self-monitoring works less well in this group, but this analysis was not prespecified, could not take into account dose of antihypertensive medication, and should be interpreted with caution. In terms of comorbidities, the effects of self-monitoring were similar whether or not an individual had a history of MI, diabetes, or CKD. In people with previous stroke, there may be a reduced effect of self-monitoring but this did not reflect more intensive treatment prior to randomisation.
To our knowledge, this is the first analysis of self-monitoring of BP to use IPD from a wide range of self-monitoring trials from North America, Australia, and Europe and including both specialist and primary care settings. IPD allowed for standardised adjustment of outcomes and sufficient power to detect differences between subgroups.
An important issue in IPD analysis is selection of studies. The BP—SMART collaboration has gained access to a large number of datasets; nevertheless, some studies were not available due to unavailability of data or lack of response. Despite this, only 4 studies eligible for the primary outcome (14% of available patient data) were unable to provide data, and sensitivity analyses suggested no material change in the results when the published aggregate data from these studies were included.
Quality of included studies was adequate in terms of randomisation sequences, appropriate allocation concealment, and analyses. Follow-up was high for most studies but only half used blinded assessment of outcome. However, a post-hoc sensitivity analysis showed no difference in results from blinding, perhaps because in most studies BP was assessed using automated monitors reducing the chance of bias.
Despite the use of IPD and the division of studies into subgroups, significant heterogeneity remained, which limited the ability to do meta-analysis. However, this does not negate the conclusion that the evidence for both BP reduction and control is stronger for higher-intensity interventions and weak for self-monitoring alone. The hypothesis that effect would vary with level of intervention was prespecified and the categorisation of studies into 4 levels of intervention was agreed to by all study investigators before results were available.
Whilst all included studies compared self-monitoring of BP to control groups without self-monitoring, inevitably different investigators used different protocols and therefore studies differed in inclusion criteria, self-monitoring regime, and target BPs. These issues could at least in part explain the remaining heterogeneity between studies. The exclusion of studies which did not use lower BP targets for self-monitored BP did not change the results, but even IPD analysis is unable to take differences between studies entirely into account and this may be reflected in the heterogeneity which remained. Significance tests should be interpreted with caution when, as in Figs 6 and 7, multiple coequal exposures are under test; however, the 3 P values ≤ 5% for heterogeneity across these 18 tests are unlikely to be all due to chance alone and the tests were prespecified.
Most outcome data were based on clinic measurement of BP, which is what was used by the majority of trials of outcome of hypertension treatment [1]. Ambulatory monitoring might reduce any attenuation to the white coat effect from repeated habituation to measurement but, whilst 6 studies used ambulatory BP as an outcome [25,32,33,43,45] (including 1 unpublished study), all but 1 of these used less intensive or no co-interventions. The single intensive study with an ambulatory outcome had data to 6 months and a positive result, whereas the remaining 4 studies showed no impact on ambulatory BP in common with the pooled results for clinic BP. Other studies have used multiple automated BP measurements in the clinic to assess habituation and have found no evidence that the BP differences are removed when the white coat effect is reduced, though further studies examining the effects of self-monitoring with intensive co-interventions on outcomes which reduce white coat effects are arguably needed [36] [35].
Even with IPD, issues such as loss to follow-up may be important. Included studies had rates of follow-up between 58% and 98% at 12 months with most studies following-up around 90%. In the main analysis, formal methods for handling missing data were not used since methods for imputation in IPD meta-analysis are in their infancy; however, the impact of each individual study on the overall results as assessed by the influence analysis suggests that factors such as differential follow-up between studies were unlikely to have affected the results [52].
The outcomes included in this review are all related to BP. Whilst BP is directly related to stroke and coronary heart disease risk, it is nevertheless an intermediate outcome. Ideally, such hard outcomes would be directly measured in trials. However, because of relatively short follow-up and small numbers of participants, few included individual trials did so.
There have been previous systematic reviews of trials of self-monitoring [4,13,14,53], including those focussing on specific outcomes such as adherence [54] or processes such as telemonitoring [55], but all previous analyses have relied on summary statistics rather than IPD. Compared to the most recent and comprehensive summary data review, the current study has provided pooled estimates of the effect of self-monitoring with different levels of co-intervention, suggests that self-monitoring alone has little impact on BP, and provides new evidence that the level of BP reduction is related to the intensity of the co-intervention [4].
Self-monitoring in the absence of such a co-intervention had little effect on BP. This is not to say that self-monitoring alone should be discouraged, for it brings other advantages both theoretical (better estimation of the underlying BP, increased self-efficacy for the patient) [6] and practical (increased adherence, reduced need for monitoring within the clinic, and identification of white coat and masked hypertension) [24,54]. These advantages are despite any potential inaccuracy caused by individuals not conforming to the recommended self-monitoring regime [9,10].
Obese patients had similar BP reductions to non-obese individuals but greater chance of BP control, which does not reflect differences in mean BP. The findings concerning patients with previous stroke and resistant hypertension require some caution, particularly the latter which was a post-hoc analysis, but have not been previously described. In the case of stroke, the results do not appear to be due to baseline intensity of antihypertensive treatment and warrant further study as more data become available.
Combining self-monitoring with increased collaboration between patient and either a nurse, physician, or pharmacist can result in important decreases in BP (6 mmHg systolic on average for the more intensive co-interventions) and improved control. The mechanisms for these reductions in BP could include lifestyle changes (no data available); increased adherence to medication (no data available) [54]; or increased prescription of medications, i.e., overcoming clinical inertia (data available from 11 studies). In order to assess the impact of enhanced medication prescription, number of medication changes was plotted against changes in BP and showed that increased numbers of medication changes were weakly correlated with reduced BP (S22 Fig). Whatever the mechanism, the literature suggests that a 6 mmHg reduction in sBP, as observed in higher-intensity interventions, would reduce subsequent stroke by more than 20% [56]. Considering the content of such interventions is an important part of decision-making in the implementation of self-monitoring. Table 1 and S1 Table describe the key characteristics of effective interventions which depend on actively intervening in terms of medication titration and/or health behaviours. Much of the effect appears to be associated with one-to-one intervention combined with medication intensification. Self-monitoring can therefore facilitate significant improvements in BP level and control but should not necessarily be seen as reducing clinical input because clinical input within the co-interventions is often required for effective BP lowering.
The recent SPRINT trial results suggest that more intensive BP interventions are likely to be important in terms of morbidity and mortality [57]. Increasing the level or intensity of intervention also increases the cost of an intervention, both directly to the health provider and also in terms of patients’ time. Understanding the relative cost-effectiveness of the different co-interventions is likely to be important in deciding policy in this area and will require further work.
The effects appear to be independent of age, sex, and a range of comorbidities (such as MI, CKD, diabetes, and obesity), but there was a suggestion that people receiving less intensive antihypertensive treatment, and those with the highest BPs (up to 170 mmHg systolic), may have the most to gain, presumably because they are not already receiving sufficient doses of medication. Conversely, with resistant hypertension there appeared to be little effect from self-monitoring. Similar results for stroke should be interpreted cautiously and warrant further study.
The data presented appear to indicate a potential attenuation of the beneficial effects of self-monitoring over time (see S6, S7 and S8 Figs). We believe that the key issue is a need for longer studies (at least 2 years, and preferably 5 years or more) that are accompanied by investigation of how best to administer a self-monitoring—based intervention in the long term, including whether it should be perhaps “topped up” with additional training over time.
Finally, we know from the individual trials that only a proportion of those with hypertension will be suitable for self-monitoring. Despite this, the numbers of people with hypertension and access to their own BP monitor are likely to be well into the tens of millions internationally and represent an important population to engage with [58,59].
Several unanswered questions remain. Ultimately, trials including cardiovascular endpoints would provide the strongest evidence for self-monitoring in the management of hypertension but may not be appropriate given the strong evidence linking BP to outcome. Further consideration of self-monitoring in the presence of comorbidities seems warranted, particularly for stroke. Furthermore, this review has not included economic outcomes (available from 6 of the included studies) or quality of life measures (available in 8 of the included studies), and these outcomes form part of the next series of investigations for this collaboration.
Self-monitoring of BP combined with co-interventions involving individually tailored support lowers clinic BP but has little effect on its own. Self-monitoring supported by such co-interventions should be recommended as part of routine clinical practice in international guidelines and further research should determine the most cost-effective means of supporting implementation.
|
10.1371/journal.pgen.1003904 | Parallel Evolution of Chordate Cis-Regulatory Code for Development | Urochordates are the closest relatives of vertebrates and at the larval stage, possess a characteristic bilateral chordate body plan. In vertebrates, the genes that orchestrate embryonic patterning are in part regulated by highly conserved non-coding elements (CNEs), yet these elements have not been identified in urochordate genomes. Consequently the evolution of the cis-regulatory code for urochordate development remains largely uncharacterised. Here, we use genome-wide comparisons between C. intestinalis and C. savignyi to identify putative urochordate cis-regulatory sequences. Ciona conserved non-coding elements (ciCNEs) are associated with largely the same key regulatory genes as vertebrate CNEs. Furthermore, some of the tested ciCNEs are able to activate reporter gene expression in both zebrafish and Ciona embryos, in a pattern that at least partially overlaps that of the gene they associate with, despite the absence of sequence identity. We also show that the ability of a ciCNE to up-regulate gene expression in vertebrate embryos can in some cases be localised to short sub-sequences, suggesting that functional cross-talk may be defined by small regions of ancestral regulatory logic, although functional sub-sequences may also be dispersed across the whole element. We conclude that the structure and organisation of cis-regulatory modules is very different between vertebrates and urochordates, reflecting their separate evolutionary histories. However, functional cross-talk still exists because the same repertoire of transcription factors has likely guided their parallel evolution, exploiting similar sets of binding sites but in different combinations.
| Vertebrates share many aspects of early development with our closest chordate ancestors, the tunicates. However, whilst the repertoire of genes that orchestrate development is essentially the same in the two lineages, the genomic code that regulates these genes appears to be very different, even though it is highly conserved within vertebrates themselves. Using comparative genomics, we have identified a parallel developmental code in tunicates and confirmed that this code, despite a lack of sequence conservation, associates with a similar repertoire of genes. However, the organisation of the code spatially is very different in the two lineages, strongly suggesting that most of it arose independently in vertebrates and tunicates, and in most cases lacking any direct sequence ancestry. We have assayed elements of the tunicate code, and found that at least some of them can regulate gene expression in zebrafish embryos. Our results suggest that regulatory code has arisen independently in different animal lineages but possesses some common functionality because its evolution has been driven by a similar cohort of developmental transcription factors. Our work helps illuminate how complex, stable gene regulatory networks evolve and become fixed within lineages.
| Gene regulation is facilitated by the binding of transcription factors to specific sites in genomic DNA. Consequently, accurate control of gene expression in any cell is largely influenced by two variables; the presence of the transcription factor proteins themselves and accessibility to regulatory sites. During animal development, a highly complex and dynamic set of regulatory interactions must be precisely articulated in order to accurately direct the patterning of the embryo. This has resulted in the establishment of stable and robust, scale free gene regulatory networks (GRNs) [1], with high information content encoded into cis-regulatory modules (CRMs), where cohorts of transcription factors bind combinatorially to define a regulatory state [2], [3].
As a result of this, the largest and most highly conserved cis-regulatory sequences identified in vertebrate genomes are associated with transcription factor genes that regulate development [4], [5], reflecting both the complexity and precision required to co-ordinate common patterning mechanisms during embryogenesis. Furthermore, the vast majority of these conserved non-coding elements (CNEs) are not conserved at the sequence level in invertebrate genomes, where parallel sets of cis-regulatory sequences have evolved [6], [7]. Interestingly, a tiny handful of vertebrate CNEs do share some sequence similarity with amphioxus elements [8], a more distant [cephalo]chordate relative, and even with elements in protostomes [9]. Recently, a number of shorter regions of sequence homology (av. 45 bp at 55% identity) have been identified between Ciona and vertebrates, although they are not generally associated with orthologous genes in the two lineages, and a majority are transcribed [10].
Nevertheless, urochordates must exploit genomic sequence, in the form of CRMs, to orchestrate their own development, deploying a similar repertoire of genes to vertebrates and other animal lineages. Indeed, patterning of the early vertebrate embryo and Ciona larva bear a strong resemblance to each other, suggesting that the many aspects of urochordate development are very similar to that of vertebrates [11], even if the rate at which their genome sequence has evolved is relatively rapid compared with amphioxus [12]. Two important questions therefore are how, and when, did complex CRMs for embryonic patterning become established in the chordate lineage. Are similarities in urochrodate and vertebrate patterning orchestrated by long established CRMs pre-dating the divergence of the chordate lineages, or have entirely different genomic sequences been recruited and deployed as CRMs in urochordates and vertebrates?
In order to address these questions we have identified a large set of urochordate (Ciona) specific CNEs (ciCNEs) through comparison of the highly diverged C. intestinalis and C. savignyi genomes, and compare them with vertebrate CNEs. The evolutionary distance between the two Ciona genomes is considered to be greater than the distance between human and chick, providing a very low background of unconstrained conservation [12]. Support for this comes from a genomewide study of vertebrate and ciona species which showed that Ciona species evolve about 50% faster than vertebrates [13], with a genomewide average amino acid distance between intestinalis and savignyi of 0.3349 (compared with values of 0.3258 and 0.3735 for human∶chick and human∶frog respectively). Many of the ciCNEs are associated with developmental regulator genes; in some cases the same genes that harbour CNEs in vertebrates, despite an absence of identifiable sequence similarity between the CNEs themselves. We test a number of these ciCNEs using two independent transgenic reporter assays in zebrafish embryos, and find that a small number drive highly specific and reproducible patterns of reporter expression. We then examine the relationship between enhancer sequence and function by further dissecting these sequences. We also assay a number of ciCNEs in C. intestinalis embryos. Our findings suggest that despite a degree of regulatory cross-talk, there is little evidence to suggest that the majority of CNEs in urochordates and vertebrates share sequence ancestry. Although it remains possible that binding site reorganization and sequence drift have resulted in very diverged homologous vertebrate and urochordate sets of CNEs, an alternative simple explanation for our findings is that the two sets of CNEs have been recruited and hardwired into the genome independently, after their divergence from a common chordate ancestor, albeit shaped by a similar repertoire of transcription factors. Functional characterization of a larger set of chordate and vertebrate CNEs would likely prove useful in distinguishing between these two scenarios.
We compared the assembled genomes of C. intestinalis and C. savignyi to identify conserved non-coding DNA sequences (Methods). Our analysis is quite different from a previous whole genome comparative analysis performed on these two genomes to identify highly conserved non-coding sequences [14] in that we removed any sequences that overlapped with known transcripts or non-coding RNAs. Consequently our dataset of 2,336 sequences (Dataset S1) represents predominantly Ciona conserved non-coding elements (ciCNEs).
The length distributions of both C. intestinalis and human CNEs are skewed to the right, with a few very long CNEs in both sets (Figure S1). ciCNEs are on average 181.6 bp long, ranging from 94 bp to 1,883 bp, with median 156 bp. For comparison, the lengths of the 1,373 human CNEs defined by alignment of the human and Fugu genomes [5] range from 93 bp to 737 bp, with a median of 177 bp. The distribution of ciCNEs is slightly more skewed than the vertebrate CNEs, reflecting a large set of short CNEs together with some extreme cases of very long CNEs. A majority of the extremely long ciCNEs overlap predicted exons (data not shown). Therefore, we expect that the extremely long ciCNEs are most likely to be at least partly un-annotated coding sequences. By comparing the sequence conservation between the two sets of CNEs, we find that ciCNEs are also less conserved than vertebrate CNEs. ciCNEs range from 71.0% to 96.8% sequence identity, with a median of 81.7%, while vertebrate CNEs range from 67.8% to 97.9%, with a median of 84.6% (based on human-Fugu pairwise comparisons). Therefore, ciCNEs are both shorter and less conserved than vertebrate CNEs, possibly reflecting a lower sequence constraint, or a simpler regulatory module structure in urochordates than in vertebrates.
We then tested whether CNEs cluster near developmental regulatory genes in the C. intestinalis genome as they do in vertebrate and nematode genomes. By assigning 2,146 ciCNEs to their closest protein coding genes (190 ciCNEs are on unplaced contigs containing no genes), we identified 1,289 ciCNE-associated genes (on average 1.7 ciCNEs per gene). Using the same approach, 1,373 human CNEs [5] are assigned to 397 CNE genes (on average 3.5 CNEs per gene). For this genome-wide comparison of human and Ciona CNEs we used proximity to assign genes to CNEs, however we expect that the numbers of CNE-associated genes are over-estimated as it is known that enhancers (and CNEs) can lie far from their targets. The number of CNE-associated genes in Ciona is likely to be exacerbated by the fact that the Ciona genome is highly fragmented. Nevertheless, in common with vertebrate CNEs, we found that ciCNE-genes are enriched for homeodomain-like (log-odds ratio = 2.03, p-value<2.2e-16), winged helix repressor (log-odds ratio = 1.71, p-value = 6.72e-11), HMG1/2 (log-odds ratio = 1.49, p-value = 1.43e-3) and zing finger C2H2 (log-odds ratio = 0.71, p-value = 1.52e-3) domains. In addition, we also found enrichment for several signalling domains that we previously saw overrepresented among nematode but not vertebrate CNE-associated genes. These domains include EGF-like (log-odds ratio = 0.68, p-value = 2.46e-3), laminin G (log-odds ratio = 1.89, p-value = 1.70e-4), cadherin (log-odds ratio = 1.71, p-value = 4.35e-4) and pleckstrin-like (log-odds ratio = 0.95, p-value = 7.92e-4). Using a compiled set of transcription factors and signalling genes in the C. intestinalis genome [15], we found that both types of genes are highly enriched in the ciCNE-associated gene set (log-odds ratio = 1.78, p-value<2.2e-16 and log-odds ratio = 1.43, p-value = 1.60e-10, respectively). Therefore, in terms of the protein domains they encode, the types of genes associated with CNEs in the C. intestinalis genome are consistent with the genes associated with CNEs in both non-chordate invertebrates [7] and vertebrates [5], perhaps reflecting the evolutionary position of C. intestinalis as an invertebrate chordate.
We then looked to see if the same genes are associated with CNEs in both urochordates and vertebrates. Among the ciCNEs-associated genes there are 32 Ciona genes orthologous to 38 human genes also associated with CNEs (orthology was determined using EnsemblCompara [16]) (Table S1). Interestingly, several of the C. intestinalis genes associated with multiple CNEs are orthologous to human genes also associated with multiple CNEs. For example, human PTCH1 is associated with 3 CNEs and its C. intestinalis orthologue is associated with 4 ciCNEs. We note that most (21/32) ciona genes have multiple orthologues in human. So for example, two paralogous human genes, MEIS1 and MEIS2, are associated with 10 and 42 CNEs respectively whilst their C. intestinalis orthologue is associated with 10 ciCNEs. The fact that orthologous genes in human and Ciona are associated with multiple CNEs further suggests that CNEs are associated with specific regulatory genes. Finally, we identified at least 45 ciCNEs that overlap with a limited number of functionally annotated cis-regulatory regions in the ANISEED database [17]. ANISEED is a database of genomic and functional information, such as gene expression patterns of genes, for ascidian genomes including those of Ciona intestinalis and Ciona savignyi.
It is intriguing that in many cases, CNEs are found next to the same gene in vertebrates and Ciona and yet they bear no observable sequence similarity to each other, despite being highly conserved within their respective lineages. Furthermore, their spatial organisation is very different. The Ciona Meis gene has 10 proximal CNEs of which 4 are upstream and the remaining 6 are dispersed across the first seven introns of the gene (Figure 1). This is in contrast with the distribution of CNEs around vertebrate MEIS1 and MEIS2, where a majority of CNEs in each case are positioned in introns towards the end of the gene or downstream of the coding sequence. In the case of the human genes, the CNEs are often hundreds of kilobases from the coding sequence. This suggests that CNEs might have evolved independently in the two lineages but have then become fixed relatively early in their history, particularly in vertebrates. Nevertheless, the genes they co-associate with play very similar roles in each lineage and so we were interested to see if ciCNEs could function as spatio-temporally specific enhancers in zebrafish embryos, a model vertebrate for this type of study.
To select a subset of ciCNEs for experimental testing of enhancer activity, we first identified all ciCNEs that are associated with genes where the orthologous human gene is also associated with CNEs. We then narrowed down the list of candidate ciCNEs by considering only those associated with genes that have known and specific expression profiles during development according to the ANISEED database [17]. We also avoided the cases where the human CNE cluster is close to multiple candidate target genes and the cases where the predicted target gene (from Woolfe et al, [5]) is not the nearest gene to the CNE. From the remaining, we selected a subset of 22 candidate ciCNEs associated with nine different Ciona genes for experiments (Table 1; Text S1).
We independently tested 21 out of the 22 Ciona CNEs (one CNE failed to amplify during PCR), firstly exploiting a co-injection strategy using a minimal beta-globin promoter [5], and secondly through direct cloning into a Tol2 vector with a c-fos promoter [18]. Whilst levels of GFP reporter expression were generally stronger using the Tol2 vector, presumably due to more efficient integration and therefore reduced mosaicism, the results were highly reproducible between the two approaches. Four out of the 21 CNEs give robust and reproducible patterns of restricted GFP expression at either 24 or 48 hours post fertilisation (hpf) using both methods (Table 1). Two of these CNEs were from the Meis gene locus, one was from the Pax6 region and the other resides within the only intron of the Hhex gene in both Ciona and vertebrates. A further four elements were able, in around 5% of embryos, to drive reporter expression in Tol2 constructs only, but these were considered too weak to merit further analysis.
We looked for consistent and reproducible patterns of GFP reporter expression in cell types other than muscle (we routinely see muscle expression in transient analyses with Tol2) in at least 10% of embryos screened for any particular ciCNE (Table 1). At 48 hpf, Pax6_ciCNE2 drives GFP expression in cranial ganglia and sensory neurons (Figure 2A, C) in 12% of screened embryos. More specifically, GFP is detectable in the sensory neurons innervating the tail fin (Figure 2B) and along the spinal cord (Figure 2D).
The two positive ciCNEs associated with the Meis gene drive very different patterns of GFP expression. Meis_ciCNE10 drives GFP expression in neuronal cells (Figure 2E) in 20% of embryos screened. At 48 hpf, GFP is readily detected in Rohon-Beard neurons (Figure 2F), including those innervating the tail fin (Figure 2G) as well as in trigeminal ganglion neurons (Figure 2H). GFP is observed in both cell bodies and axonal projections. A more detailed confocal analysis shows strong GFP fluorescence into the projections of the Rohon-Beard neurons and trigeminal ganglion, extending to the hindbrain (Figures 2I, J).
Injection of Meis_ciCNE1 drives a very robust pattern of GFP expression (Figure 3A–I). Remarkably, in over 50% of embryos screened, GFP expression is detected in motor neurons (Figure 3A, C) and interneurons (Figure 3B, D). Confocal microscopy allowed us to identify morphological subtypes of interneurons and motor neurons. Two classes of descending interneurons (Figure 3E), one class of ascending interneurons (Figure 3F) and one class of bifurcating interneurons (Figure 3G) were GFP positive, as were as at least two subtypes of primary motor neurons (Figure 3H, I). It should be noted that meis1 has been identified as a gene potentially involved in interneuron migration [19].
Finally, in embryos injected with the Hhex_ciCNE1 GFP expression was detected in cells of the hematopoietic lineage (Figure 3J) The size and morphology of the cells resemble macrophages (Figure 3K, L). In zebrafish, a specific lineage of early macrophages differentiate in the yolk sac before the onset of blood circulation [20].
Previously we have shown that evolutionarily conserved aspects of enhancer function often reside in core regions of a CNE sequence [21]. In order to examine whether sub-regions of ciCNEs are sufficient to drive GFP expression in zebrafish, we carried out an extensive functional analysis of sub-sequences of the pax6 ciCNE and the two meis ciCNEs.
Pax6_ciCNE2 is 413 nucleotides (nt) in length and was initially divided into three non-overlapping segments (Figure 4A) and the relative activity of each sub-region compared with the whole ciCNE (Figure 4B). Only the first two regions are able to activate GFP expression (Figure 4C, D), with the first 171 nt being most active. A further delineated region spanning nt 88–244 is able to drive the same patterns of GFP reporter expression as the entire ciCNE, but in a little under half the number of embryos (Figure 4E). Meis_ciCNE1, a particularly strong enhancer in zebrafish, is 457 nt in length and was similarly initially divided into three non-overlapping segments (Figure 5A). Only the most 3′ region shows any activity (Figure 5C) and this is both anteriorly restricted and observed in ten times fewer embryos than the full length element (Figure 5B). Whilst fusing the middle and 3′ regions together gave a small increase in the number of embryos driving GFP (Figure 5D), a larger central core region, encompassing nt 97–384, is able to drive more robust and comprehensive expression, re-capitulating the pattern driven by the full-length ciCNE (Figure 5E).
Meis_ciCNE10 is a relatively short element, just 108 nt in length. Initially, this element was divided into two overlapping sub-regions ΔCNE10-1 and ΔCNE10-2 (Figure 6A), where all the detectable enhancer activity was confined to the second segment (Figure 6B, C). Further definition of the ciCNE resulted in a 3′ fragment encompassing nt 71–108 (ΔCNE10-2-2) which retains the same enhancer potential as the full element (Figure 6D, E). Deletion of a putative Pbx-Hox site at nt 71–79 from ΔCNE10-2 (ΔCNE10-2-1) or from the full length ciCNE (ΔCNE10-3) results in loss of enhancer potential. However, enhancer activity is also lost on deletion of nt 83–92 from the full length ciCNE (ΔCNE10-4). Further synthetic constructs were then made by annealing complementary oligonucleotides representing nt 71–108 (ΔCNE10-2 oligo1 (Figure 6F)), nt 71–94 (ΔCNE10-2 oligo2 (Figure 6G, H)) and nt 82–108 (ΔCNE10-2 oligo3) resulting in the delineation of a minimal sequence of just 24 nucleotides (nt 71–94, 5′ tgattaatatttcataatgcacta 3′) that is sufficient to re-capitulate both the strength and varied pattern of GFP expression of the full length element.
Trinucleotide site-directed mutagenesis across this region (Figure 7) identifies a critical 12 nucleotide motif, (5′ ttaatatttcat 3′) rich in A+T, and containing strong binding sites for helix-turn-helix homeodomain transcription factors, a diverse group of proteins that play important roles in developmental patterning. However, expression is considerably weaker at all mutated positions across the 24mer, suggesting that, as is generally the case, any homeodomain binding protein might be binding co-operatively alongside other factors across this site. Of particular note is the fact that the first 8 nucleotides of the 24mer sequence represent a perfect canonical Pbx/Hox site (tgatnnat), a bipartite site that itself forms a close relationship with meis proteins, and a motif that is strongly enriched in vertebrate CNEs [22].
We searched for sequence similarity to the 24 nt sequence (5′ tgattaatatttcataatgcacta 3′) that drives highly specific neuronal expression in zebrafish embryos and found no identical sequences in any of the organisms in Ensembl [23] except for the known match close to the Meis gene in C. intestinalis. We also profiled the 24mer for transcription factor binding sites in JASPAR [24] and TRANSFAC [25], predicting a large number of possible sites, including a 13 nt match to the binding site of the Oct domain binding transcription factor POU3F2, a protein known to be involved in neurogenesis in the central nervous system (CNS) [26].
The above experiments demonstrate that despite the absence of sequence conservation between ciona and vertebrate CNEs, 4 out of 21 ciona elements can act as enhancers in zebrafish. Extensive analysis of subsequences of these elements shows that in all cases the minimal functional CNE is at least 12 nt long. This suggests that these ciona elements are recognized and co-ordinately bound by more than one transcription factor in order for them to act as robust developmental enhancers in zebrafish.
We next assessed the activity of selected ciCNEs in Ciona embryos. We focused on well-annotated genes, particularly those with known expression patterns at the tailbud stage of development when major tissue types have been established and transgene assays are viable. Seventeen ciCNEs were assessed, three that had shown activity in zebrafish assays (Pax6_ciCNE2; Meis_ciCNE1; Meis_ciCNE10) and fourteen others (only Pax6_ciCNE1, Meis_ciCNE7, Zfhx_ciCNE1 and Hhex_ciCNE1 were not tested). These were cloned into the reporter vector pCES and electroporated into Ciona zygotes.
At the tailbud stage Ciona Pax6 is expressed in the central nervous system, including both the brain and spinal cord [27]. Pax6_ciCNE2 drove reporter expression into a subset of this domain in the ventral sensory vesicle, a part of the brain (Figure 8A). Both Meis_ciCNE1 and Meis_ciCNE10 drove expression into the ventral sensory vesicle (Figure 8B, C) and anterior tail epidermis (Figure 8D), in a pattern similar to the endogenous expression pattern of the Ciona Meis gene [15], [17]. All three transgenes also drove expression into the endoderm located to the posterior ventral part of the trunk. This is a common ectopic site of expression observed with the pCES vector, reflecting the expression of the gene from which the minimal promoter is derived. The remaining ciCNEs had little or no activity in tail bud stage Ciona embryos: only Nkx2.2/2.4_ciCNE2 showed reproducible expression, with this in the posterior ventral trunk endoderm as described above (data not shown). These cells are distinct from the cells to which the mRNA for this gene localises [15].
These results show that the Pax6 and Meis elements, which drive transgene expression in zebrafish, are capable of driving reporter expression in Ciona in a pattern reflecting the endogenous mRNA domain. Although Nkx2.2/2.4_ciCNE2 was able to increase the residual activity of the basal promoter, it did not drive expression in a pattern related to the expression of Ciona Nkx2.2/2.4. Other ciCNEs failed to activate expression. These may be CNEs associated with gene expression at other points in the life cycle, and hence not active in tailbud stage embryos.
We have identified a genome-wide set of non-coding elements that are conserved between two representatives of the urochordates, C. intestinalis and C. savignyi. Due to the rapid rate of neutral evolution of their genomes, these two species are ideal candidates for the identification of functionally constrained sequences, and have enabled the generation of a valuable data set for comparison with vertebrate CNEs. ciCNEs are slightly shorter and less well conserved on average than vertebrate CNEs, despite the divergence between the Ciona genomes being somewhat less than that between fish and mammals [12]. This suggests, given that these regions are in general candidate cis-regulatory elements, that the complexity of cis-regulation (as measured by the numbers of transcription factors that bind combinatorially to an element at any one time) might be greater in vertebrates than urochordates. This in turn may reflect the increased number of paralogous transcription factors in vertebrate genomes generated through gene/genome duplications and a greater number of tissue types. A further indication of increased vertebrate complexity, at least associated with developmental regulation, is the larger numbers of CNEs that cluster around individual gene loci; for example there are 10 ciCNEs, compared with a total of 52 vertebrate CNEs, associated with the Meis genes (Figure 1).
To try to further understand the relationship between ciCNEs and vertebrate CNEs from a functional perspective, we assayed a set of C. intestinalis CNEs located adjacent to genes that have orthologues in vertebrates that also harbour CNEs. We first adopted a co-injection strategy that has been used previously to characterise vertebrate CNEs in zebrafish embryos, using a minimal beta globin promoter fused to the GFP gene. We then re-assayed all 21 ciCNEs using the well-established Tol2 transgenesis approach, using a vector containing a cfos promoter, again fused to GFP. Although the co-injection strategy resulted in highly mosaic and consequently rather weak GFP expression, we found the same four elements to be active using the Tol2 approach plus another four ciCNEs that drive weaker expression in a small number of embryos. Thus, we believe the results obtained, at least for the four ciCNEs positive in both assays, are robust and reliable and independent of promoter used. Notably we routinely obtained some non-specific ‘ectopic’ muscle expression using the Tol2 vector, but this has been previously documented [28].
The positive ciCNE from the pax6 locus (Pax6_ciCNE2) drives expression in sensory neurons in the spinal cord and cranial ganglia in zebrafish embryos at 48 hpf. GFP expression extends caudally as far as sensory neurons innervating the tail. In zebrafish, pax6 (represented by two genes, pax6a and pax6b) is expressed in sensory placodes, the eye and throughout the CNS during neurogenesis although not specifically in cranial ganglia [29]. Furthermore, whilst pax6 expression is strongest in the ventral spinal cord, sensory neurons tend to originate more dorsally [30]. Similarly, in Ciona, Pax6 is expressed throughout the CNS at early tailbud stage [27], [31]. When the Pax6_ciCNE2 was electroporated in Ciona embryos, expression was observed in the ventral sensory vesicle, the most anterior portion of the Ciona CNS, and related to the vertebrate forebrain. Thus Pax6_ciCNE2 drives reporter expression consistent with the endogenous pattern of expression of the Ciona Pax6 gene, and in a pattern that partially overlaps pax6a expression in zebrafish embryos. However, the same element drives somewhat different patterns of reporter gene expression in the two different organisms. Pax6_ciCNE2 is a relatively large sequence (413 nt) and efforts to dissect it were largely unproductive, although a core region encompassing nt 88–244 is able to drive the same pattern of expression as the whole element but in a smaller proportion of injected embryos, suggesting that this core region is either less stable or a weaker enhancer. Interestingly, Pax6_ciCNE2 has been identified in Ciona previously but was only tested as part of a larger fragment that encompasses the entire intron 4 region in C. intestinalis and as such it does not possess enhancer activity [31].
The Ciona Meis gene has 10 CNEs, and two of these exhibit strong and specific enhancer activity in zebrafish embryos. Although the expression patterns driven by Meis_ciCNE1 and Meis_ciCNE10 in zebrafish embryos are very different, both sequences activate expression in neuronal cells. Meis_ciCNE1 in particular activates reporter gene expression in at least four different classes of interneurons and two classes of motor neurons throughout the CNS and is by some margin the strongest enhancer in either assay. Meis genes act as Hox/Pbx co-factors [reviewed in 32] and whilst particularly associated with hindbrain development in vertebrates [33], zebrafish meis genes are expressed throughout the brain and spinal cord as well as in the developing eye [34]. Ciona Meis is expressed in the ventral sensory vesicle and the anterior epidermis of the tail and posterior trunk at the tailbud stage [15], [35] and the two Meis ciCNEs direct patterns of reporter gene expression consistent with this pattern.
Dissection of Meis_ciCNE1 resulted in the identification of a large core region (nt 97–384 out of 457) of 288 nt that is sufficient to activate the same pattern of reporter gene expression as the whole element, despite a smaller core region comprising nt 156–310 having no enhancer activity. Similar to Pax6_ciCNE2, the larger core region appears to be a weak enhancer, driving expression in less than half the number of embryos than when the whole element is injected. Both the whole element and core region (nt 97–384) are highly active throughout the spinal cord and hindbrain, consistent with a prominent role for Meis genes in hindbrain development, although the core region activates a smaller percentage of injected embryos. Note that there is very limited reporter expression more rostrally in the mid- or forebrain. Contrastingly, the 3′ region of Meis_ciCNE1 (nt 311–457) is able to activate reporter expression more rostrally in the mid-to-forebrain region of the embryo yet not in the hindbrain or spinal cord. A construct combining the middle and 3′ regions of the ciCNE (nt 156–457) however, results in loss of the rostral expression and restoration of primarily the hindbrain, but also the spinal cord expression patterns. Thus it would appear that this ciCNE has the potential to drive expression in the fore- and midbrain encoded in the 3′ region but that this is repressed by upstream sequences in the ciCNE.
Meis_ciCNE10 is already a comparatively small element at just 108 nt in length and as a consequence was initially dissected into two overlapping regions of approximately 70 nt. Firstly, it was apparent that a majority of the activation potential of this ciCNE was located in the 3′ region. Attention focused on a small core region where loss of different motifs resulted in loss of enhancer activity. Strikingly, a minimal region of just 24 nucleotides (nt 71–94) is able to drive reporter expression in the same pattern as the full length element. However this minimal region was no longer able to activate expression in Ciona tailbud embryos. This suggests that mechanisms of activation are subtly different between Ciona and zebrafish.
Hhex_ciCNE1 is located in the single intron of the Ciona Hex gene. This ciCNE drives reporter expression very specifically (in both assays) in a small population of cells located either in the yolk sac or in the circulatory system, with a morphology reminiscent of monocytes or macrophages. This would reflect a role for Hhex_ciCNE1 consistent with that of zebrafish hhex in early haematopoeisis [20], [36].
We also note that the three ciCNEs that tested positive in Ciona tailbud embryos also showed the strongest phenotypes in zebrafish embryos, while the ciCNEs that were negative in Ciona tailbud embryos had limited or no impact in zebrafish. Whether this apparent association is meaningful is difficult to determine, as Ciona transgenesis only assesses construct activity up to a specific point in the life cycle, the tailbud stage. However this stage does present the canonical chordate bodyplan and active neuronal differentiation. One possibility is that Ciona enhancers active at this time point are more likely to also be active cross-species, for example reflecting constraint on underlying regulatory circuitry imposed by the use of similar suites of transcription factors to establish the conserved chordate body plan in the two lineages. In this respect we note that one of the few previous studies to demonstrate cross-species enhancer activity between Ciona and vertebrates also found tailbud stage enhancers to be active in vertebrates, in this case for two enhancers associated with the Ciona Hox1 gene [37]. However, we cannot unequivocally conclude this without exhaustive testing for activity amongst the other ciCNEs at other life cycle stages in both Ciona and zebrafish, and as such it must remain speculative.
In summary, these results demonstrate that at least some of the regulatory logic encoded in ciCNEs can be recognised and deciphered in a vertebrate embryo, directing specific and reproducible patterns of expression in distinct populations of cells during early development. This is in agreement with other studies that have shown that developmental enhancers can function in heterologous contexts in different species (e.g. [38]). However, as we would predict if there has been extensive CRM remodeling, the patterns of expression activated by ciCNEs in zebrafish embryos are not wholly consistent with the endogenous expression of their associated gene, and can in at least one case be driven by a very small sub-region within a ciCNE. Furthermore, it has been established that trans-regulatory changes (i.e. the ability of one species to interpret the cis-regulatory code from another species) also play a role in the reproducibility of enhancer activity [39]. Our data are supported by another recent study that assayed three putative Ciona regulatory elements in zebrafish embryos [10], and suggests that the CRM architecture of vertebrate and urochordate CNEs is very different. We speculate that control is mediated by regulatory cross-talk via a limited number of transcription factors, rather than accurate deciphering of whole ciCNEs as CRMs.
In the vertebrate lineage it is now well established that the most highly conserved regulatory elements are associated with developmental transcription factors, remaining largely unchanged at least since the divergence of cartilaginous fish around 500 million years ago (MYA). However, with just a few exceptions [8], [9], vertebrate CNEs do not show strong sequence similarity to non-vertebrate sequences. In this paper we have tried to examine the reasons for this paradox. Recently, a comparison of vertebrate and Ciona conserved non-coding sequences identified between 150 and 200 short stretches of conservation, termed oCNEs (av. 45 bp at 55% identity) [10]. Surprisingly, oCNEs are not found in syntenic locations in vertebrates and urochordates, but are located close to different developmental regulator genes, suggesting they have been co-opted into novel CRMs and regulatory networks, possibly as a result of major re-arrangement events. 65 oCNEs are embedded in our ciCNE set (we would expect no overlap by chance), in agreement with our hypothesis that CRMs have been extensively re-modeled, and that even the small minority of shared sequence ancestry has been re-deployed into new regulatory elements and networks. Indeed, these two complementary datasets hint at the extent of re-structuring of gene regulatory networks early in chordate history, and contribute to our understanding of the processes of evolution within gene regulatory networks in different lineages. A second important contributing factor to CRM re-modeling might be the result of the continued and rapid evolution of ancestral chordate CNE sequences in the urochordate lineage but many more urochordate genome sequences are necessary to measure this. Finally, the location and spatial organisation of multiple CNEs around genes, such as at the meis loci, also show no obvious relationship between lineages. A majority of CNEs are downstream of vertebrate meis1 and meis2 genes or in 3′ introns, whereas the Ciona meis gene has no downstream CNEs and all are either upstream or in 5′ introns. If vertebrate and urochordate CNEs have evolved from the same ancestral sequences then there must have been a great deal of local rearrangement of these sequences in early urochordate evolution (vertebrate CNEs remain co-linear across all species and the ciCNEs are co-linear between C. intestinalis and C. savignyi). Given these observations, we conclude that urochordate and vertebrate CNEs emerged and evolved largely independently.
Conservation of CRM function in the absence of sequence conservation or ancestry is not surprising. There are many well-documented examples of regulatory conservation with low or no sequence conservation [40], [41]. Because transcription factor binding sites are small and degenerate, they can easily arise by chance within a short stretch of genomic sequence thereby making existing binding sites redundant [42], [43]. In this way, previously established regulatory regions can become highly divergent, or new sequence regions may be recruited as regulatory sites, without an overall change in function. Alternatively, extensive re-wiring of the regulatory code can create a new set of CRMs that still co-operate within the GRN to achieve the same output. This is supported by the fact that divergent expression profiles of orthologous sets of zebrafish and Ciona genes can still result in similar body plans [44].
Despite an apparent lack of direct sequence ancestry, CNEs from vertebrate and urochordate genomes will not have evolved completely independently. They are associated with the same genes and regulatory networks. Consequently, as we demonstrate here, a number of ciCNEs tested are recognised, at least in part, by specific developmental regulatory states (i.e. a set of transcription factors) when injected into the genome of a species that has been evolving independently for over 500 million years. In essence, this reveals that in several cases vertebrate and urochordate CNEs represent different solutions to the same problem, ensuring that similar cohorts of transcription factors active in a particular cell type switch on the same target gene.
The C. intestinalis repeat-masked genome (version v2.0) was retrieved from the Joint Genome Institute website (http://genome.jgi-psf.org/Cioin2/Cioin2.info.html). The C. savignyi repeat-masked genome (version v2.1) was retrieved from the Sidow lab website (http://mendel.stanford.edu/sidowlab/Ciona.html) at Stanford University Medical Centre [45]. For the BLAST similarity search, the C. savignyi scaffolds were split into 500 kb fragments overlapping by 200 bp using the EMBOSS [46] program splitter. The C. savignyi fragments were then searched for similarity against the sequence of the C. intestinalis genome using MegaBLAST [47]. MegaBLAST was run with word seed length 20 bp, mismatch penalty -2 and e-value threshold 0.001, as described previously for the Fugu-human whole genome comparison [5]. This search returned 177,708 matches between the C. savignyi sequence fragments and the C. intestinalis genome. In line with Fugu∶human comparisons [5], only alignments at least 100 bp long were retained, thus reducing the set to 73,728 sequence hits.
The C. intestinalis conserved sequence elements were first screened for evidence of transcription according to Ensembl C. intestinalis annotation (release v40) using Ensembl Perl API [23]. Elements overlapping exons or containing in total more than 10 bp of repeats were removed. Conserved elements were further filtered by searching for similarity against the EMBL EST, Rfam and the microRNA registry using MegaBLAST (e-value cut-off 0.001). All elements with matches to the non-coding RNA databases were removed and elements with more than three matches to expressed transcripts from EMBL were also removed. In addition, because analysis of duplicated elements was beyond the scope of this manuscript, C. intestinalis elements matching multiple locations in C. savignyi were removed, too. The final set consists of 2,336 C. intestinalis CNEs (ciCNEs), where for 1,817 ciCNEs there is no evidence of transcription and for 519 there is little evidence of transcription (up to 3 matches to expressed transcripts from EMBL)
The nearest protein-coding genes (i.e. genes with the nearest TSS) to ciCNEs were retrieved using Ensembl Perl API. 190 of the 2,336 cCNEs were in sequence fragments that did not contain any genes. The remaining 2,146 ciCNEs were assigned to 1,289 protein-coding genes. The human orthologs of the ciCNE-associated genes were retrieved using Ensembl Perl API accessing the Ensembl Compara database, C. intestinalis Ensembl Core and H. sapiens Ensembl core database (Ensembl release v43).
This was performed as previously described for nematode CNEs [7]. In brief, we downloaded InterPro domains of all human and ciona genes from Ensembl [22]. Using a custom Perl script we converted all domains to their top-level parent domain based on InterPro annotation hierarchy [48]. We removed domains present in fewer than 10 genes. We calculated the enrichment of each domain in CNE-associated genes versus the rest using the log-odds ratio test in R and accounted for multiple testing using the Benjamini and Hochberg method [49].
CNEs were amplified from Ciona genomic DNA by PCR and assayed in zebrafish using the Tol2 system [50]. The PCR products were cloned into the pCR8/GW/TOPO vector (Invitrogen) and then into a Tol2GFP construct [51], using the Gateway LR Clonase II enzyme (Invitrogen). Transient transgenic zebrafish embryos were screened for GFP expression at 24 hpf and 48 hpf.
Mutations in the 24 nt sequence of Meis_ciCNE10 were generated by mutating the wild type sequence already inserted into the tol2 vector using the ‘QuickChange II Site-Directed Mutagenesis Kit’ (Agilent Technologies).
Putative ciCNE fragments were directionally cloned in 5′ to 3′ orientation into the β-galactosidase based reporter vector pCES, which uses a minimal promoter derived from the C. intestinalis FoxAa gene [52]. Adult C. intestinalis type B were collected from marinas on Hayling Island, South England, and maintained in a re-circulating sea water aquarium at 12°C. Gametes were removed separately by dissection, eggs fertilised in vitro and the chorion removed chemically [53] within 15 mins of fertilisation. Electroporation of fertilised eggs was carried out as described, [54], with modifications [55], using 40 µg of construct DNA. Embryos were cultured until the tail bud stage before fixation in 0.2% glutaraldehyde for 30 minutes in sea water, two washes in PBS and transfer to staining buffer (3 mM K5Fe(CN)6, 3 mM K3Fe(CN6), 1 mM MgCl2). They were stained in staining buffer containing 4 mg ml−1 Xgal at 37°C for 12 to 72 hours. All experiments included a negative control (the pCES vector without an enhancer inserted) and a positive control (the Ciona βγ-crystallin enhancer [55] in pCES). All negative controls showed no reporter expression, and positive controls showed at least 50% of embryos with palp and/or sensory vesicle expression, reflecting a typical rate of successful electroporation by this method [56].
|
10.1371/journal.pntd.0001260 | Resolution of Praziquantel | Praziquantel remains the drug of choice for the worldwide treatment and control of schistosomiasis. The drug is synthesized and administered as a racemate. Use of the pure active enantiomer would be desirable since the inactive enantiomer is associated with side effects and is responsible for the extremely bitter taste of the pill.
We have identified two resolution approaches toward the production of praziquantel as a single enantiomer. One approach starts with commercially available praziquantel and involves a hydrolysis to an intermediate amine, which is resolved with a derivative of tartaric acid. This method was discovered through an open collaboration on the internet. The second method, identified by a contract research organisation, employs a different intermediate that may be resolved with tartaric acid itself.
Both resolution procedures identified show promise for the large-scale, economically viable production of praziquantel as a single enantiomer for a low price. Additionally, they may be employed by laboratories for the production of smaller amounts of enantiopure drug for research purposes that should be useful in, for example, elucidation of the drug's mechanism of action.
| The drug praziquantel (PZQ) is used very widely in both animal and human medicine, where it is the mainstay of the treatment of the neglected tropical disease schistosomiasis. The drug is currently manufactured and administered as a racemate (1∶1 mixture of enantiomers) but for various reasons the large-scale production of PZQ as the single active enantiomer is very desirable. We describe here the preparation of praziquantel as a single enantiomer using classical resolution. The protocols are experimentally simple and inexpensive. One method was found and validated by an unusual research mechanism—open science—where the details of the collaboration (involving academic and industrial partners) and all research data were available on the web as they were acquired, and anyone could participate. The other route was found in parallel by a contract research organisation. Besides being possible routes by which praziquantel may be produced in large quantities for the affected communities, it is also hoped that these methods can be used for the production of smaller quantities of enantiopure PZQ for pharmacological studies.
| Schistosomiasis (bilharziosis) is termed a "neglected" tropical disease owing to the continuing low level of investment in treatments, prevention and research, yet the disease accounts for an extraordinarily high level of suffering around the world.[1], [2] Schistosomiasis has been called a "silent pandemic".[3]
Over the past decades several compounds have been used for the treatment of schistosomiasis,[4]–[6] but today there is only one drug of choice, a highly effective small molecule called praziquantel (PZQ).[7], [8] PZQ is produced on a very large scale (300 metric tons worth of API per year) and is used primarily in veterinary medicine. In human medicine, PZQ is used essentially as preventive (mass) chemotherapy for all forms of schistosomiasis - whereby school-aged children or entire communities are given a dose of PZQ once a year. Such mass treatment programs (e.g. that coordinated by the Schistosomiasis Control Initiative)[9] deploy 100 million tablets annually, and there is expected to be a further large growth in the demand for PZQ in the coming years.[10]
With increasing use comes an increased risk of the development of resistance or tolerance by the parasite. Decreased drug sensitivity was developed via an artificial selection experiment in the laboratory,[11] and reports of similar decreases have already been noted in the field.[12]–[15] Reliance on a single drug for intensive mass treatment is risky. While there have been attempts to find bioactive PZQ analogs,[16]–[18] as well as the discovery of new compounds for the treatment of schistosomiasis based on different modes of action,[19], [20] in the short term it is sensible to continue to use PZQ in a way that maximizes its life as a useful drug.
PZQ is synthesized and administered as a racemate. The L-(–)-enantiomer is the eutomer[21]–[24] and has the (R) configuration.[7], [23] Administration of the pure eutomer resulted in fewer side effects than the racemate.[22] The inactive (+)-enantiomer is associated with side effects and is also primarily responsible for the extremely bitter taste of the pill;[25] factors such as taste and large pill size contribute to there being a compliance problem with PZQ in the affected communities.[26]–[27] The typical dose per pill (40 mg kg−1, pill contains 600 mg active pharmaceutical ingredient (API)) is large. The pill is difficult to swallow for children (who are the main target of mass chemotherapy campaigns) often requiring tablets to be split and crushed, which brings out the bitter taste even further. Decreasing the pill size, reducing side effects and removing the bitter taste, while having the right amount of the active ingredient, could be accomplished were the drug to be made available as a single enantiomer. For these reasons investigations into the viability of a process-scale route to enantiopure PZQ were included in the WHO/TDR business plan for 2008–2013.[28] Availability of the separate enantiomers would be a valuable tool for the elucidation of the mechanism of action of the drug, still unknown after more than 30 years of use;[29] in such experiments the inactive enantiomer would act as the perfect control.
There are typically four methods available for the conversion of a racemic synthesis to one that generates a single enantiomer (Figure 1): 1) Enantioselective synthesis, 2) Chromatographic separation, 3) Stereoablation (destruction and selective reconstruction of the stereocenter) and 4) Resolution. To date, reports of the preparation of enantiopure PZQ either have insufficient detail to allow for their appraisal, or likely do not have the potential for the large-scale production of the drug given the severe price constraint;[25], [30]–[34] for example with enantioselective chromatographic approaches significant quantities of solvents would be required.
Racemic praziquantel is off-patent. Through market forces, and its production on such a large scale, the racemate is available for approximately US10¢ per gram. A reasonable price for wide distribution of the enantiopure compound would therefore be approximately US20¢ per gram since half the dose is required of the enantiopure compound. Can enantiopure PZQ be obtained without a significant increase in price? It is a challenge for either academia or industry to solve this problem, for different reasons. This article describes the discovery of two solutions, one by an open collaborative project that starts from rac-PZQ, and one conducted by a contract research organization that employs a commercially-available precursor. We term these “resolutions of praziquantel” for simplicity. Formally both processes involve the resolution of a precursor or derivative that is then converted to praziquantel.
(rac)-PZQ was a gift from WHO/TDR; the sample was originally synthesized by Merck. Analytical samples of (R)- and (S)-PZQ were a gift from Intervet Innovation GmbH and were prepared according to the literature method.[25]
To a solution of (R)-(–)-PZQ (c = 1, EtOH) was added a solution of (S)-(+)-PZQ (c = 1, EtOH) to give a total volume of 1 mL (Table S1). The results indicate a linear relationship between optical purity and optical rotation (Figure S5).
To a solution of (S)-(+)-PZQamine (c = 1, DCM) was added a solution of rac-PZQamine (c = 1, DCM) to give a total volume of 1 mL (Table S2). The results indicate a linear relationship between optical purity and optical rotation (Figure S6).
rac-Praziquanamine (10.0 g, 49.5 mmol) and (–)-dibenzoyl-L-tartaric acid•2 i-PrOH (23.7 g, 49.5 mmol) were dissolved in isopropanol (450 mL) and water (90 mL) by heating the stirred mixture. The solution was allowed to cool to rt and after 2 h the colourless crystals were filtered and dried to yield the salt as pale yellow crystals (12.1 g, 44%). m.p. 145.5–147.5°C. Small-scale liberation of amine (procedure below) gave [α]D20 (liberated amine) = -242° (c = 1, DCM), implying 79% ee (determined by polarimetry).
The salt was recrystallized from a mixture of isopropanol (180 mL) and water (90 mL). The crystalline precipitate was kept at 5°C for 12 h before filtration, though the crystallization is essentially complete after 2 h. This procedure gave the salt as colourless spicular crystals (10.2 g, 85% from this procedure, 37% overall). m.p. 147.3–148.5°C.
To an ice-cooled solution of R-(–)-PZQamine (3.27 g, 16.2 mmol) and triethylamine (2.45 g, 3.38 mL, 24.3 mmol, 1.5 eq.) in dichloromethane (80 mL) was added dropwise cyclohexanoyl chloride (2.62 g, 2.39 mL, 17.8 mmol, 1.1 eq.) at 0°C and stirring was continued for 14 h at rt. The solution was quenched with water (10 mL) and stirred for a further 30 min. The layers were separated and the organic layer was washed with saturated sodium carbonate solution, 0.5 N HCl solution and brine, dried over magnesium sulfate and concentrated under reduced pressure. The remaining yellow oil became solid after drying under high vacuum and storing at 5°C. The pale yellow solid was recrystallized from acetone/hexane (35 mL, 1∶1 mixture) and two further batches of analytically identical crystals were obtained from the mother liquor after concentration and recrystallization. R-(–)-PZQ was thus obtained as colourless crystals (4.56 g, 90%, 97% ee). (Figure S11) m.p. 113.5–114.5°C. [α]D20 = -136° (c = 1, EtOH).
An outline description of this procedure can be found online.[50]
A coordination website was created on which was posted the problem of the preparation of praziquantel as a single enantiomer.[51] While suggestions were received, input was initially low. In mid-2008 the project was funded by a government/NGO consortium. The resulting raw experimental data were posted in full to an open, online electronic lab notebook (based on the open source electronic lab notebook system, Labtrove, developed by the University of Southampton, UK. [52]) Periodic updates were posted on the coordination website, and the project was popularised to increase traffic (For a description of how the open science project was conducted, see the accompanying paper [53]).
Two approaches were begun in the laboratory that have so far proved intractable. The first, a community suggestion, relied on oxidation of PZQ to an enamide, which was to be subsequently hydrogenated asymmetrically[54] (a similar approach was described in a patent, employing Raney Nickel modified with tartaric acid, giving products with low optical purities – see reference [34]); this is a strong approach owing to the highly effective use of asymmetric hydrogenation in process chemistry.[55] Through an online collaborative process, catalysts are being screened for this reduction[56] but the reaction is difficult owing to the lack of a well-placed coordinating group able to direct the metal catalyst to the double bond. The second approach was based on an asymmetric Pictet-Spengler reaction.[57] Catalysts for similar reactions are known,[58] and the relevant starting material (a peptide acetal) is an intermediate in two known syntheses of PZQ.[39], [59] Unfortunately this substrate contains an unreactive aromatic ring (i.e. lacking electron donating groups), and at the time of writing no known asymmetric catalyst has given conversion to PZQ.
The third possibility was resolution. Such an approach is widely used in the process-scale production of enantioenriched intermediates because the relevant chiral resolving agents are frequently inexpensive and/or can be recycled. Inputs to the collaborative website and elsewhere suggested this approach was more likely to lead to an economically viable solution to the problem.[60] In response, this approach was prioritized.
To effect a resolution, PZQ should be hydrolysed to praziquanamine (PZQamine, Figure 2A). The process must employ only crystallizations (rather than chromatography) to be practicable. The use of procedures that avoid the synthesis of complex catalysts, chromatographic purifications and NMR-based assessments of purity would also assist laboratories in underdeveloped countries to access enantiopure PZQ locally on smaller scales.
In the corresponding author's laboratory, PZQamine could be generated with ease, but the enantiomers could not be baseline separated by enantioselective HPLC due to a limited range of chiral stationary phases being available. This precluded a convenient local assay for resolution trials. In addition several attempts to resolve PZQamine with a range of chiral acids had met with mixed success.[61]–[62] To find a suitable chromatographic assay, an appeal for assistance was posted in several online discussion boards. In particular, the Process Chemists Group on LinkedIn furnished multiple offers of help. One company, Syncom B.V., a contract research organization in the Netherlands, additionally offered to perform a free screen of chiral acids for the resolution of PZQamine in order to discover a lead structure for the project. One gram of racemic PZQamine was shipped to Syncom. An effective chiral stationary phase was found,[36] followed by a chiral resolving agent ((–)-di-p-anisoyl-L-tartaric acid) that permitted the isolation of the desired (R)-enantiomer of PZQamine from the mother liquor in ca. 66% ee, which could be increased to 95% ee after one recrystallization.[63]
With this lead in hand, optimization of the process was carried out. The resolving agent in question is commercially available (reasonably expensive on a small scale) but could be synthesised from tartaric acid. However, purification away from the p-methoxybenzoic acid byproduct formed during its synthesis was non-trivial. It was also thought that isolation of the desired enantiomer of PZQamine from the resolved solid, rather than the mother liquor, would be more desirable; hence (+)-di-p-anisoyl-D-tartaric acid was prepared[46] and used for the resolution of praziquanamine to give the desired (R)-(–)-praziquanamine in the first-precipitated salt.[64] Such an approach is sub-optimal since this resolving agent must be obtained from unnatural enantiomer of tartaric acid. It was found that the simple expedient of using (–)-dibenzoyl-L-tartaric acid solved both these problems, allowing the isolation of (R)-PZQamine in 44% yield and 80% ee without recrystallization and 33% yield and 97% ee after one recrystallization (the maximum yield for a resolution is 50%). Although we did not expect (–)-dibenzoyl-L-tartaric acid and (–)-di-p-anisoyl-L-tartaric acid to give opposite enantiomers of praziquanamine in the first-precipitated salts based on our experience with Dutch resolution and the “family” behaviour of resolving agents, this is not an isolated example and non-familiar behaviour has been observed in other resolutions.[65]–[66] No Horeau effect[67] is observed for either PZQ[41] or PZQamine[42] in common solvents and concentrations at room temperature, meaning that for analytically pure samples of either, optical activity can be used as an assessment of optical purity in laboratories without access to enantioselective HPLC.
(R)-PZQamine can be converted to (R)-PZQ with commercially-available cyclohexanoyl chloride in 90% yield,[49] thus completing the formal resolution of PZQ. The resolving agent can be recycled in 89% yield. Conditions to effect the racemization of the undesired (+)-PZQamine are now being sought.[68]
At the same time as this procedure was being discovered by an open approach, another contract research organisation was asked (by WHO/TDR) to look into solutions to the same problem without communication to the open project. This led to the discovery of a complementary resolution (Figure 2B). From an investigation of compounds available in bulk, a resolution of a commercially-available intermediate (3) was assayed. Tartaric acid could effect this resolution to provide the enantioenriched intermediate in 37% yield and 94% ee. PZQ could be synthesized from enantioenriched 3 by cyclization with chloroacetyl chloride and removal of the benzoyl group, generating (R)-PZQamine, which can be taken on to provide enantiopure (R)-PZQ as before. A summary of this method was posted to the coordination website when complete.[50]
Full experimental details for the open process may be found in this paper. Readers are encouraged to review, evaluate and contribute to refining the resolutions online by addressing current weaknesses (e.g., the need for a chlorinated solvent extraction process in the initial PZQ hydrolysis). Both processes show sufficient promise in terms of cost on a lab scale (simple methodology, inexpensive resolving agents, good yields and efficiencies) that costs approaching those needed should be attainable upon scale-up; the processes are therefore being examined by WHO/TDR on a kilogram scale for economic viability. The routes found are quite similar. An advantage of the approach discovered by the CRO is its use of tartaric acid itself, as opposed to a derivative, but the derivatization employed in the open approach is straightforward. Which route is adopted depends to some extent on the method(s) currently employed in the commercial manufacture of the API, and perhaps surprisingly this information is not readily available. The ton-scale availability of 3 implies its use in the synthesis of PZQ, presumably via the original Merck process,[7] yet to the best of our knowledge the CRO manufacturing PZQ for the Schistosomiasis Control Initiative (Shin Poong, South Korea) were employing a different approach[39] that generated PZQamine 2 as an intermediate, implying a similar availability of that material in quantity. The open source approach is the basis of an educational project in which students from around the world are encouraged to collaborate in further optimization. (Interested students and laboratory instructors can view the experiments and collaborate on the relevant website[69]).
It is clear that the availability of a new synthetic route, even if it is economically viable, does not translate automatically into a product available to the end-user. Additional elements must be taken into consideration including the regulatory requirements for further studies (chemistry, manufacturing & control; non-clinical; and clinical). While the time and costs associated with this process are expected to be significantly less than with a typical new chemical entity, they are yet to be quantified and supported. WHO/TDR is actively seeking commercial partners potentially interested in pursuing this project as well as sources of funding.
|
10.1371/journal.pcbi.1002765 | Structural Insights into the Inhibition of Actin-Capping Protein by Interactions with Phosphatidic Acid and Phosphatidylinositol (4,5)-Bisphosphate | The actin cytoskeleton is a dynamic structure that coordinates numerous fundamental processes in eukaryotic cells. Dozens of actin-binding proteins are known to be involved in the regulation of actin filament organization or turnover and many of these are stimulus-response regulators of phospholipid signaling. One of these proteins is the heterodimeric actin-capping protein (CP) which binds the barbed end of actin filaments with high affinity and inhibits both addition and loss of actin monomers at this end. The ability of CP to bind filaments is regulated by signaling phospholipids, which inhibit the activity of CP; however, the exact mechanism of this regulation and the residues on CP responsible for lipid interactions is not fully resolved. Here, we focus on the interaction of CP with two signaling phospholipids, phosphatidic acid (PA) and phosphatidylinositol (4,5)-bisphosphate (PIP2). Using different methods of computational biology such as homology modeling, molecular docking and coarse-grained molecular dynamics, we uncovered specific modes of high affinity interaction between membranes containing PA/phosphatidylcholine (PC) and plant CP, as well as between PIP2/PC and animal CP. In particular, we identified differences in the binding of membrane lipids by animal and plant CP, explaining previously published experimental results. Furthermore, we pinpoint the critical importance of the C-terminal part of plant CPα subunit for CP–membrane interactions. We prepared a GST-fusion protein for the C-terminal domain of plant α subunit and verified this hypothesis with lipid-binding assays in vitro.
| The actin cytoskeleton is a prominent feature of eukaryotes and plays a central role in many essential aspects of their lives. This highly malleable structure responds to a wide range of stimuli with rapid changes in organization or dynamics. These responses are thought to be mediated by dozens of actin-binding proteins, the biochemical activities of which have been demonstrated to be tightly controlled by other proteins and/or signal transduction mediators. In this study, we investigated the structural aspects of inhibition of actin-capping protein (CP) by phosphatidic acid (PA) and phosphatidylinositol (4,5)-bisphosphate (PIP2). We employed diverse computational methods in combination with experimental approaches to reveal mechanistic details of the direct interaction of CP with the phospholipid membrane containing either PA or PIP2. Importantly, we found several differences between PA/PIP2–CP interactions from two distinct species, Arabidopsis and chicken, that enable us to explain and expand upon previously published results. Our new data shed light on the nature of interactions between peripheral membrane proteins and PA-containing lipid bilayers. In addition to a description of the phospholipid-mediated regulation of CP activity, our work also significantly contributes to the ongoing debate on structural details of protein interactions with phospholipids.
| The actin cytoskeleton represents part of a complex network that is essential for cell motility, organelle movements and cell polarity. Actin filaments are dynamic structures in general and, in plant cells, they serve as tracks for some of the fastest movements on earth. To regulate actin cytoskeleton organization and dynamics, cells use more than a hundred classes of actin-binding proteins (ABPs). To a limited extent, these proteins can be classified based on their binding properties and activities in vitro. Some ABPs bind actin monomers regulating the size and activity of the polymerizable actin pool, whereas others bind to the sides of actin filaments. Side-binding proteins can create higher-order filament structures like meshworks and bundles, or they can create breaks and sever filaments. Another group of ABPs interacts with actin filament ends and regulates the stability and dynamics of polymer assembly/disassembly [1]. A conserved member of this latter group is actin-capping protein (CP or CapZ), which inhibits the addition and loss of actin subunits at the barbed end of actin filaments [2], [3].
CP is a heterodimeric protein with a mushroom-like structure [4]. Each monomer, α and β subunit (CPα and CPβ), has a molecular weight of approx. 30 kDa and despite their sequence divergence, they have similar structural folds [4]. Several recent studies describe a mode of interaction between CP and the actin filament barbed end [5], [6], highlighting the importance of C-terminal domains from both subunits. These C-terminal parts form so-called tentacles laying on the top of the protein and are mainly composed from amphipathic helices [4]. It has been shown previously that binding of CP to actin filaments is regulated by several other proteins, either by competition for filament ends or by direct protein-protein interactions and allosteric regulation [7]. Another set of key regulators that inhibit CP activity are the signaling phospholipids, phosphatidylinositol (4,5)-bisphosphate (PIP2) and phosphatidic acid (PA) [8]–[12].
Phospholipids are part of the complex lipid-signaling language of eukaryotic cells and enable communication between plasma membrane, endomembrane compartments and cytoplasm. The role of phosphoinositides (PPIs) as signaling molecules was established many years ago [13]. More recently, PA has emerged as an important signaling messenger, especially in plant responses to biotic and abiotic stress [14]. This acidic phospholipid often functions by recruiting effector proteins to membranes in a spatio-temporally specific manner and/or it affects the biophysical properties of membranes [15]. One characteristic feature of PA and PPIs is their rapid turnover, which is mediated by particular enzymes producing and degrading them [16]. Despite the fact that both PA and PIP2 have important signaling functions, they significantly differ in their biophysical properties. PIP2 contains a bulky headgroup, with net charge ranging from −3 to −5 under physiological pH. and an inverted conical shape that promotes positive curvature of membranes. On the other hand, PA has a tiny headgroup with net charge ranging from −1 to −2 and it may induce formation of membrane structures with negative curvature [17], [18]. Although PIP2 binding by proteins is generally very well described and diverse binding-domains have been discovered [19], [20], much less is known about PA-protein interactions [14].
The ability of PIP2 to regulate CP has been known for a long time [7]; however, there is still some controversy about the exact binding site on CP. Kim et al. [11] performed an exhaustive site-directed and truncation mutagenesis of chicken CP (GgCP). These authors report that mutation of basic amino acids located on the α tentacle (R256, K260) as well as on the β subunit (R225) caused a reduction in PIP2 binding by about 4-fold. A similar reduction in PIP2 binding was observed following deletion of the last 28 C-terminal residues from the α tentacle. Although these results clearly show the importance of the α tentacle for binding to phospholipids, neither mutations or truncations totally abolished PIP2 binding. Kuhn and Pollard [12] studied fission yeast CP and its interactions with PPIs. These authors did not find any effect of various PPIs, including PIP2, on Schizosaccharomyces pombe CP activity. They constructed a homology model for CP from several species and, based on the comparison of electrostatic potentials mapped onto these structures, they hypothesize that a positively-charged patch located on CPβ close to the basic cluster on the α tentacle (which is absent in S. pombe CP) also contributes to the interaction with PPIs. Identification of a PA-binding site on CP remains more elusive; two seminal works that describe the effect of signaling phospholipids on mammalian CP, indicate that PA is not able to inhibit and/or dissociate this protein from actin filaments [8], [9]. However, we showed that mouse CP was able to bind PA, but with lower affinity than Arabidopsis thaliana CP (AtCP). We also demonstrated that PA is a potent inhibitor of AtCP activity, preventing it from interacting with filament barbed ends [10].
In this study, we focus on the interaction between AtCP, GgCP, PA and PIP2 in the context of phospholipid bilayers. To gain a structural perspective about these interactions, we utilized a combination of different computational methods and experimental approaches. We used the recently described MARTINI force field [21], [22] to investigate dynamics of CP binding to phospholipid bilayers containing PA or PIP2. We show different preferences of animal and plant CP towards distinct signaling phospholipids. Our results clearly reveal the importance of C-terminal tentacles from both subunits in these interactions. We further confirm the importance of the α subunit tentacle from AtCP in the PA interaction with an in vitro binding experiment using a GST-fusion protein. Altogether, our results explain and significantly expand upon previously published results [10]–[12].
Given that CP has been identified as one of the major regulators of actin dynamics in different species, such as animals, fungi and plants [7], we asked whether CP is a generally distributed actin-regulating protein in eukaryotes. To achieve this goal, we searched more than 50 genomes for different species covering members of almost all eukaryotic superkingdoms [23]. Both CP subunits are well conserved in most eukaryotic lineages and are mostly present as single-copy genes. Nevertheless, in some organisms CP genes are multiplied; for example, vertebrates have three different genes for the α subunit and Trichomonas vaginalis has five genes for the β subunit (Figure 1). Moreover, the vertebrate gene for β subunit undergoes alternative splicing, producing additional variability [7]. It is worth noting that there is no organism with just one subunit gene for the heterodimer, i.e. an α gene but no β gene, or vice versa; this finding correlates well with genetic and biochemical data indicating strict dependency between α and β subunits. Surprisingly, we have not found CP genes for either subunit in sequenced genomes of green algae, red algae and in certain parasites such as Toxoplasma gondii. Some of these organisms probably lost CP genes during evolution, mainly because of their life strategies, i.e. parasites or extremophiles. The overall phylogeny of both CP subunits mainly follows organismal evolution (Figure 1). Metazoan genes, together with Choanoflagellate Monosiga brevicollis as a basal clade, cluster with Fungi in the case of both CP subunits. Plant sequences also form well supported groups. The phylogenetic relationships between other sequences of CPα (from Chromalveolata, Excavata and Amoebozoa groups) are not so clear. In the case of CPβ, Ameobozoa and Excavata sequences form well supported clusters. We also tried to find homologs of the eukaryotic protein in eubacteria and archeabacteria using more sensitive search tools, such as PSI-BLAST [24], but we did not found any obvious homologous sequences. Therefore, it is reasonable to speculate that CP is an eukaryotic innovation, similar to other ABPs, e.g. formins [25].
To clarify the mode of animal CP binding to PIP2 and to compare it with the binding of CP from different species to PA and PIP2, we utilized diverse methods of computational structural biology. First, we constructed a homology model for AtCP using the crystal structure of GgCP α1β1 (also known as CapZ; [4]) as a template (Figure 2A). A comparison of electrostatic surface potential for both structures shows marked differences in the distribution of charged residues. AtCP is much more negatively charged than the chicken protein (Figure 2B), but it contains one positively charged patch corresponding to the PIP2-binding region on GgCP identified by Kim et al. [11]. To further test the binding modes between PA and PIP2 binding by AtCP and GgCP, we used a computational molecular docking approach similar to that of Kim et al. [11]. Results for the docking of truncated PA (diacetyl-PA) to AtCP ended with a single prediction of binding site and correlate well with the positively-charged patch located on the α tentacle (Figure S1). We also computed the docking of a truncated PIP2 molecule to AtCP with the same results. As a control for these experiments, we used phosphatidylcholine (PC) and docking of this molecule did not result in any single prediction.
Phospholipids spontaneously form more complex systems, such as membranes or vesicles; therefore, we thought it important to ask what is the mode of CP binding to signaling phospholipids in the context of a lipid bilayer. Molecular dynamics (MD) simulation provides a useful and powerful tool to study complex biological systems, such as membranes or proteins [26]–[28]. We employed coarse-grained MD (CG-MD) with the MARTINI force field [21], [22]; this allowed us to simulate larger systems for longer periods of time and has been successfully applied to describe processes like raft-like structure formation, membrane protein dynamics or SNARE-mediated vesicle fusion [28]–[30]. We modeled self-assembly of a lipid bilayer in the presence of CP protein, as this procedure has been shown to be advantageous for the characterization of peripheral membrane protein dynamics [31], [32]. Specifically, we simulated several systems comprising different concentrations of 1-palmitoyl-2-oleoyl-phosphatidic acid/1-palmitoyl-2-oleoyl-phosphatidylinositol (4,5)-bisphosphate and 1-palmitoyl-2-oleoyl-phosphatidylcholine (POPA/POPIP2 and POPC) in the presence of AtCP or GgCP (Table 1). Snapshots from 100 ns of self-assembly of a lipid bilayer containing 20% POPA in POPC in the presence of AtCP are shown in Figure 3. We observed formation of a lipid bilayer within approx. 30 ns in all simulations. This is similar to the time required for membrane formation as described by previous studies [32], [33]. The membrane initially aggregates in the vicinity of CP (Figure 3B); however, the protein is very quickly pushed from the core of the lipid bilayer (Figure 3C, D). CP is peripherally bound to the membrane after approx. 50 ns and remains closely attached to the membrane for an additional 50 ns (Figure 3D). In all simulations performed (i.e. either AtCP or GgCP, and either POPA or POPIP2 in POPC membranes), the CP protein faces towards the lipid bilayer via its tentacles (Figure 3D), but the involvement of the tentacles in the interaction with the membrane is slightly different for particular simulations. Importantly, the protein always ends in this position independent of its initial orientation in the simulation box.
After 500 ns of simulation, clear differences in the binding mode between AtCP and GgCP proteins and the POPA/POPC lipid bilayer were observed (Figure 4 and Figure S2). We found that the binding of AtCP to membranes composed from POPA/POPC is dependent on the concentration of POPA and on the PA charge, −1 or −2. In the case of POPA with a charge −1, AtCP only binds membranes with a high content of POPA (50%). By contrast, AtCP binds to membranes comprising 20% POPA with the charge −2 (Figure 4B), but not to 10% POPA. In all positive cases, AtCP binds the membrane via the α tentacle (Figure 4B and Figure S2A). Moreover, and in good agreement with docking results, residues from the positively-charged patch of the α tentacle (K273, R276, K277, K278 and R283) interact with POPA (Figure 5A). Furthermore, the amphipathic helix at the very end of the α tentacle is embedded in the membrane (Figure 4B) via its hydrophobic residues (Figure 4B, L279, V281, L285, F286 and W288). On the other hand, GgCP binds membranes containing POPA solely via the β tentacle (Figure 4C) and interacts with the membrane mainly by nonpolar contacts (Figure 5C).
To study the mode of CP binding to POPIP2/POPC membranes, we used two different concentrations of POPIP2 (1 and 5%). AtCP interacts with 5% POPIP2 membranes with both tentacles (Figure 4E) and, similarly to POPA, the majority of polar interactions are mediated by the positively-charged region on the α tentacle (Figure 5B). However, we observed a decreased number of nonpolar contacts between AtCP and membranes containing 5% POPIP2/POPC (Figure 5B) compared to 20% POPA/POPC (Figure 5A). This correlates very well with density profiles computed for these two simulated systems (Figure S3), where we found that the α tentacle is much more embedded into the hydrophobic part of the phospholipid bilayer comprising 20% POPA/POPC. Intriguingly, we did not found any preferential binding site when simulating AtCP with membranes containing 1% POPIP2 but rather observed that protein rotates closely to the membrane (Figure S2B). Conversely, we observed GgCP binding to membranes with both concentrations of POPIP2 (Table 1). The interaction of GgCP with membranes containing 5% POPIP2/POPC is mediated by both tentacles (Figure 4F and Figure 5D). Interestingly, we observed that the binding is mediated just by the α tentacle when we used a lower amount of POPIP2 in the membrane (1%, Figure S2C). We also performed self-assembly simulations and subsequent extension for conditions without any signaling lipid in the membrane; in this case we did not observe any binding between CP and POPC bilayers (Figure S2D).
In summary, we observed that AtCP differs from its vertebrate counterpart GgCP in the way it interacts with membranes containing POPA/POPC or POPIP2/POPC (Table 1). The interaction between AtCP – POPA/POPC membrane is mediated solely by the α tentacle and the binding is provided by the combination of polar and nonpolar interactions (Figure 4B and Figure 5A). On the other hand, GgCP interacts with the lipid bilayer containing POPA/POPC with the β tentacle and the interaction seems to be mediated preferentially by nonpolar contacts (Figure 4C and Figure 5C). The interaction of either AtCP or GgCP with the membrane consisted of POPIP2/POPC is mediated by both tentacles (Figure 4E and 4F), although there are also significant differences in the POPIP2 binding by AtCP and GgCP. In particular, the longer β tentacle of GgCP provides more nonpolar contacts with the POPIP2-containing bilayer in comparison with AtCP (Figure 5B and 5D).
To further confirm the importance of the α tentacle for association of AtCP with POPA/POPC membranes, we performed in silico mutagenesis of two residues with the greatest number of polar (CPα-K278A and CPα-R283A) as well as for the two most important nonpolar contacts (CPα-F286S and CPα-W288S). We simulated three 500 ns runs of CG-MD as described above and computed minimal distances between AtCP and membrane during these simulations. As shown in Figure 6A, wild-type AtCP always remains closely associated with the membrane. On the other hand, mutation of the polar residue K278 to alanine leads to complete disruption of AtCP-POPA/POPC association. Similar but weaker effects can be observed for the CPα-R283A mutation. Interestingly, CPα-W288S mutation was also able to disrupt binding of AtCP to the POPA/POPC membrane, although not in every run. On the other hand, we did not observe any effect caused by mutation of CPα-F286S. We also performed analogous simulations for the mutated AtCP proteins with POPIP2/POPC membranes (Figure 6B). In this case, we found that only mutation of W288 has an effect on the association of AtCP with the membrane. Collectively, these results further confirm the critical importance of the CP α tentacle for PA binding that is mediated by interaction site containing positively charged residues K278 and R283. The effect of the W288S mutation on both POPA/POPC and POPIP2/POPC-binding supports the hypothesis of structural importance of W288 (homologous to W271 in GgCP) for stability of the α tentacle as proposed by Kim et al [6].
Previously, we described dissociation constant (Kd) values for plant and mouse CP binding to PA and PIP2 micelles, as analyzed by changes in endogenous tryptophan fluorescence [10]. The findings show that AtCP has a somewhat higher apparent affinity for PIP2 micelles than for PA (11 µM versus 17 µM, respectively). The apparent affinities of the animal protein for PA and PIP2 are markedly different, with mouse CP showing a higher affinity for PIP2 (8 µM for PIP2 versus 59 µM for PA). Here, we employed the potential of mean force (PMF) calculation with the umbrella sampling protocol [34] to gain insight into the quantitative aspects of the computed interactions. We used steered molecular dynamics to pull the protein away from the membrane and to generate sampling windows for PMF calculation. For this type of pulling experiment, we applied position restraints on the lipids to keep them in the membrane. Figure 7 shows PMF curves for four selected systems. We found that GgCP interacts most tightly with membranes containing 5% PIP2/POPC with ΔG −236 kJ/mol. AtCP interacts with membranes of the same composition with ΔG −185 kJ/mol. In comparison to GgCP (ΔG −69 kJ/mol), AtCP interacts more strongly with membranes composed from 20% POPA/POPC (ΔG −112 kJ/mol). Importantly, this is a similar trend compared to the experimental data; there is a huge difference between the binding of PA and PIP2 for GgCP and a much smaller difference in the case of AtCP.
A direct alignment of the primary sequences for the C-terminal tentacles from CP proteins across diverse eukaryotes (Figure 8A and Figure S4) revealed that although the positively-charged region located on the α tentacle is generally well conserved, several lineage-specific differences could be identified, which might explain distinct binding properties of AtCP and GgCP. Plant sequences generally have longer α tentacles (Figure 8A) with a conserved lysine (K278, in GgCP this is Q261), that shows the greatest number of polar contacts with PA (Figure 5A). Moreover, plant α tentacles contain leucine, proline and asparagine (L285, P287 and N290) instead of lysine, aspartate and lysine in vertebrate sequences (K268, D270 and K273), resulting in a decrease of polar residues in this region compared to animal CP. These amino acid changes facilitate the observed embedding of the plant α tentacle into PA-containing membranes (Figure 8B and Figure S3). Intriguingly, higher plants also have a shorter β tentacle and thus lack a major part of the amphiphatic helix located at this position in vertebrate CP (Figure S4).
To further confirm whether the AtCP α tentacle constitutes a PA-binding domain, we prepared a recombinant fusion protein between GST and the C-terminal 38 amino acids from AtCP α subunit (GST-CPα-Cterm). Protein-lipid overlay assays showed strong binding of the GST-CPα-Cterm to PA (Figure 8C), similar to our previous observations with full-length AtCP protein [10]. In addition, the interaction of GST-CPα-Cterm with a subset of PPIs including PIP2 and phosphatidylinositol (3,4,5)-trisphosphate (PIP3), as well as with cardiolipin and sulfatidate was also observed in this assay. Interestingly, cardiolipin and sulfatidate contain a phosphate/sulphate group and thus resemble PA and PPIs to some extent. However, the binding of PIP3, cardiolipin and sulfatidate to GST-CPα-Cterm is most probably non-physiological, as PIP3 is not present in plant membranes and cardiolipin is found only in bacteria and in the inner membrane of mitochondria. We also found that GST-CPα-Cterm binds to lipid vesicles containing 20% PA and PC in co-sedimentation experiments (Figure 8D). These two complementary approaches clearly demonstrate that the AtCP α tentacle is sufficient for PA binding.
We previously described different binding affinities for plant and animal CP interacting with two distinct signaling phospholipids, PA and PIP2 [10]. Here, we focused on the structural aspects of these interactions by employing diverse methods of structural bioinformatics. It has been shown that these methods, and particularly CG-MD simulation, can play a crucial role in our understanding of general principles of processes such as lipid bilayer formation, peptide segregation into raft-like structures in the membrane, and characterization of protein-lipid interactions with both integral- and peripheral-membrane proteins [28]. Recently, the combination of homology modeling and CG-MD was used to investigate interactions between diverse voltage sensors and lipid bilayers [35]. Initial all-atom MD studies done on GgCP, in the absence of membranes, revealed that the α tentacle is rather immobile and remains stationary on the protein surface during the simulation [36]. This immobility is mainly stabilized by the interaction of W271 of the amphiphatic helix with the core of the animal protein. Interestingly, we observed that the homologous tryptophan in AtCP (W288), together with other hydrophobic residues of the α tentacle, is embedded into the membrane after 500 ns MD simulation (Figure 8B). These data support the hypothesis of Wear and Cooper [37], that proposes the induction of α tentacle mobility by non-ionic detergent. We suggest that a lipid bilayer could have a similar effect on the mobility of the α tentacle and facilitate embedding of hydrophobic residues.
In this report, we describe differences between AtCP and GgCP for both C-terminal tentacles (Figure 8A and Figure S4), which may reflect distinct properties of CP–actin interaction between organisms. Alternatively, given that plant cells contain 10- to 100-fold lower amounts of PIP2 than PA [38], [39], one can speculate that differences in the tentacles is an adaptation to the distinct levels of PA and PIP2 in mammals and plants, i.e. increased binding properties of AtCP towards PA. As discussed above, we observed the embedding of the AtCP α tentacle into membranes containing PA. Consistent with this observation, we found a decreased number of polar residues in this tentacle. It is important to note that this difference is rather subtle, but mutations leading to a more nonpolar α tentacle could reduce actin binding [6]. We also observed that plant CPs have a shorter β tentacle and thus they lack the majority of the amphiphatic helix located in this region (Figure S4). We hypothesize that the PA- and actin-binding properties of plant CP have co-evolved to keep the right balance between actin regulation and responses to lipid signaling.
Kooijman et al. [18] described remarkable properties of PA and proposed a model for the electrostatics/hydrogen bond switch, where arginine and lysine residues on binding peptides can increase the charge of PA to −2. The authors also performed all-atom MD simulation of K8 and R8 peptides with bilayers formed from DOPC/DOPA and found that simulations where DOPA had charge −2, were in better agreement with experimental results. In our simulations, we observed the dependence of AtCP binding on the charge of PA, but it is important to note, that when we observed the interaction, the binding mode was very similar for each system regardless of the PA charge (Figure 4B and Figure S2A). Moreover, PA has a unique cone shape under physiological conditions and it has been proposed that PA could facilitate the insertion of hydrophobic protein domains into a bilayer [18]. Consistent with this hypothesis, we observed insertion of hydrophobic parts of the AtCP α tentacle into membranes containing PA (Figure 8B and Figure S3).
In our CG-MD simulations with membranes containing 5% POPIP2 and POPC, we observed the involvement of both tentacles with either animal or plant CP (Figure 4E, F and Figure 5B, D), suggesting cooperativity between both tentacles. When we simulated the system containing 1% POPIP2/POPC, we found that GgCP binds the phospholipid bilayer preferentially by the α tentacle (Table 1). Altogether, these results clearly show the importance of a positively-charged patch located on the α tentacle in both AtCP and GgCP. This region corresponds to lipid-binding site identified by Kim et al. [11]. We did not observe the involvement of the second putative PIP2-binding site proposed by Kuhn and Pollard [12]. Moreover, the latter positively-charged region is completely lacking in AtCP.
Importantly, we obtained very similar quantitative trends for the interactions studied herein when compared to experimental approaches [10]. We found a much smaller difference between the binding of PA and PIP2 by AtCP when compared to GgCP. The energies of the interactions computed from experimentally determined Kd values vary from −24 to −29 kJ/mol, whereas from the umbrella sampling protocol, we computed the energy ranging from −62 to −236 kJ/mol. These discrepancies could be explained by different composition of the membrane (experimental Kds were determined for the system with just one phospholipid, i.e. PA or PIP2, and the lipids were in micelles rather than bilayers).
The most recent information on CP–actin interactions comes from a study by Kim et al. [6], who combined computational approaches with a large scale site-directed mutagenesis. They propose a model in which GgCP interacts with actin mainly via its tentacles and faces the actin filament barbed end with the top of the mushroom structure. The authors identified 49 residues of mammalian CP (18 on CPα and 31 on CPβ). They mutated 45 of these residues and found that only 10 showed more than a 3-fold increase in Kd. A direct comparison of these residues between GgCP and AtCP shows that 7 residues are highly conserved (these residues include CPα-E200, CPα-K256, CPα-R260, CPα-K268, CPβ-R195, CPβ-K223 and CPβ-R225 of mammalian CP). Interestingly, AtCP completely lacks nonpolar residues located on the β tentacle (L258, L262, L266) which are responsible for the interaction with the hydrophobic cleft in actin. In our computed modes of the CP-membrane interaction, we observed that CP binds membranes mainly via its tentacles. Therefore, it is tempting to speculate that steric hindrance imposed by CP–membrane binding prevents actin binding. Interestingly, GgCP bound to the PA-containing membrane has the α tentacle and the top of the mushroom-like structure unoccupied (Figure 4C). This could be an explanation why PA has not been described as an inhibitor of the activity of the animal CP [8], [9].
In summary, our results provide structural insight into the regulation of CP by two signaling phospholipids, PA and PIP2. A prominent role for the α and β C-terminal tentacles located on the top of the CP structure is apparent. We have shown differences of PA and PIP2 binding between AtCP and GgCP explaining published experimental data. Our results represent a comprehensive view of the interaction between CP and PA- or PIP2-containing membranes and reveal the mode of binding with structural implications for CP regulation. We also identified the PA-binding domain of AtCP and experimentally showed that it is sufficient for binding membranes in vitro. Our results call for intensive future research involving, in particular, a detailed mechanistic description of the phospholipid-induced uncapping of actin filaments. We also suggest that it would be relevant to examine the possible synergistic effects of distinct phospholipids on the inhibition of CP activity.
CP protein sequences were identified by gapped BLAST or PSI-BLAST [24] searching against the non-redundant protein database at the National Center for Biotechnology Information (http://blast.ncbi.nlm.nih.gov/Blast.cgi) using Arabidopsis annotated sequences with default settings. In addition, blast searches were conducted using Phytozome web page and DOE Joint Genome Institute (http://www.phytozome.net/; http://www.jgi.doe.gov/). In most cases, the search parameters were set at the default values; however, occasionally, modifications were used (the changed parameters included mostly length of the word and type of scoring matrice). Putative genes were initially identified based on the automatic annotation at the aforementioned databases. Since gene models based on computer annotations often contain errors, exon-intron structures were manually curated with the aid of experimentally-verified sequences or sequences from closely related species.
Multiple alignments were constructed with mafft algorithm (in einsi mode) [40] and manually adjusted. Maximum likelihood method using PhyML program [41] was employed for phylogeny inference with the WAG matrix, Γ-corrected among-site rate variation with four rate site categories plus a category for invariable sites, all parameters estimated from the data. Bayesian tree searches were performed using MrBayes 3.1 [42] with a WAG amino acid model, where all analyzes were performed with four chains and 1 000 000 generations per analysis and trees sampled every 100 generations. All four runs asymptotically approached the same stationarity after first 500 000 generations which were omitted from the final analysis. The remaining trees were used to infer the posterior probabilities for individual clades.
A homology model for AtCP was built on the X-ray structure for GgCP (rcsb 1IZN). The manually edited alignment obtained by PSIPRED [43] was used as input for MODELLER 9v8 [44]. As template contains shorter C-terminus of α subunit, residues ranging from 288 to 302 were forced to α-helix formation according to secondary structure prediction. The best model was selected on the energy and constraint violation values of MODELLER and further evaluated by PROSA and WHAT IF algorithms [45], [46]. APBS program [47] was used to compute electrostatic potential of CP.
To simulate self-assembly of lipid bilayers in the presence of protein, the MARTINI CG force field was used [21], [22]. The protein was described according to ELNEDIN representation [48] with Rc 0.9 nm and K 500 kJ·mol−1·nm−2. CG model for POPIP2 molecule was prepared according to [49]. GROMACS 4.0.5 was used for all MD simulations [50]. Lenard-Jones and electrostatic interactions were shifted to 0 between 9 and 12 Å and between 0 and 12 Å, respectively. A relative dielectric constant of 15 was used. Simulations were run in NPT ensemble. The temperature of protein, lipids, and solvent was coupled separately at 310 K using the Berendsen algorithm, with a coupling constant 1.0 ps. The system pressure was coupled using the same algorithm with a coupling constant 3.0 ps, compressibility of 3·10−5 and reference pressure 1 bar. Simulations were performed using a 20 fs integration time step. The protein, lipids and water were placed randomly in the simulation box. Na+ ions were added to ensure electroneutrality of the system. The whole system was energy-minimized using steepest descent method up to maximum of 500 steps and production runs were performed. In cases where some lipids remained apart from the lipid bilayer, CG water particles were used to replace them and the whole system was again energy-minimized. These systems or the final states of self-assembly were subsequently prolonged under the same conditions as self-assembly simulations. All simulations were repeated 3–5 times.
The final configurations of four selected systems were used as inputs for the pulling experiments. The simulation box was extended in the z direction to capture the proposed trajectory of the pulling. Additional CG water particles were added to this extended space. The extended system was energy-minimized and short simulation for 50 ns was run. The CP was extracted from the membrane by applying a constraint force to the centre of mass (COM) of the protein in a direction coincident with z axis. Lipid molecules were restrained by position restraints during the pulling experiment (kpr = 1000 kJ mol−1 nm−2). CP was pulled at a rate of 0.5 nm ns−1 and COM pulling was carried out until the COM of CP was 4 nm apart from COM of the lipid bilayer. Snapshots along the pulling trajectory were extracted at COM spacing of 0.1 nm to generate starting configurations for umbrella sampling windows. For umbrella sampling calculation, we used approx. 40 windows from the pulling experiment described above. All generated configurations (windows) were equilibrated for 50 ns before PMF calculation. Afterwards, for each window a 100 ns long simulation was performed with the biasing potential applied to restrain COM of CP in a required distance from COM of the lipid bilayer. PMF curves were obtained using the WHAM algorithm [51].
It is important to note that times reported in this study are computational times. It was shown that effective times for CG simulations are longer; for proteins and lipids in MARTINI force field, the speed up factor is about four-fold [52], i.e. 500 ns simulation time would correspond to 2 µs real time.
The C-terminus of AtCP α subunit (AtCPα-Cterm, aa 270–308) was amplified by PCR using Phusion DNA polymerase (Finnzymes) and cloned into the pGEX-KG vector. The resulting plasmid (GST–AtCPα-Cterm) was transformed into Escherichia coli strain BL21 and cells were grown overnight at 37°C. After sub-culturing into fresh medium, cells were grown at 37°C to an OD600 of approximately 1.5, then induced for 4 h with 0.4 mM isopropyl thio-β-D-galactoside. Recombinant proteins were purified on glutathione-Sepharose (GE Healthcare) according to the manufacturer's instructions. Protein-lipid overlay assays with membrane lipid strips (Echelon) were performed according to manufacturer's instructions with protein concentration 0.5 µg/ml. To detect lipid binding in vesicles, we used the procedure described by [18] with slight differences; binding buffer comprised 125 mM KCl, 25 mM Tris, pH 7.8, 1 mM dithiothreitol and 0.5 mM EDTA. To reveal lipid binding, we incubated 400 nmol of lipids with 1 µg of GST-tagged protein.
|
10.1371/journal.pntd.0005893 | Molecular genomic characterization of tick- and human-derived severe fever with thrombocytopenia syndrome virus isolates from South Korea | Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne viral disease caused by the SFTS virus (SFTSV) from Bunyaviridae that is endemic in East Asia. However, the genetic and evolutionary characteristics shared between tick- and human-derived Korean SFTSV strains are still limited.
In this study we identify, for the first time, the genome sequence of a tick (Haemaphysalis longicornis)-derived Korean SFTSV strain (designated as KAGWT) and compare this virus with recent human SFTSV isolates to identify the genetic variations and relationships among SFTSV strains. The genome of the KAGWT strain is consistent with the described genome of other members of the genus Phlebovirus with 6,368 nucleotides (nt), 3,378 nt, and 1,746 nt in the Large (L), Medium (M) and Small (S) segments, respectively. Compared with other completely sequenced human-derived Korean SFTSV strains, the KAGWT strain had highest sequence identities at the nucleotide and deduced amino acid level in each segment with the KAGWH3 strain which was isolated from SFTS patient within the same region, although there is one unique amino acid substitution in the Gn protein (A66S). Phylogenetic analyses of complete genome sequences revealed that at least four different genotypes of SFTSV are co-circulating in South Korea, and that the tick- and human-derived Korean SFTSV strains (genotype B) are closely related to one another. Although we could not detect reassortant, which are commonly observed in segmented viruses, further large-scale surveillance and detailed genomic analysis studies are needed to better understand the molecular epidemiology, genetic diversity, and evolution of SFTSV.
Full-length sequence analysis revealed a clear association between the genetic origins of tick- and human-derived SFTSV strains. While the most prevalent Korean SFTSV is genotype B, at least four different genotypes of SFTSV strains are co-circulating in South Korea. These findings provide information regarding the molecular epidemiology, genetic diversity, and evolution of SFTSV in East Asia.
| Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne viral disease caused by the SFTS virus (SFTSV). During entomological surveillance of SFTSV infection in Korean ticks collected from SFTS outbreak areas, we isolated a single SFTSV strain which we designated KAGWT. In addition, we isolated three SFTSVs from human patients with typical SFTS symptoms. In this study, we report the genomic sequences of each of these isolates and compare the genetic and evolutionary characteristics between tick- and human-derived Korean SFTSV isolates. Genetic and phylogenetic analyses of these sequences revealed that the tick-derived Korean SFTSV strain is clustered into genotype B, the most prevalent genotype in South Korea, and was closely related to other SFTSV in the same group. Furthermore, our results show that at least four different genotypes of SFTSV strains are co-circulating in South Korea.
| Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne viral disease characterized by fever, gastrointestinal symptoms, leukopenia, and thrombocytopenia. It was first reported in China in 2010 [1] and was subsequently identified in South Korea and Japan in 2013 [2–4]. The causative agent, SFTS virus (SFTSV), belongs to the genus Phlebovirus in the family Bunyaviridae [1]. Other novel tick-borne phleboviruses, including Heartland virus (HRTV), Malsoor virus (MV) and Hunter Island Group virus (HIGV), which are genetically related to but distinctly different from SFTSV, have been isolated from humans and ticks in the United States [5, 6], bats in India [7] and ticks in Australia [8], respectively.
Like other members of the genus Phlebovirus, SFTSV contains a tripartite RNA genome consisting of three single-stranded RNA segments of negative polarity, designated large (L), medium (M), and small (S). The L, M, and S segments encode the RNA-dependent RNA polymerase (RdRp), the viral envelope glycoproteins (Gn and Gc) and both a nucleoprotein (NP) and a nonstructural protein (NSs) in an ambisense orientation, respectively [1, 9].
Although human-to-human transmission of SFTSV through contact with blood and/or body secretions of patients has been reported [10–12], the virus is generally transmitted to humans by tick bites. Several studies have reported SFTSV isolation or detection from tick species including Haemaphysalis longicornis, Rhipicephalus microplus, H. flava, H. concinna, Amblyomma testudinarium, and Ixodes nipponensis [1, 13–18]. Further, this virus has also been isolated or detected from domestic animals (e.g., cattle, goats, dogs, chickens and cats), small mammals such as rodents and shrews [19–22], and reptiles [23].
Since the first reported fatal case in 2012 in Gangwon Province in South Korea [2], concern regarding SFTSV has grown as the numbers of SFTS patients has increased annually with 36 cases reported in 2013, 55 cases in 2014, 79 cases in 2015, and 165 cases in 2016 [24]. Moreover, the mean mortality rate of these cases was approximately 21.8%. During our previous survey of carrier ticks from affected areas, a single SFTSV strain (designated KAGWT) was initially isolated from H. longicornis nymphs collected from the Samcheok-si, Gangwon Province [15]. In addition, we also isolated SFTSV from two recovered- and one fatal-human cases that presented with typical SFTS symptoms. In this study, we analyzed the whole genome sequence of the first tick-derived Korean SFTSV and human SFTSV strains to compare the genetic and evolutionary characteristics between tick- and human-derived Korean SFTSV isolates. Genetic characterization revealed that the tick-derived SFTSV is closely associated with recent KAGWH3 Korean human isolates and that at least four different genotypes of SFTSV strain are co-circulating in South Korea. Taken together, our results suggest that more intensive and continuous surveillance of SFTSV is essential for better understanding of the molecular epidemiology, genetic diversity, and evolution of this virus in East Asia.
Chungbuk National University Hospital received written consent for sample collection from each patient with SFTSV infection. All participants were adults and this study was approved by the institutional review board (IRB) of Chungbuk National University Hospital (IRB no. 2015-08-009-001).
The KAGWT strain was isolated from H. longicornis ticks collected from the Samcheok-si, Gangwon Province in South Korea as described previously [15]. The CB1, CB2, and CB3 strains were isolated from the sera of SFTS patients who were hospitalized with typical SFTS symptoms at Chungbuk National University Hospital.
For virus propagation, the virus was passaged five times on confluent monolayers of Vero E6 cells (ATCC No. CRL-1586; American Type Culture Collection, Manassas, VA) in Dulbecco’s Modified Eagle Medium (DMEM; Gibco, Grand Island, NY) containing 8% fetal bovine serum (FBS; Gibco) with penicillin (100IU/mL) and streptomycin (100μg/mL; P/S, Gibco) placed in 37°C incubator supplemented with 5% CO2. Cell culture supernatant was collected after seven days and stored at -80°C as the working virus stock for whole genome sequencing.
Viral RNA was extracted from 140 μL of the viral stock using QIAamp Viral RNA Mini Kits (Qiagen, Hilden, Germany) according to the manufacturer’s instructions. Single strand cDNA was synthesized using viral RNA and primers specific for each segment using a cDNA synthesis kit (Cosmogenetech co, Ltd., Seoul, Korea) according to the manufacturer’s instructions.
For the whole genome sequencing, one to six overlapping PCR fragments covering the SFTSV genome containing full-length L, M, and S segments were amplified by PCR from cDNA using SP-Taq DNA polymerase (Cosmogenetech, Seoul, Korea). PCR for each segment was performed at 95°C for 5 minutes, followed by 45 cycles of amplification consisting of 95°C for 30 seconds, annealing at 53 or 60°C for 30 seconds according to the primer sets for each segment, and 72°C for 2 minutes, with a final extension at 72°C for 5 minutes. The non-coding 5′ and 3′ ends of the viral genome were determined by rapid amplification of cDNA ends (RACE) method as described previously [25]. PCR products were then purified using a QIAquick Gel Extraction Kit (Qiagen) according to the manufacturer’s instructions and direct sequenced using ABI Prism BigDye Terminator Cycle Sequencing Kits (Applied Biosystems, Foster City, CA) and an ABI 3730x1 sequencer (Applied Biosystems) at Cosmogenetech Co, Ltd. The nucleotide sequences obtained from this study were assembled using the SeqMan program in the DNASTAR software (version 5.0.6; DNASTAR Inc., Madison, WI) to determine the complete genomic sequence.
Genetic and phylogenetic analyses were conducted by aligning published full-length sequences of SFTSV obtained from China, Japan and Korea isolates, which are available in GenBank, together with the closely related SFTSV sequences obtained from the basic local alignment search tool results. A total of 41 full-length sequences of the L, M, and S segments, including the strains isolated in this study (Table 1), were included in the analyses.
Multiple sequence alignments were performed using the Clustal W algorithm in DNASTAR version 5.0.6 or MEGA version 6.0 [27]. The aligned nucleotide and deduced amino acid sequences were analyzed using the MegAlign program of DNASTAR to compare the sequence homologies and amino acid substitutions.
Phylogenetic analyses were performed based on the full-length L, M, and S segments of SFTSVs. Phylogenetic trees were constructed with MEGA version 6.0 software using the Maximum Likelihood (ML) method based on the Kimura 2-parameter model. The reliability of the ML tree was evaluated by the bootstrap test with 1,000 replications.
The full-length L, M, and S segment sequences of tick- and human-derived strains determined in this study have been deposited in GenBank under the following accession numbers: KAGWT (KY273136 to KY273138), CB1 (KY789433, KY789436, and KY789439), CB2 (KY789434, KY789437, and KY789440), and CB3 (KY789435, KY789438, and KY789441) strains.
The genomes of tick- and human-derived Korean SFTSV strains analyzed in this study were consistent with 6,368 nucleotide (nt), 3,378 nt and 1,746 nt present in the L, M and S segments respectively, consistent with what has been reported for other members of the genus Phlebovirus [9]. The L segment encodes a 6,255 nt long ORF [2,084 amino acids (aa)] for the RNA-dependent RNA polymerase gene, the M segment comprises a 3,222 nt precursor of the glycoprotein gene, coding for Gn (516 aa) and Gc (511 aa) proteins, while the S segment contains 882 nt and 738 nt long ORFs, which translate into a nonstructural protein (293 aa) and a nucleoprotein (245 aa), respectively. The non-coding regions (NCRs) of the L, M and S segments at the 5′ termini are 16 nt, 18 nt and 43 nt, respectively; and at the 3′ termini are 97 nt, 138 nt and 29 nt, respectively. As shown in Fig 1, complementary sequences within the 5′ and 3′ NCRs of the three segments are highly conserved between tick- and human-derived Korean SFTSV strains.
Alignment and pairwise comparisons of each segment between the tick-derived KAGWT strain and other fully sequenced human-derived Korean SFTSV strains showed nucleotide (and deduced amino acid) sequence homology ranging from 95.9 (99.3) to 99.9 (100), 94.0 (98.4) to 99.8 (99.8), and 94.9 (98.3) to 99.9 (100)% for L, M and S segments, respectively (Table 2). This high homology between the tick-derived KAGWT and other human-derived SFTSV strains reflects the close level of relatedness of these strains to each other.
In particular, the KAGWT strain showed the highest sequence identities at the nucleotide and deduced amino acid levels with the KAGWH3 strain isolated in 2014 from the serum of a patient from Gangwon Province, South Korea who experienced typical SFTS symptoms [25]. These results indicate that tick- and human-derived Korean SFTSV strains are most closely related with one another. Moreover, the deduced amino acid sequence of KAGWT revealed two amino acid variations (A66S, I89V) in the Gn protein compared with the KAGWH3 strain. In particular, one unique amino acid substitution in the Gn protein, at position 66 (alanine to serine), was found only in the KAGWT strain compared with other human SFTSVs. According to previous reports, a change from phenylalanine to serine at position 330 in the M segment polyprotein (F330S) occurred in cell culture-adapted SFTSV and resulted in the large-focus phenotype [28]. However, all SFTSV strains analyzed in this study had a conserved amino acid sequence (F) at position 330 of the M segment.
Compared with the other fully sequenced SFTSV strains, KAGWT showed higher identity with the genotype B SFTSV strains than with the other genotypes at each segment as shown in S1 Table. The nucleotide sequence identities between the tick-derived KAGWT and other SFTSV strains from China, Japan, and South Korea belonging to the genotype B were 96.4 to 99.9% (L), 95.8 to 99.8% (M), 94.9 to 99.9% (NP) and 95.8 to 99.8% (NSs) similar, respectively. However, KAGWT showed relatively low nucleotide sequence identity ranging from 95.7 to 96.3% (L genes), 93.6 to 94.4% (M genes), 95.1 to 96.2% (NP genes) and 94.8 to 95.9% (NSs genes), respectively with other genotypes of SFTSV (Table 2 and S1 Table).
In comparison with the other fully sequenced SFTSV strains, CB2 was identified as the first genotype A SFTSV strain out of 14 full-length sequenced Korean SFTSV isolates. Genetic comparison results showed that the CB2 strain had nucleotide sequence identity ranging from 97.9% (L), 97.9 to 98.0% (M), 98% (NP) and 97.2% (NSs), respectively with the other genotype A SFTSV strains circulating in China. However, the CB2 strain showed relatively low nucleotide sequence identity ranging from 96.1 to 96.7% (L genes), 93.5 to 96.2% (M genes), 95.3 to 97.6% (NP genes) and 93.9 to 96.8% (NSs genes), respectively with other SFTSV strains (Table 2 and S1 Table).
To investigate the genetic evolutionary origins and relationship between the tick- and human-derived Korean SFTSV strains, phylogenetic analyses were conducted with full-length complete sequences of previous SFTSV isolates from China, Japan, and South Korea. Phylogenetic analyses of complete genome sequences (L, M, and S segments) indicated that the tick-derived Korean SFTSV strain (KAGWT) was clustered with the genotype B SFTSV strains circulating in humans in China, Japan, and South Korea (Figs 2–4), although KAGWT also exhibited a close relationship with the KAGWH3 human isolate [25].
Furthermore, the CB1 and CB3 human isolates also clustered with genotype B viruses although they are separated into a different sub-node from tick-driven Korean SFTSV strains (KAGWT) and are closely related with recent Korean SFTSV human isolates, KADGH and KACNH3. In addition, the CB2 human isolate was clustered together with recent China genotype A viruses.
Overall the phylogenetic trees revealed that Korean SFTSV strains can be classified with high branch support into four distinct genotypes (designated A, B, D, and F) out of the six genotypes described previously [26]. To investigate the prevalence of each genotype of SFTSV, we analyzed the SFTSV sequences available in the GenBank database. As shown in Table 3, most SFTSV strains from South Korea belong to genotype B (11 out of 14 isolates) and only one isolate each was reported from genotype A, D, and F. It is noteworthy that the CB2 is the first report case of a genotype A virus in South Korea. Isolates from China were diverse with all six genotypes being represented, although the most prevalent was genotype F followed by genotypes A and D (Table 3). In contrast, only three SFTSV strain genotypes (B, C, and D) observed in Japan and the most prevalent genotype was B.
SFTS is an emerging tick-borne infectious disease caused by a novel Phlebovirus, SFTSV that is highly endemic in China, Japan, and South Korea [1, 2, 4].
In this study, we determined the whole genome sequences of the first tick and patient-derived Korean SFTSV strains and compared them with other available whole genomic SFTSV sequences. To the best of our knowledge, this is the first report of the whole genomic sequence of an SFTSV strain isolated from ticks and of a genotype A SFTSV strain collected from South Korea. Comparison of the whole genome sequence of tick- and human-derived strains revealed that all SFTSVs consist of three segments with 6,368 nt in the L segment, 3,378 nt in the M segment, and 1,746 nt in the S segment. Thus, the genome organization of Korean strains are consistent with the known genome organization of SFTSV belonging to the Phlebovirus genus [9]. Complementary sequences within the 5′ and 3′ NCRs of the three segments analyzed in this study are highly conserved, as in other SFTSV strains (Fig 1). According to previous studies of Bunyaviruses, these complementary sequences form panhandle-like structures and may have important roles in transcription initiation, viral RNA replication, viral RNA encapsidation, and viral genome packaging [29, 30]. All Korean SFTSV strains have 53 unique amino acid variations in the L, Gn, Gc, N, and NSs proteins compared with other SFTSV strains. Among them, one unique amino acid substitution in the Gn protein at position 66 (alanine to serine) was found only in the tick-driven KAGWT strain. Since little information is known about SFTSV, additional research regarding the association of amino acid variations in each gene product with the pathogenesis of SFTSV infection and detailed molecular studies utilizing reverse genetics will be required to further explain the functional role of these unique substitutions and in particular, the Gn protein substitution (66S).
Genetic and phylogenetic analysis revealed that the tick-driven KAGWT strain isolated from H. longicornis nymphs in 2013 [15] has high sequence identity in each segment and is closely related with the KAGWH3 strain isolated in 2014 from the serum of an SFTS patient who experienced high fever, vomiting, diarrhea, and fatigue. This is not surprising given that both viruses were from the Samcheok-si, Gangwon Province [25], and further suggests that tick- and human-derived Korean SFTSV strains are closely related to each other.
Although several studies about SFTSV classification have been reported [26, 31–34], the uniform classification of SFTSVs has not yet to be established. Therefore, to use unified nomenclature of SFTSV genotypes, more mutual understanding and discussion are needed. In this study, we adapted the nomenclature as previously described by Fu et al. [26] which described results of most large numbers (205 SFTSV strains) and complete sequences. That paper showed SFTSV strains can be divided into six distinct genotypes. Through full-length sequence analysis of human-driven SFTSV isolates we detected the first genotype A strain (CB2) from a patient exhibiting severe SFTSV-like symptoms in the Chungbuk Province of South Korea in 2015. The Korean strains belong to four of these (genotypes A, B, D, and F) of which genotype B strains appear to be predominant. It should be noted that in China all six genotypes of SFTSV have been reported, while only one isolate was reported for each genotype A, D and F in Korea. Thus, further detailed surveillance of tick- and human-derived SFTSVs are needed to understand the actual prevalence and possible transmission of each genotype of SFTSV strain into South Korea. Due to the segmented nature of the Phlebovirus, genetic reassortment is known to be an important process resulting in the sequence diversity necessary for viral evolution [35]. Previous studies have shown reassortment occurs in Phleboviruses including Hantavirus [36, 37], Rift Valley fever virus (RVFV) [38], and SFTSV [26, 31–34]. Although SPL087A and AHL/China/2011 SFTSV strains isolated in Japan and China, respectively, were identified as reassortants which contained different genotype segments in the same SFTSV strain [26, 31–34], no reassortant was identified in the Korean SFTSV strains tested in this study. Therefore, continuous analysis consisting of in-depth surveillance and genome sequencing will be needed in South Korea to obtain more detailed information of the molecular epidemiology, genetic diversity, and evolution of SFTSV.
In conclusion, we report here the first whole genomic sequence of the tick-derived SFTSV KAGWT strain isolated from the ticks in the Gangwon province of South Korea and the comparison of the genetic characteristics of this virus with recent human SFTSV isolates. Genetic and phylogenetic analyses with full-length genome sequences revealed that tick- and human-derived Korean SFTSV strains are closely related to one another and that at least four different genotypes of SFTSV are co-circulating in South Korea. These results provide insight into the genetic origins of human of SFTSV strains as well as shed light on the molecular epidemiology, genetic diversity, and evolution of SFTSV. Furthermore, these associations will have important implications for the design of diagnostic procedures and vaccines against SFTSV.
|
10.1371/journal.ppat.1003283 | ATM and ATR Activities Maintain Replication Fork Integrity during SV40 Chromatin Replication | Mutation of DNA damage checkpoint signaling kinases ataxia telangiectasia-mutated (ATM) or ATM- and Rad3-related (ATR) results in genomic instability disorders. However, it is not well understood how the instability observed in these syndromes relates to DNA replication/repair defects and failed checkpoint control of cell cycling. As a simple model to address this question, we have studied SV40 chromatin replication in infected cells in the presence of inhibitors of ATM and ATR activities. Two-dimensional gel electrophoresis and southern blotting of SV40 chromatin replication products reveal that ATM activity prevents accumulation of unidirectional replication products, implying that ATM promotes repair of replication-associated double strand breaks. ATR activity alleviates breakage of a functional fork as it converges with a stalled fork. The results suggest that during SV40 chromatin replication, endogenous replication stress activates ATM and ATR signaling, orchestrating the assembly of genome maintenance machinery on viral replication intermediates.
| All cells have evolved pathways to maintain the integrity of the genetic information stored in their chromosomes. Endogenous and exogenous agents induce mutations and other damage in DNA, most frequently during DNA replication. Such DNA damage is under surveillance by a complex network of proteins that interact with one another to signal damage, arrest DNA replication, and restore genomic integrity before replication resumes. Many viruses that replicate in the nucleus of mammalian host cells have evolved to disable or evade this surveillance system, but others, e.g. polyomaviruses like SV40, activate it and somehow harness it to facilitate robust replication of viral progeny. We have sought to determine how SV40 induces and deploys host DNA damage signaling in infected cells to promote viral chromosome replication. Here we present evidence that, like host DNA, replicating viral DNA suffers damage that activates surveillance and repair pathways. Unlike host replication, viral DNA replication persists despite damage signaling, allowing defective replication products to accumulate. In the presence of host DNA damage signaling, these defective viral products attract proteins of the host damage surveillance network that correct the defects, thus maximizing viral propagation.
| Faithful duplication of the genome is vital for cell proliferation. In metazoans, the consequences of inaccurate genome replication include cell death, premature aging syndromes, neuro-degeneration disorders, and susceptibility to cancer [1], [2]. The DNA damage signaling protein kinases ataxia telangiectasia-mutated (ATM) and ATM- and Rad3-related kinase (ATR), members of the phosphoinositide-3 kinase-like kinase (PIKK) family, act to ensure that cells with incompletely replicated or damaged DNA do not progress through the cell cycle [1]. ATM and DNA-dependent protein kinase (DNA-PK) respond primarily to DNA double strand breaks (DSB) that are associated with either Mre11/NBS1/Rad50 (MRN) [3] or Ku70/80 [4], respectively. Additionally, intracellular oxidation or alterations in chromatin structure can activate ATM kinase [5], [6]. In contrast, single-stranded DNA (ssDNA) bound by RPA activate ATR [7], [8]. When activated, ATM and ATR phosphorylate consensus SQ/TQ motifs in target proteins at sites of damage, e.g. the histone H2AX, which facilitates recruitment of repair proteins and activation of downstream kinases Chk1 and Chk2 that enforce the checkpoint [8], [9].
Failure to activate DNA damage checkpoints results in genome instability syndromes. Mutations in the human ATM gene can cause the cancer-prone disorder ataxia telangiectasia. Hypomorphic mutations in the ATR gene can cause the genomic instability disorder Seckel Syndrome, but complete loss of ATR results in cell death [10], [11]. The central roles of ATM and ATR in genome maintenance suggest the potential to manipulate their activity for cancer chemotherapy, fueling the development of potent small molecules that specifically inhibit ATM and ATR activities in cellulo [12], [13].
Interestingly, multiple animal viruses have evolved to manipulate DNA damage signaling pathways to facilitate viral propagation [14]. Some viruses, e.g. Herpes simplex, evade or disable DNA damage response pathways that result in inappropriate processing of viral DNA [15], [16]. In other cases, viral infection appears to activate checkpoint signaling and harness it to promote the infection. HIV, human papillomaviruses, and polyomaviruses induce and depend on ATM signaling for viral propagation [17], [18], [19], [20], [21], [22]. However, mechanistic understanding of how these viruses activate damage signaling and exploit it for viral propagation is limited.
Simian Virus 40 (SV40), a polyomavirus that propagates in monkey kidney cells, has served as a powerful model to study eukaryotic replication proteins and mechanisms in vivo and in vitro [23], [24], [25], [26], [27]. Checkpoint signaling proteins are dispensable for SV40 DNA replication in vitro, yet in infected cells, ATM or ATR knockdown, over-expression of kinase-dead variant proteins, or chemical inhibition of checkpoint signaling clearly decreases or delays SV40 chromatin replication [26], [28], [29], [30]. To determine how checkpoint signaling facilitates viral replication in SV40-infected primate cells, we have utilized small molecule inhibitors of the PIKK family members ATM, ATR, and DNA-PK to suppress checkpoint signaling in host cells during three specific time windows after SV40 infection. Characterization of the resulting viral DNA replication products reveals that inhibition of ATM or ATR, but not DNA-PK, reduced the yield of unit length viral replication products and caused aberrant viral DNA species to accumulate. ATM inhibition led to unidirectional SV40 DNA replication and concatemeric products, whereas ATR inhibition markedly increased broken SV40 DNA replication forks. Our results strongly suggest that unperturbed viral chromatin replication in infected cells results in double strand breaks, activating checkpoint signaling and fork repair to generate unit length viral replication products.
Replicating SV40 chromatin in infected cells has been visualized by fluorescence microscopy in prominent subnuclear foci that co-localize with Tag and several host proteins essential for viral DNA replication in vitro, suggesting that these foci may represent viral chromatin replication centers [26], [29], [31]. However, SV40 infection activates ATM and ATR signaling, and several DNA damage signaling proteins, e.g. MRN, γH2AX, ATRIP, and 53BP1, co-localize with Tag in these foci [28], [29], [30], [32], implying a link between SV40 replication and damage signaling. On the other hand, interaction of ectopically expressed Tag with the spindle checkpoint protein Bub1 can also induce cellular chromosome breaks [33], indicating that Tag interference with host mitotic checkpoint proteins may suffice to damage genomic DNA in uninfected cells.
As a first step to assess a potential link between SV40 chromatin replication and DNA damage signaling, viral replication centers in SV40-infected BSC40 monkey cells were characterized in detail. Chromatin-bound Tag was visualized in subnuclear foci as expected and colocalized with newly replicated DNA that had incorporated the deoxynucleoside EdU (Figures 1A and S1A). Chromatin-bound PCNA, DNA polymerase δ, and the clamp-loader RFC, host proteins that are essential for viral DNA replication in vitro, colocalized with Tag foci in both BSC40 and human U2OS cells at 48 hours post infection (hpi) (Figures 1A, S1B–D, and S2). In contrast, Cdc45, an essential component of the CMG host replicative helicase that colocalized with replicating chromatin in mock-infected U2OS cells (Figure S2C, D), was virtually excluded from viral replication centers (Figures 1A, S1E, and S2C, D). The results strongly suggest that in infected cells, these chromatin-bound Tag foci represent sites of viral, rather than host, chromatin replication.
We next asked whether SV40 DNA replication itself might induce DNA damage signaling in the absence of viral infection. Toward this end, the plasmids pMini SV40-wt, and its replication-defective variants lacking Tag helicase activity (D474N) [34], or containing a single base pair insertion that inactivates the viral origin (In-1) [35], were transfected into BSC40 monkey cells (Figure 1B). As expected, all three plasmids expressed Tag, but only the SV40-wt plasmid replicated (Figure 1C, D). SV40-wt activated phosphorylation of Chk1 and Chk2 more robustly than either of the replication-defective constructs (Figure 1C, compare lane 1 to lanes 2–3). Moreover, prominent γH2AX foci, a marker of DNA damage signaling in chromatin [36], colocalized with chromatin-bound Tag in viral replication centers in SV40-wt transfected cells (Figure 1E). In contrast, the few γH2AX foci detected in cells transfected with the replication defective plasmids did not colocalize with Tag. Thus, in the context of transfected cells, viral DNA replication, but not SV40-driven Tag expression, is sufficient to induce DNA damage signaling, suggesting that DNA breaks in replicating viral chromatin may activate checkpoint signaling.
To determine the temporal requirements for ATM activity during infection, we exposed infected cells to the specific ATM chemical inhibitor Ku-55933 [12] during the early phase (virus entry, Tag expression, host DNA synthesis), late phase (viral DNA replication, late gene expression, and virion assembly), or throughout a 48-hour infection (Figure 2A). Infected cells exposed to the Ku-55933 solvent, DMSO, served as a positive control. Mock-infected cells not treated with inhibitor served as a negative control. ATM activity was stimulated by infection, as indicated by phosphorylated Nbs1 and Chk2 in western blots (Figure 2B, compare lane 1 to lane 5), reduced by the presence of Ku-55933 in either the early or late phase of infection (Figure 2B, compare lanes 2, 3 to lane 1), and nearly abolished by the presence of Ku-55933 throughout infection (Figure 2B, lane 4).
To assess the impact of ATM inhibition during each phase of infection on viral chromatin replication, we visualized viral replication centers and DNA damage signaling in each infected cell population using immunofluorescence microscopy (Figure 2C). In infected cells exposed to DMSO, the normal, brightly stained viral replication centers with colocalized Tag, EdU, and γH2AX were observed (Figure 2C). When Ku-55933 was present only during the early phase of infection, about half of the cells displayed normal replication centers with colocalized Tag, EdU and γH2AX foci (Figure 2C and D). However, aberrant pan-nuclear staining of Tag, EdU, and γH2AX predominated when Ku-55933 was present during the late phase or throughout infection (Figure 2C and D). Taken together, the results demonstrate that ATM activity was beneficial but not essential during the early phase of infection, whereas it was vital for the assembly and/or stability of viral replication centers during the late phase of infection.
The links between ATM activity and SV40 replication centers led us to hypothesize that inhibition of ATM might affect not only the level, but perhaps also the nature of the viral DNA replication products. To investigate this possibility, we used southern blotting to analyze total intracellular DNA from SV40-infected BSC40 cells that had been treated with DMSO or Ku-55933 throughout infection (Figure 3A). Inhibition of ATM reduced the level of 5.2 kbp viral DNA products migrating as form I (supercoiled), form II (nicked), and form III (linear), relative to that in the DMSO-treated control infections (Figure 3A, compare lanes 1–4 to 5–8). However, ATM inhibition also caused accumulation of high molecular weight SV40 DNA products too large to enter the gel (Figure 3A, compare lanes 3, 4 to lanes 7, 8). These large products failed to migrate into the gel after restriction digestion with enzymes that cut host DNA but not SV40 DNA. In contrast, most of these products collapsed into unit length linear SV40 DNA after digestion with an enzyme that cleaves SV40 DNA once (Figure S3A), indicating that the large DNA products contain head-to-tail repeats of unit length viral DNA.
To quantify the data in Figure 3A, the signal in SV40 monomer bands (forms I, II, and III) in each sample was normalized to that of mitochondrial DNA (Mito) in the same sample. This normalized monomer signal in each sample was then compared to that of the normalized monomer bands in the positive control at 72 hpi. (Figure 3A, lane 4) and graphed in Figure 3B. The graph reveals that ATM inhibition reduced unit length SV40 product by at least 5-fold compared to the DMSO control infections (Figure 3B). Quantification of the concatemeric SV40 DNA in each sample relative to that of the total SV40 signal in the same sample revealed that ATM inhibition increased accumulation of viral DNA concatemers by an order of magnitude compared to that in the DMSO control samples (Figure 3C). Thus, inhibition of ATM throughout infection reduced monomeric and increased concatemeric SV40 DNA products.
To determine what stage of SV40 infection required ATM activity, total intracellular DNA was extracted from infected BSC40 cells exposed to Ku-55933 during three time windows, as diagrammed in Figure 2A. The purified DNA was separated by gel electrophoresis and analyzed in southern blots (Figure 3D). Inhibition of ATM either early or throughout infection reproducibly reduced the level of total viral DNA and monomeric DNA products by 50–80% relative to that generated in the DMSO-treated control infection (Figure 3D, E). Similarly, in the late phase of infection, inhibition reduced viral DNA monomers to a level comparable to that observed when ATM was inhibited during the early phase, yet total viral DNA was only insignificantly decreased compared to DMSO-treated cells (Figure 3D, E). SV40 monomers comprised about 80% of the total viral DNA signal in samples from infected cells exposed to DMSO or Ku-55933 during early phase (Figure 3F). In contrast, monomers comprised only 64% of the total signal in samples treated with Ku-55933 late or throughout infection (Figure 3F). When Ku-55933 was applied either during the late phase or throughout infection, the fraction of total viral DNA in concatemers increased 10- and 11-fold, respectively, relative to the fraction in DMSO-treated infected cells (Figure 3G). The fraction of total SV40 DNA migrating at 20-kbp linear also increased in cells treated with Ku-55933 late or throughout infection, relative to that in DMSO-treated control infections (Figure 3G).
To confirm these findings in a different cell background, the temporal requirements for ATM activity were also determined in SV40-infected human U2OS cells, with similar results (Figure S3B–E). Taking the results together, we infer that SV40-infected cells require ATM signaling, primarily during the late phase of infection, to favor production of unit-length genomes rather than aberrant products.
To better understand how the aberrant viral replication products arise, we compared replication intermediates generated with and without Ku-55933 during the late phase of infection. The total DNA was first digested with a restriction nuclease that cleaves SV40 once in the viral origin (BglI) or once in the region of termination (BamHI). Neutral two-dimensional (2 d) gel electrophoresis was then used to separate viral replication intermediates from the accumulated non-replicating unit-mass SV40 DNA, followed by southern blotting using the whole SV40 genome as the probe [37]. Replicating viral DNA is present in the form of circular, converging forks known as Cairns intermediates (Figure 4B). The digestion of Cairns intermediates with BglI or BamHI results in double Ys or bubbles, respectively (Figure 4A, B). In the BglI-cleaved DNA from DMSO-treated control infections, the bubble arc was absent and the unit-mass viral DNA migrated in the 1 n spot as expected (Figure 4A–C). Also as expected, an intense double Y arc indicative of converging forks and an X structure signal indicative of hemi-catenates or Holliday junctions were observed (Figure 4C). In addition, the simple Y arc signal revealed some unidirectional replicating forks (Figure 4C) that can be most easily explained by rolling circle replication. When BamHI-cleaved DNA from DMSO-treated infected cells was analyzed by 2 d gel electrophoresis, the bubble arc was detected and the double Y arc was absent, as expected (Figure 4D). Similar to BglI digestion, both an X structure and a weaker simple Y arc were present (Figure 4D).
In contrast, the pattern of BglI-digested viral replication intermediates generated in the presence of Ku-55933 displayed a much fainter double Y arc and a more intense simple Y arc (compare Figure 4E with C). Similarly, X structures and D-loops, or other complex branched intermediates (red star), were more prominent when ATM was inhibited (compare Figure 4E with C), consistent with increased Holliday junction formation between replicating rolling circles [38], [39]. Likewise, BamHI-cleaved replication intermediates from Ku-55933-treated infections displayed a robust simple Y arc and a corresponding decrease in the bubble arc (Figure 4F). Moreover, the intense X structure and D-loop arcs were retained (Figure 4F). These patterns suggest that inhibition of ATM sharply increased the frequency of rolling circle replication (Figure 4G). Quantification of the signal present in the simple Y, X structure, D-loop, and double Y arcs from BglI-digested DNA (Figure 4C, E boxes) showed that ATM inhibition increased the abundance of simple Ys, X structures, and D-loop arcs relative to the double Y arc by six, three, and eight-fold, respectively, from three to four independent experiments (Figure 4H). Analogously, quantification of BamHI-digested DNA (Figure 4D, F boxes) revealed ATM inhibition increased the quantities of simple Ys, X structures, and D-loop arcs relative to the bubble arc (Figure 4I). We conclude that the ATM inhibitor Ku-55933 increased both rolling circle replication and strand invasion events at the expense of bidirectional SV40 chromatin replication.
The importance of ATM activity in SV40 chromatin replication suggested the possibility that other checkpoint kinases might also contribute to viral infection. To further explore this question, we treated SV40-infected BSC40 cells with caffeine, a less selective inhibitor of both ATM and ATR in vitro and of the S/G2 checkpoints in vivo [40]. Of note, caffeine is structurally unrelated to the more potent Ku-55933 and ATR inhibitors [12], [13]. As expected, caffeine inhibited phosphorylation of Chk1 and Chk2 when present during the late phase or throughout infection (Figure S4A, B) but also hyper-activated DNA-PK (Figure S4B, compare lane 1 with lanes 2–4) [41]. Caffeine reduced the level of total viral DNA products in SV40-infected BSC40 cells to less than 1% of the control level when caffeine was present throughout infection (Figure S5A, B). Exposure to caffeine late or throughout infection reduced the fraction of total viral DNA signal in monomers (form I, II, III) and increased the fraction in concatemers and other aberrant products (Figure S5A, C, D). Similarly, in SV40-infected U2OS cells, caffeine reduced total viral replication products and increased the fraction of aberrant products (Figure S5E–H). The results further confirm the role of ATM activity in SV40 chromatin replication in infected cells and suggest that ATR and/or DNA-PK activity may stimulate viral replication.
Although SV40 infection did not activate DNA-PK, it was activated in infected cells exposed to Ku-55933 or caffeine, as evidenced by DNA break-dependent auto-phosphorylation of DNA-PK at S2056 [41] (Figures 2B, S4B). To test for a potential role of DNA-PK activity in viral chromatin replication, SV40-infected BSC40 cells were exposed to small molecule inhibitors of DNA-PK during the early or late phase, or throughout infection and total intracellular DNA was analyzed by southern blotting (Figure S6A–C). When DNA-PK was inhibited with either Nu7441 or Nu7026, the levels of viral monomer and aberrant viral DNA products closely resembled those in SV40-infected BSC40 cells (Figure S6D). Moreover, inhibition of DNA-PK had little or no effect on viral replication centers (data not shown). Thus, it is unlikely that DNA-PK has a major role in viral chromatin replication in unperturbed infected cells.
The role of ATR kinase activity in infection was directly examined by treating SV40-infected BSC40 cells with a specific small molecule inhibitor of ATR, VE-821 (ATRi) [13], during three different time windows of infection (Figure S7A). As expected, ATRi caused a third of the cells to lose viability over 48 h, but SV40-infected and mock-infected cells were equally sensitive (Figure S7B). SV40 infection activated Chk1, as indicated by phosphorylation of Ser317 (Figure S7C, compare lane 1 with lane 5), and ATRi effectively suppressed ATR activation during each time window (Figure S7C, lanes 2–4).
Viral DNA replication products from the four cell populations and mock-infected cells were analyzed by southern blotting and quantified relative to mitochondrial DNA in the same samples. In the presence of ATRi, the level of total viral DNA replication products declined markedly relative to that in DMSO-treated control infections, amounting to only 10% of the control when ATRi was present for the full 48 h (Figure 5B, C). In cells exposed to ATRi during the late phase or throughout infection, the fraction of viral DNA products in monomers (forms I, II and III) dropped, whereas that in concatemers and other aberrant products rose (Figure 5B–E and Figure S8A). Analysis of viral replication products from SV40-infected U2OS cells exposed to ATRi demonstrated a similar requirement for ATR activity (Figure S8B–D). Taken together, these results indicate that infected cells require ATR activity before, as well as during viral chromatin replication, for normal accumulation of viral genomes.
The structures of viral replication intermediates generated in the presence and absence of ATR kinase activity were characterized by using neutral 2 d gel electrophoresis and southern blotting. As expected, BglI-digested SV40 replication intermediates from control infections displayed a strong double Y arc indicative of converging forks, X structures, and a weaker simple Y arc with both legs of similar intensity (Figure 6B). In contrast, BglI-digested replication intermediates from ATRi-treated cells yielded a novel pattern (Figure 6C). Although the double Y and X structure arcs closely resembled those in the DMSO control, the simple Y arc displayed much greater intensity in the leg closer to the 1 n linear DNA (Figure 6B and C, zoomed box) than in the other leg closer to 2 n linear DNA. This pattern is not consistent with rolling circle replication, which generates a uniformly intense simple Y arc (Figure 4) or with two stalled replication forks, of which one breaks, creating an asymmetric simple Y [42]. The observed pattern is also inconsistent with one normal replication fork and one slower moving fork, which would converge asymmetrically to generate a cone-shaped signal between the X structure arc and the Y arc [43]. However, the novel pattern observed could arise if one fork stalls prematurely (Figure 6F, I, II), while the other fork progresses until it encounters the stalled fork and then breaks, generating a broken late Cairns intermediate (Figure 6F, III, IV) [37]. Close inspection of the intense leg of the Y arc reveals that its intensity is uneven, suggesting that it may arise from a series of closely spaced break sites along the Y arc (Figure 6C). If the break sites reside 2.5 kb or less from the BglI cleavage site, the intensity of signals would be greater in the right leg of the simple Y arc, as observed (Figure 6C, box). This interpretation predicts that if replication products from the ATRi-treated infection were digested with BamHI, which cleaves 2.5 kb from the BglI site, the sites of breakage, and hence greater signal intensity, should shift to the left leg of the simple Y arc, closer to the 2 n linear DNA (Figure 6A, E). Indeed, this shift was observed (compare Figure 6D with E), confirming that when the moving replication fork encountered a fork that had stalled in the presence of ATRi, the moving fork broke (Figures 6F and S9).
This study presents several lines of evidence that SV40 harnesses host DNA damage signaling for quality control of viral chromatin replication. We show that viral DNA replication in vivo is sufficient to induce DNA damage signaling at viral replication centers (Figures 1, S1, S2), suggesting that DNA lesions may arise in unperturbed replicating viral DNA. Importantly, damage signaling is vital to maintain viral replication centers (Figures 1, 2). Furthermore, suppression of ATM and/or ATR signaling increases the level of aberrant viral replication products at the expense of unit length viral DNA (Figures 3–5, S3, S5, S8), implying that viral replication-associated damage in infected cells requires ATM and ATR signaling to promote repair of viral replication forks. Lastly, our results indicate that the defective replication intermediates resulting from inhibition of ATM (Figure 4) and ATR (Figures 6, S9) are distinctive. Taken together, our results support a model in which ATM and ATR serve different but complementary roles in orchestrating repair at viral replication forks (Figure 7).
SV40 chromatin replication centers resemble over-sized host DNA damage response foci (for a comparison, see Figure 1 in ref [29]), where diverse damage signaling and DNA repair proteins assemble on chromatin at a DNA lesion and dissociate when repair is completed [1], [44]. Many of the same signaling and repair proteins are found at both viral replication centers and host damage response foci [18], [21], [22], [28], [29], [30], [32], [33] (Sowd, unpublished). However, unlike the prominent viral replication centers, the punctate host damage response foci encompass megabase regions of chromatin, raising the question of how SV40 mini-chromosomes give rise to the large subnuclear foci observed in the microscope. The size of SV40 replication centers increases with the number of incoming viral genomes and with time post-infection in permissive primate cells [29], suggesting that our ability to detect viral replication centers depends on the ability of each infected cell to generate 10–100 thousand daughter genomes [45]. Moreover, unperturbed viral replication centers display nascent ssDNA (Sowd, unpublished) and DNA breaks that are likely responsible for activating checkpoint signaling, analogous to lesions that nucleate host damage response foci.
A major difference between SV40 replication centers and host damage response foci is that checkpoint signaling does not inhibit the viral replication machinery, whereas Chk2 phosphorylation of the purified host replicative helicase Cdc45/Mcm2-7/GINS inhibits its helicase activity in vitro [46] and Chk1 inhibits Cdc45 recruitment to chromatin to initiate replication in vivo [47]. Based on these considerations, we suggest that SV40 replication centers serve as hubs where host replication and repair factors efficiently service many client viral genomes in close proximity. These hubs are nucleated and maintained by the assembly of the ATM and ATR signaling complexes at sites of viral replication stress, followed by recruitment of downstream repair factors [1]. Of note, all of the host proteins needed for SV40 DNA replication in vitro [23], [24], [25] also function in host DNA repair [23], [25], [48], [49]. Thus SV40, though it encodes only a single essential replication protein, has evolved a rather remarkable strategy to generate viral replication compartments.
Recent studies in several laboratories, including ours, established that knockdown or inhibition of ATM in polyomavirus-infected cells reduced production of unit length viral genomes [21], [22], [28], [29]. Since these studies evaluated only unit length viral DNA, the aberrant viral replication products generated by unidirectional replication forks were overlooked (Figures 3, 4, S3). Interestingly, total intracellular DNA from unperturbed infected CV1P cells has also been reported to contain head-to-tail SV40 DNA repeats of 50 to 100 kbp at very late times after infection [45]. These observations indicate that concatemers may be a normal product of viral replication, and suggests that inhibition of ATM activity might simply increase the frequency of unidirectional replication, advance its timing, or both.
Although replication-associated breaks may be a rare event during unperturbed viral DNA replication, the large number of replicating viral genomes would facilitate their detection, particularly when ATM activity is suppressed. Yet surprisingly, when undigested total intracellular DNA from an ATM-inhibited infection was analyzed by 2 d gel electrophoresis, bidirectional replication was still observed (data not shown) and unit length viral DNA remained the predominant product when ATM was inhibited (Figures 3 and S3). These observations can be most simply explained by a model in which theta-form SV40 replication intermediates (Figure 7, I–III) break, giving rise to unidirectional forks that amplify the break by generating concatemers and branched concatemers [38], [39] (Figure 7, V, VI). Our data suggest that ATM kinase activity is crucial for the repair of one-ended replication-associated DSBs to reassemble bidirectional replication intermediates (Figure 7, VII) [49], [50], [51].
It is interesting to consider a possible role for unidirectional viral replication and its large concatemeric products in the tumorigenic activity of SV40, and more broadly of polyoma- and papillomaviruses. Concatemeric genomes of Merkel cell carcinoma virus and HPV are often integrated into human chromosomal DNA in tumors associated with these viruses [52], [53], [54]. The integration events and the consequences of long-term viral oncogene expression are primary risk factors for such cancers. It seems likely that in an infected cell under conditions of insufficient ATM activity, the level of viral concatemers would rise. With inadequate ATM activity, breaks in host chromosomal DNA would also be less frequently repaired through accurate, homology-dependent repair. Thus one can speculate that viral DNA concatemers generated under conditions of insufficient DNA damage signaling might be inaccurately joined with broken host chromatin, contributing to viral tumorigenesis [55].
SV40 chromatin replication was highly sensitive to inhibition of ATR throughout a 48 h infection (Figures 5, S8). One consequence of ATR inhibition was that infected cells continued to cycle throughout infection, rather than arresting in late S phase where viral DNA replication would be favored [30]. However, the most prominent SV40 replication defect induced by ATRi was the tendency of converging replication forks to stall and break (Figures 6, 7, S9). Our data imply that after initiating replication at the viral origin, one replisome encounters an unknown replication block at variable positions in the viral genome (Figure 6F, S9, I and II, red triangle). Since the two sister Tag helicases need not remain coupled after initiation, they can proceed asynchronously as they replicate the viral genome bidirectionally [26], [56], [57], [58], [59]. Thus, the functional, unstalled replisome continues replication until it approaches the stalled fork (Figure 6F, III). We suggest that without ATR activity, the unstalled fork cannot converge with the stalled fork and breaks, yielding the pattern observed on the simple Y arc (Figure 6C, E, F, IV–VI). Consistent with this interpretation, fork convergence is well known to represent a slow step during unperturbed SV40 DNA replication in infected cells and to occur in a ∼1 kbp region around the BamHI site [60], [61], [62], suggesting that specialized host proteins and ATR-dependent modifications may be needed to complete replication.
Our observation that ATRi renders SV40 fork convergence prone to DNA breakage is reminiscent of common fragile sites in the human genome, which suffer gaps and breaks in Seckel Syndrome cells that express defective ATR alleles [63]. Thus SV40 and other small DNA tumor virus genomes may harbor a potential fragile site in the region where the two viral replication forks converge. Consistent with this speculation, C-terminal truncation of the polyomaviral T antigen encoded in the “fragile site” could render an integrated viral genome replication-defective and perhaps more tumorigenic [52], [64], [65], [66]. Similarly, the viral “fragile site” where replication forks converge would correspond to common viral genome breakpoints in integrated high risk papillomaviral genomes in cervical cancer [67], [68], [69].
For details not described below, please refer to the online Supporting Methods (Protocol S1).
Ku-55933, kindly provided by Astra-Zeneca, was used as described [12], [29]. Importantly, Ku-55933 did not inhibit sixty off-target kinases. It specifically inhibits purified ATM with an IC50 of 12.9 nM, whereas it inhibits the related kinases mTOR and DNA-PK with IC50 values of 2500 nM and 9300 nM, respectively, in vitro [12]. Caffeine (Sigma) was dissolved to 24 mM in DMEM and used at a final concentration of 8 mM to inhibit ATM and ATR [40]. ATRi and Nu7441 were generous gifts from Dr. David Cortez. ATRi dissolved in DMSO at 5 mM was used at a final concentration of 5 µM [13]. ATRi selectively inhibits ATR with a Ki of 13 nM, whereas at least a 100-fold higher concentration is required in vitro to inhibit the related kinases ATM (Ki = 16000 nM), DNA-PK (Ki = 2200 nM), mTOR (Ki = 1000 nM), and PI3Kgamma (Ki = 3900 nM) [13]. Nu7441 was dissolved in DMSO to 2 mM and applied to cells at 1 µM [70], [71]. Nu7026 (EMD) was dissolved to 5 mM in DMSO and used at a final concentration of 10 µM [72].
DMEM containing inhibitor or solvent was added to cells 30 min prior to infection. At time zero, DMEM with inhibitor or solvent was removed, and fresh warm DMEM containing inhibitor or solvent and SV40 was added to cells. Cells were gently rocked every 15 min during the first 2 hpi. At 2 hpi, complete DMEM containing inhibitor or solvent was added to each dish of cells. At 20 hpi, medium was aspirated and cells were washed once with PBS to remove residual inhibitor or solvent. Fresh medium containing inhibitor or solvent was then added to cells and infections were allowed to proceed until the chosen endpoint. Solvent control treatments utilized the solvent concentration present in the inhibitor-treated medium.
Total intracellular DNA was prepared from infected and mock-infected cells. For each experiment, all samples were prepared from an equal number of cells. Cell pellets were resuspended in 0.4 ml of TE (10 mM Tris pH 8.0, 1 mM EDTA). SDS, RNase A, proteinase K, and Tris pH 7.5 were added to a final concentration of 0.4%, 0.2 mg/ml, 50 ug/ml and 100 mM, respectively, in a total volume of 0.5 ml. Following overnight digestion at 37°C, each sample was extracted twice with Tris-saturated phenol (pH 7.9) and once with 24∶1 chloroform: isoamyl alcohol. DNA was precipitated with sodium acetate and ethanol. DNA was allowed to dissolve in T0.1E (10 mM Tris pH 8.0, 0.1 mM EDTA) for 2 days, and then digested overnight at 37°C with 40 U of SacI-HF and XbaI (both from New England Biolabs). Digested DNA was re-precipitated and then dissolved in 50 µL of T0.1E per 2.5×105 cells. Equal volumes of DNA were loaded on gels for southern blots unless otherwise indicated.
One-dimensional 0.7% agarose gels in 1× TAE were electrophoresed at 10 V/cm for 1.5 h. Neutral 2 d gel electrophoresis was performed as previously described [37] with the following modifications. The first dimension of the gel was electrophoresed at 1 V/cm through a 0.4% 1× TAE for 22 h. 1× TAE was found to enhance separation of D-loop arc (data not shown). The second dimension was electrophoresed at 5.5 V/cm through a 1.1% 1× TBE gel containing 0.5 ng/ml ethidium bromide for 5.5 h with circulation.
Southern blotting was performed using radiolabeled probes for SV40 and BSC40 mitochondrial DNA as described [34]. A probe for human mitochondrial DNA was generated by PCR amplification (primers: U2OS Mito-F ACG CGA TAG CAT TGC GAG AC; U2OS Mito-R CTT TGG GGT TTG GTT GGT TCG), followed by random priming. Hybridized blots were visualized using a Typhoon Trio laser scanning imager (GE Healthcare) and quantified using ImageQuant 5.2 (GE Healthcare).
Bands or arcs corresponding to each DNA structure of interest were quantified and the value from a region of the blot without signal, e.g. Mock for SV40 probe, was subtracted as background. To compare the level of a DNA structure after a given treatment (e.g. DNA structure (% of Total DNA)), the total signals for the DNA were summed, and the signal of a discrete DNA structure (e.g. form I monomer) were divided by the total signal in the lane (e.g. [form I monomer signal]/[total signal in the lane]). To quantify variations in replication between treatments, all SV40 DNA signals were normalized using the respective mitochondrial DNA signal. Normalized signals were then divided by the normalized signal present in the infected solvent control to yield the DNA signal (% of DMSO).
The southern blot signals from an equal area of each arc in neutral 2 d gels were quantified (boxed areas in Figure 4C, D, E, F). Background signal in an area of equal size was subtracted, and the values for each arc were normalized to the value for the double Y (Figure 4H) or bubble arc (Figure 4I).
Statistics were performed in Microsoft Excel using the data analysis package. Prior to t-test, single factor ANOVA analysis was performed. If ANOVA resulted in p<0.5, a two sample t-test assuming unequal variances was performed. One-tailed p values from student's t test are denoted by the number of asterisk(s): * p<0.05 ** p<0.01 *** p<0.001 **** p<0.0001. All one tailed p values were generated by comparing data from SV40 infection in the presence of inhibitor to that from SV40 infection in the presence of DMSO. Bar graphs present the average of 3 to 4 independent experiments and error bars represent standard deviation.
|
10.1371/journal.pcbi.0030026 | Functional Impact of Missense Variants in BRCA1 Predicted by Supervised Learning | Many individuals tested for inherited cancer susceptibility at the BRCA1 gene locus are discovered to have variants of unknown clinical significance (UCVs). Most UCVs cause a single amino acid residue (missense) change in the BRCA1 protein. They can be biochemically assayed, but such evaluations are time-consuming and labor-intensive. Computational methods that classify and suggest explanations for UCV impact on protein function can complement functional tests. Here we describe a supervised learning approach to classification of BRCA1 UCVs. Using a novel combination of 16 predictive features, the algorithms were applied to retrospectively classify the impact of 36 BRCA1 C-terminal (BRCT) domain UCVs biochemically assayed to measure transactivation function and to blindly classify 54 documented UCVs. Majority vote of three supervised learning algorithms is in agreement with the assay for more than 94% of the UCVs. Two UCVs found deleterious by both the assay and the classifiers reveal a previously uncharacterized putative binding site. Clinicians may soon be able to use computational classifiers such as those described here to better inform patients. These classifiers can be adapted to other cancer susceptibility genes and systematically applied to prioritize the growing number of potential causative loci and variants found by large-scale disease association studies.
| A significant number of breast and ovarian cancers are due to inherited mutations in the BRCA1 and BRCA2 genes. Many women who receive genetic testing for these mutations are found to have variants of the genes that result in changed amino acids in the BRCA1 or BRCA2 proteins. The effect of these variants on cancer risk is not well-understood, posing a problem for patients and their health providers. We describe computational biology methods that predict and analyze the impact of 36 BRCA1 variants on protein function. The predictions are validated by biochemical assays of BRCA1 in yeast and mammalian cell cultures. The speed and accuracy of the computational methods is well-suited to rapid evaluation of large numbers of variants in genes that predispose to inherited diseases.
| The BRCA1 gene encodes a large multifunction protein involved in cell-cycle and centrosome control, transcriptional regulation, and in the DNA damage response [1–3]. Inherited mutations in this gene have been associated with an increased lifetime risk of breast and ovarian cancer (6–8 times that of the general population) [4]. There are several thousand known deleterious BRCA1 mutations that result in frameshifts and/or premature stop codons, producing a truncated protein product [5]. In contrast, the functional impact of most missense variants that result in a single amino acid residue change in BRCA1 protein is not known. The Breast Cancer Information Core database (http://research.nhgri.nih.gov/bic/), a central repository of BRCA1 and BRCA2 mutations identified in genetic tests, currently contains 487 unique missense BRCA1 variants (April 2006), of which only 17 have sufficient genetic/epidemiological evidence to be classified as deleterious (Clinically Important) and 33 as neutral or of little clinical importance (Not Clinically Important). As genetic testing for inherited disease predispositions becomes more commonplace, predicting the clinical significance of missense variants and other UCVs will be increasingly important for risk assessment.
Because most UCVs in BRCA1 and BRCA2 occur at very low population frequencies (<0.0001) [6], direct epidemiological measures, such as familial cosegregation with disease, are often not sufficiently powerful to identify the variants associated with cancer predisposition. A promising approach is to supplement epidemiological and clinical analysis of UCVs with indirect approaches such as biochemical studies of protein function and bioinformatics analysis [6–8]. In the future, physicians and genetic counselors may be able to rely on all these sources of information about UCVs when counseling their patients.
Previous bioinformatics analysis of BRCA1 UCVs has depended primarily on measures of evolutionary conservation in multiple sequence alignments of human BRCA1 and related proteins from other organisms [9–11]. Two groups have attempted to include information about BRCA1 protein structure. Williams et al. predicted the impact of 25 missense variants in BRCA1′s C-terminal BRCT domains by considering both conservation and location of variant amino acid residues in an X-ray crystal structure [12]. Variants were predicted deleterious if their properties were similar to properties of biochemically characterized deleterious variants in Escherichia coli Lac Repressor and bacteriophage T4 lysozyme. Mirkovic et al. developed a set of hierarchical rules (Rule-based decision tree) based on the conservation, variant structural location, and amino acid residue physiochemical properties of 30 deleterious and seven neutral biochemically characterized BRCA1 missense variants [7].
We have developed a novel combination of 16 predictive features that describe conservation, impact of mutation on protein structure, and amino acid residue properties, and used them as input to computational supervised learning algorithms. These algorithms are trained to learn a generic classification of amino acid residue substitutions and positional contexts. The training set is composed of 618 missense variants in the transcription factor TP53 biochemically characterized as functional or nonfunctional in a transactivation assay [13]. TP53 is a tumor suppressor gene that is inactivated in the majority of human cancers.
Our validation set is composed of 36 missense variants in BRCA1′s BRCT domains that were biochemically characterized with a transactivation assay [14]. These 36 variants were selected because they occur in individuals from families with breast or ovarian cancer in which no other deleterious mutation in BRCA1 or BRCA2 was found and were functionally tested under the same protocols and conditions, yielding standardized measurements of each variant's transactivation activity with respect to wild-type. We use the validation set to assess the supervised learners and compare them with algorithms based on evolutionarily allowed amino acid residues or empirically derived rules. The algorithms with greatest correlation between assay and computational predictions are the supervised learners Naïve Bayes [15], Support Vector Machine [16], and Random Forest [17].
Given a protein X-ray crystal structure, the supervised learning approach can quickly and accurately predict the outcome of our BRCA1 transactivation assay with greater than 94% accuracy on tested missense variants in the BRCT domains. We have applied the best performing supervised learners to blind prediction of the functional impact of 54 UCVs found in BIC and occurring in the BRCA1 BRCT domains. For each of these UCVs, we produce a consensus prediction and, where possible, a molecular explanation for the impact of the variant.
Next, we describe the protocol used to train and validate the supervised learning algorithms, the selection of 16 features used to represent each missense variant to the algorithms, implementation details of each algorithm, and performance assessment criteria (Methods). We then show how a combination of sequence- and structure-based features in a supervised learning setting obviates some of the problems with evolutionary analysis and empirically derived rules, providing specific examples of the strengths and weaknesses of each approach (Results, Discussion). We show that two of the variants found to be deleterious by both the assay and the classifiers may be at a previously uncharacterized protein binding site and that electrostatic changes at the site may weaken the interactions of BRCA1 and protein partners that are important for its functions (Discussion). Finally, we discuss the generalizability of our methods to other cancer susceptibility genes and to large-scale disease association studies (Discussion).
We trained four supervised learning algorithms to discriminate between a set of 398 deleterious/nonfunctional and 220 neutral/functional TP53 missense variants, biochemically characterized in a transactivation assay [13]. The variants were downloaded from the IARC TP53 website (http://www-p53.iarc.fr). We only used variants capable or incapable of activating transcription for all eight of the TP53 promoters tested in the transactivation assay and located in the core DNA binding domain of TP53.
The 36 BRCA1 BRCT missense variants described in our companion paper [14] were used as an independent validation set for the supervised learners. These variants were also classified by sequence-analysis methods based on evolutionarily allowed amino acid residues: Align Grantham Variation Grantham Deviation (Align-GVGD) [18], Sorting Intolerant from Tolerant (SIFT) [19], Ancestral Sequence [9,11], and empirically derived rules encoded in a decision tree (Rule-based decision tree) [7]. Each method was evaluated by its agreement with the BRCA1 transactivation assay on the validation set, according to accuracy (fraction of all variants correctly classified), sensitivity or true positive rate (fraction of all nonfunctional variants correctly classified), specificity or true negative rate (fraction of all functional variants correctly classified), Matthews correlation coefficient [20], and coverage (fraction of variants for which a prediction was made) (Table 1). Matthews correlation coefficient is defined as
and ranges from −1.0 (worst) to 1.0 (best). A coefficient of 0 is equivalent to a random prediction, and less than 0 indicates a worse than random prediction. TP is the number of correctly classified nonfunctional variants, TN the number of correctly classified functional variants, FP the number of incorrectly classified nonfunctional variants, and FN the number of incorrectly classified functional variants.
For the Naïve Bayes, Support Vector Machine, Random Forest, Decision Tree, Align-GVGD, and SIFT classifiers, we computed a receiver operating characteristic (ROC) curve that quantifies the tradeoff between coverage of detected nonfunctional variants (true positive rate) and misclassified functional variants (false positive rate = 1 − specificity). ROC analysis was not possible for the Rule-based decision tree and Ancestral Sequence algorithms, which predict the class of a missense variant but do not provide an associated score.
The supervised learning algorithms (Naïve Bayes, Support Vector Machine, Random Forest, Decision Tree) were trained by associating each amino acid residue substitution in the TP53 training set with 16 carefully selected predictive features (Table 2). A vector of features for a single substitution is denoted as
. The features describe properties of variant and wild-type residues: local structural environment; physiochemical attributes; and evolutionary conservation. To compute the features, we used DSSP (a program that calculates a variety of geometrical properties for each amino acid residue in a protein structure) [21], MODELLER for comparative protein structure modeling [22], SAM-T2K for protein sequence alignments and hidden Markov models [23], and in-house PERL code.
We began with a core set of 13 features selected by a correlation analysis between features and classes (functional or nonfunctional) of the TP53 variants, as described previously [24] (Table S3). An additional 18 candidate features were evaluated by adding them to the core set and doing 10-fold cross-validation tests of Support Vector Machine performance. Three features were found to improve performance and were added to the optimal feature set; the others were rejected (Tables S4 and S5). Next we evaluated Support Vector Machine performance with each of the best 16 features held out. In each case, the 10-fold cross-validation test yielded decreased performance.
We used X-ray crystal structures from the Protein Data Bank [25] for the BRCA1 BRCT domains (1t29 chain A in complex with BACH1 peptide) [26] and the DNA binding domain of TP53 (1kzy) [27]. We performed in silico mutations on the structures with the MUTATE_MODEL routine of MODELLER (available as a Python script at http://salilab.org/modeller/wiki/Mutate_model). MUTATE_MODEL substitutes the wild-type amino acid residue at a position of interest with a variant amino acid residue, and optimizes the coordinates of the variant's backbone and sidechain atoms with an initial conjugate gradient minimization, molecular dynamics optimization with simulated annealing, and a final conjugate gradient minimization (E. Feyfant, 2004, private communication).
The amino acid residue sequences of human TP53 (P04637) and BRCA1 (P38398) were downloaded from UNIPROT [28], and each was used as a seed sequence for the SAM-T2K iterative alignment-building algorithm [23]. For BRCA1, only amino acid residues in the BRCT domains (1649–1859) were aligned. We used the SAM w0.5 program to apply sequence weighting and regularization with Dirichlet mixtures [29] to each resulting alignment and to produce a profile hidden Markov model [30]. The TP53 and BRCA1 alignments and hidden Markov models are available upon request.
We trained a soft margin Support Vector Machine classifier with a radial basis kernel using the e1071 package in R [31]. The Support Vector Machine algorithm optimizes a vector of weights
(one weight for each training example) and a bias parameter b. The parameters g (radial basis kernel width) and C (penalty for violating the soft margin) were optimized on the training set with grid search using default parameters. Each of the 28 missense variants
was then scored with the discriminant function
where l is the number of examples in the training set, yi is the class label of each example in the training set (for deleterious/nonfunctional variants yi = −1 and for neutral/functional variants yi = 1), and
is the value of the radial basis kernel function given
and training example
. Variants are classified as deleterious/nonfunctional if
and neutral/functional if
.
The Naïve Bayes algorithm estimates the probability that each variant belongs to deleterious or neutral classes
by applying the Bayes rule:
where the prior class probability P(C) is the fraction of deleterious (or neutral) missense variants in the training set and each feature Xi is assumed to be conditionally independent of the k − 1 other features, given its class membership, so that
where P(Xi | C) is estimated from the training set. We used the Naïve Bayes method in R's e1071 package. Each feature was approximated to be normally distributed and no smoothing was applied to the feature distributions.
We used the rpart package in R [32] to train a Decision Tree with the following parameters: minsplit = 20 (minimum number of observations required at a tree node before a split is attempted) and cp = 0 (no pruning of tree regardless of whether a split will improve model fit). To reduce overfitting, we pruned the resulting tree using the standard heuristic “1 Standard Error rule” [33] and 10-fold cross-validation. According to the 1 Standard Error rule, the pruned tree with best generalization properties has a cross-validation error on the training set 1 Standard Error worse than the tree with the lowest cross-validation error. The pruning process yielded a reduced set of features: Φ and Ψ mainchain dihedral angles, normalized solvent accessibility of wild-type, Grantham difference, volume change, relative entropy, and positional hidden Markov model conservation score.
We used the randomForest package in R [34] to train a Random Forest, an algorithm based on a majority vote of a large number of decision trees, in which the candidate features at each tree node are randomly sampled [17]. The user-defined input parameters to randomForest are total number of trees in the forest and mtry (number of randomly sampled features considered as candidates for a split at each tree node). Both were selected with grid-search optimization as described for the Support Vector Machine [31].
Predictions of Naïve Bayes, Decision Tree, and Random Forest are in the form of class conditional probabilities, where the two classes are D (deleterious/nonfunctional) and N (neutral/functional). For each example, the classifiers report P(D |
) (probability that the variant is deleterious, given feature vector
) and P(N |
) (probability that the variant is neutral, given feature vector
). To evaluate accuracy, true positive rate, true negative rate, and Matthews correlation coefficient, we classified variants as deleterious if P(D |
) > 0.5 and neutral otherwise. To compute ROC curves, we used the log likelihood ratio
as the output score of Naïve Bayes, Decision Tree, and Random Forest.
The SIFT algorithm [19] predicts the probability that a missense mutation occurs at a given alignment position. Variants that occur at conserved alignment positions are expected to be tolerated less than those that occur at diverse positions. The algorithm uses a modified version of PSI-BLAST [35] and Dirichlet mixture regularization [29] to construct a multiple sequence alignment of proteins that can be globally aligned to the query sequence and belong to the same clade. We used the SIFT server (http://blocks.fhcrc.org/sift/SIFT.html), with PSI-BLAST search set to the Swissprot-TrEMBL protein sequence database [36]. Both the full-length human BRCA1 sequence (amino acid residues 1–1863) and the BRCT C-terminal domain sequence only (amino acid residues 1649–1859) were submitted to the server. For the full-length sequence, SIFT reported low confidence predictions for 34 out of 36 missense variants. Consequently, we based our SIFT predictions on the C-terminal domain sequence. To compute accuracy, true positive rate, true negative rate, and Matthews correlation coefficient, we used the binary class predictions of the SIFT server (deleterious or neutral), based on the default SIFT threshold (tolerated mutation probability > 0.05). For ROC analysis, we used the raw SIFT probabilities, which range from 0 to 1.
The Ancestral Sequence classifications were computed as described [9,11]. Each position in an alignment of eight mammalian BRCA1 orthologs identified as giving best results by Pavlicek et al. was categorized as fixed (completely conserved), conserved (substitution of similar amino acid residues), or nonconserved (dissimilar amino acid residues or gaps). Any substitution at a fixed position and any nonconservative substitution at a conserved position is classified as deleterious. Amino acid residue similarity is based on the Gonnet PAM250 score (i.e., the likelihood that amino acid residue A has mutated into amino acid residue B in a pair of sequences that have diverged by 250 mutations per 100 amino acid residues of sequence) [37].
The Align-GVGD method calculates two scores for each amino acid residue substitution, Grantham Deviation (GD) and Grantham Variation (GV), based on a modified Grantham distance measure [18,38]. The scores define four categories of missense variants: “Enriched deleterious 1” variants occur at invariant alignment positions for which the substitution is outside the range of variation observed at the position (GV = 0, GD > 0); “Enriched deleterious 2” occur at variable alignment positions containing physiochemically similar amino acid residues where the substitution is outside the range of observed variation (0 < GV < 61.3, GD = 0); “Enriched neutral 1” occur at variable positions containing physiochemically similar amino acid residues where the substitution is inside the range of variation (GV > 0, GD = 0); and “Enriched neutral 2” occur at variable positions containing dissimilar amino acid residues where the substitution is slightly outside the range of variation (GV > 61.3, 0 < GD < 61.3). We classified variants using first an alignment of placental mammals, a marsupial (gray short-tailed opposum), chicken, frog, and the pufferfish Tetraodon (“Align-GVGD Tnig”), and second an alignment that also includes the sea urchin Strongylocentrotus purpuratus (“Align-GVGD Spur”). Accuracy, true positive rate, true negative rate, and Matthews correlation coefficient were evaluated by reducing the four categories to deleterious/nonfunctional or neutral/functional. Variants may have GV and GD values that do not match any of the four categories (e.g., a variant with GV = 80 and GD = 80), which lowers coverage, Matthews correlation coefficient, true positive, and false positive rates. For ROC analysis, rather than fixing thresholds on GV and GD at 61.3 for each substitution, we considered the number of true positives and false positives over a range of thresholds, from the smallest to largest values of GV and GD in our dataset (0 to 215).
A Rule-based decision tree is a classification tree with human-designed rules that uses both structure- and sequence-based information, implemented in PERL [7]. Rule-based Decision Tree classifies a missense variant as either deleterious/nonfunctional or neutral/functional, but does not compute numerical scores.
The structural models of all BRCA1 BRCT missense variants were visually compared with the wild-type structure (1t29) using the molecular graphics program Chimera [39]. We explored changes in hydrogen bonding patterns and geometric properties of the molecular surface with Chimera's FindHBond and MSMS routines. To visualize the distribution of amino acid residue conservation on the protein surface, the RenderByAttribute routine was used, with coloration defined by percent conserved in a hand-edited SAM-T2K alignment of BRCA1 orthologs. Species used in this alignment were Homo sapiens (AAA 73985), Pan troglodytes (AAG43492), Gorilla gorilla (AAT44835), Pongo pygmaeus (AAT44834), Macaca mullata (AAT44833), Canis familiaris (AAC48663), Bos taurus (AAL76094), Monodelphis domestica (AAX92675), Mus musculus (AAD00168), Rattus norvegicus (AAC36493), Gallus gallus (AAK83825), Xenopus laevis (AAL13037), and Tetraodon nigroviridis (AAR89523).
A highly conserved surface patch was identified as a possible binding site and subjected to further analysis. We used DELPHI [40] to compute the electrostatic surface potential at the putative binding site for the wild-type structure and for models of two solvent-exposed variants characterized as deleterious in our functional assays (T1685A and R1753T) [14]. The solvent relative dielectric constant was set to 4.0, the protein relative dielectric constant to 20.0, and ionic strength to the physiological value of 0.2 mM. Charges were estimated with the united atom AMBER model [41]. The proteins were prepared for DELPHI by adding heavy atoms missing from the 1t29 crystal structure with MODELLER's COMPLETE_PDB routine and adding hydrogens with REDUCE [42], then visualized in Chimera with a GRASP surface representation [43].
We compared the transactivation activity of the wild-type LexA DBD-BRCA1 or GAL4DBD-BRCA1 fusion construct in both yeast and mammalian cells with the activity of constructs containing 36 single missense variants in the BRCT domains [14]. Variant constructs presenting 50% or more of wild-type activity are characterized as neutral and those with 45% or less are characterized as deleterious, thresholds that are in agreement with available genetic evidence. These functional characterizations were used as a standard to evaluate the reliability of nine computational classifiers (Figure 1). We also provide a post-prediction analysis of these variants (Table S1). Three classifiers with the highest correlation to the functional assay were applied to predict the impact of 54 UCVs in the BRCT domains currently listed in the Breast Information Core database.
Based on ROC analysis, the supervised learners Random Forest, Support Vector Machine, and Naïve Bayes yield the most reliable computational classifications of the 36 variants (Figure 2). The area under the ROC curve (AUC) quantifies the probability that a classifier will give a randomly drawn deleterious example a lower score than a randomly drawn neutral example. AUC is 0.992 for Random Forest, 0.947 for Support Vector Machine and Naïve Bayes, 0.86 for Align-GVGD Tnig, 0.852 for Align-GVGD Spur, 0.783 for SIFT, and 0.738 for Decision Tree (Figure 2). The Decision Tree algorithm appears to overfit the training set and generalizes less well than the other supervised learners.
Three of the supervised learning algorithms (Naïve Bayes, Support Vector Machine, and Random Forest) produce the best classifications of the 36 variants, as measured by accuracy, true positive rate, true negative rate, Matthews correlation coefficient, and coverage, using default thresholds (Table 1). According to these statistical measures, the best sequence analysis methods are Ancestral Sequence and Align-GVGD. Random Forest, Naïve Bayes, and Support Vector Machine are the most accurate scoring predictors, according to the AUC. The methods rankings are slightly different when evaluated by threshold-dependent statistics that reduce predictive scores to deleterious/neutral classes or by the score-based and threshold-independent ROC statistic of AUC.
We applied the top performing algorithms (Naïve Bayes, Support Vector Machine, and Random Forest) to predict the impact of 54 BRCA1 UCVs listed in BIC that (a) are located in the BRCT domains, and (b) have not been functionally characterized by our transactivation assays. Based on a majority vote of the computational predictors, we computed a “consensus prediction” for each UCV (Figure 3). We provide structural explanations for the impact of as many variants as possible, and indicate where the predictions are supported by biochemical experiments found in the literature (Table S2).
The predicted deleterious UCVs are predominantly in the core secondary structure elements of the BRCT domains, rather than in loops, particularly in the β sheet of BRCT-N (β1, β3, and β4), and in helix α′3 and the turn connecting helix α′1 and strand β′2 in BRCT-C (Figure 4).
We observed a patch of highly conserved amino acid residues that form a groove on the BRCA1 surface, on the opposite face from the known phosphopeptide binding cleft (Figure 5A–5C). These residues are T1684, T1685, H1686, K1711, W1712, and R1753. Both T1685 and H1686 have been shown to be highly sensitive to mutation, and our companion paper [14] contains new experimental evidence that R1753T has markedly reduced transactivation activity in both yeast and mammalian cells. The groove residues form hydrogen bonds with each other and several other conserved residues, including S1651, V1687, T1681, G1706, E1731, E1735, and P1749, producing two hydrogen bonding networks. The first network is found in BRCT-N (S1651, T1684, T1685, H1686, V1687, G1706, K1711, W1712, E1731) and the second network connects BRCT-N residues with the linker region that connects BRCT-N and BRCT-C (E1735, P1749, R1753) (Figure 5A–5C). All the residues lining this groove are completely conserved in our alignment of BRCA1 orthologs, except for T1684, which is conserved in all orthologs except for Tetraodon (pufferfish), the organism most distant from human in our alignment (Figure 5A). Previous studies have shown that G1706 and P1749 are also sensitive to mutation [44,45] (S. Marsillac, 2006, private communication). The proposed binding site would be specific to BRCA1, as most of these positions (except for H1686) are not highly conserved across tandem BRCT repeats in MDC1, PTIP, BARD1, and 53BP1.
The solvent-exposed missense variant R1753T found at the proposed binding site has <20% of the wild-type transactivation activity in yeast and <5% in mammalian cells [14], suggesting that the wild-type arginine amino acid residue might be important for binding of BRCA1 to a protein partner (or nucleic acid ligand). Although the mechanism of the BRCA1 BRCT domains in transactivation is not known, it is believed to depend on interactions with a variety of partners [46,47]. The mutation of R1753 to a threonine changes the local electrostatic surface potential from primarily positive and neutral (depicted as blue and white) to negative (red) (Figure 5D). This change may weaken the binding of protein partner(s) or nucleic acid ligand(s) necessary for transactivation.
We have developed an approach to rapid characterization of inherited missense variants in the BRCT domains of BRCA1 that is able to retrospectively predict the outcome of a functional transactivation assay with greater than 94% accuracy. Our method makes no a priori assumptions about which predictive features might best inform a classification algorithm. Rather, we hypothesize that given a large sample of deleterious and neutral variants, we can quantitatively measure the most informative predictors and learn to distinguish between these two functional classes with a supervised learning algorithm. We discuss (a) how our approach compares with similar work [7,12,18,19,48–50], (b) how prediction of deleterious variants can identify putative binding sites, (c) how computational classifiers can save time and money required for biochemical assays of many candidate variants, and (d) the possibilities of generalizing the methods to large numbers of disease-associated genes.
Supervised learning algorithms have previously been applied to predicting the functional impact of missense variants in TP53 with a four-body “potential” based on Delauney tessellation [51], to engineered variants in E. Coli Lac Repressor, HIV protease, and T4 bacteriophage lysozyme [24,52,53], and to large sets of single nucleotide polymorphisms [54–56]. Much of this work has been limited by overfitting problems. Benchmarking of Support Vector Machines and Decision Trees in several studies has shown that high numbers of false positive and false negative classification errors (~0.30) are generated when the learners are applied to proteins other than those in their training sets [24,52]. Supervised learning using four-body potentials is further limited in application, because each missense variant is represented by a profile of n features (amino-acid residue potential scores), where n is the number of amino acid residues in the protein. Supervised learning algorithms require fixed-length feature vectors; thus, an algorithm trained on missense variants represented by n features can only classify missense variants that are represented by n features. For example, if the training set is composed of missense variants in Lac Repressor (327 amino acid residues), the algorithm cannot be used to classify mutants in Lysozyme (164 amino acid residues).
Here we identify a set of 16 predictive features that, in combination with Support Vector Machine, Random Forest, and Naïve Bayes supervised learning algorithms, avoids the overfitting problem when the training set is composed of TP53 variants and the validation set is composed of BRCA1 missense variants. Initially, we used 31 deleterious and eight neutral BRCA1 BRCT variants that had been functionally tested as our training set. However, this approach yielded poor classification performance in a cross-validation test, presumably because of small sample size (unpublished data). As an alternative, we selected our features and performed supervised learning with a training set of 600+ artificially engineered TP53 missense variants. The ability of computational learning algorithms trained on TP53 variants to classify BRCA1 missense variants in agreement with the BRCA1 functional assay (94%+) suggests that mechanisms underlying structural and functional defects may be similar in TP53 and BRCA1.
In comparison, an approach based on sequence analysis and expected frequencies of structural features inferred from mutagenesis studies of E. Coli lac repressor and T4 lysozyme resulted in only 75% agreement with a BRCA1 BRCT trypsin sensitivity assay of 22 variants [12,50]. We find that the best supervised learners are in greater agreement with the BRCA1 transactivation assay than several sequence analysis methods and an empirically designed set of rules and thresholds [7].
The sequence analysis methods that incorporate physiochemical properties of amino acid residues as well as evolutionary conservation (Ancestral Sequence and Align-GVGD) are more accurate than SIFT, which only considers evolutionary conservation. A weakness of these methods is that, for purposes of classifying deleterious variants, there is no principled way to choose the optimal set of evolutionarily related sequences to align and analyze. In this work, we used sets of aligned sequences taken from published work (Ancestral Sequence) [9], the SIFT and Align GVGD webservers [18,57], and a deep alignment (out to the sea urchin Strongylocentrotus purpuratus) generated by the creators of Align GVGD. Different sequence sets produce different classifications of the variants, and choice is biased by available genomes and decisions about appropriate thresholds of relatedness. The problem is illustrated with classifications of the BRCA1 BRCT missense variant V1665M. Align-GVGD Tnig, Align-GVGD Spur, and Ancestral Sequence incorrectly classify V1665M as deleterious, because valine is completely conserved in their multiple sequence alignments. In contrast, SIFT constructs an alignment that includes two Arabidopsis proteins containing BRCT domains (UNIPROT Q9ZWC2, Q3E7F4) with a methionine aligned at this position, and thus correctly classifies the variant as neutral.
SIFT uses Dirichlet mixture pseudocounts to estimate allowed amino acid residues at each alignment position [29]. Pseudocount approaches compensate for incomplete sequence sampling in a multiple sequence alignment by adding counts for imaginary amino acid residues that are statistically likely to occur at each position. Our analysis indicates that several SIFT errors are the result of poorly estimated pseudocounts (Table S1). Importantly, SIFT is the only sequence analysis method that includes an automated alignment algorithm. The accuracy of Ancestral Sequence and Align GVGD depends on manual sequence selection, so these methods cannot automatically be applied to other cancer susceptibility genes or to whole genome analysis.
Despite their limitations, sequence analysis methods have the advantage that they can be applied to any position in BRCA1, whereas our supervised learners require protein structure information and are thus limited to regions for which accurate protein structure or structural models are available. For example, the BRCA1 BRCT missense variant E1794D is misclassified by Support Vector Machine, Random Forest, Rule-based decision tree, and Decision Tree. Analysis of X-ray crystal protein structure quality with MolProbity and PROCHECK [58,59] indicates that in the 1t29 structure, the atomic coordinates at this position, may be incorrect. The corresponding backbone dihedral angles Φ and Ψ of 143.8° and 112.9°, respectively, are statistical outliers. Using sequence analysis alone, with multiple sequence alignments in which Tetraodon BRCA1 contains an aspartic acid residue at this position, Align-GVGD, Ancestral Sequence, and SIFT correctly classify the variant as neutral.
The empirically designed rules (Rule-based decision tree) are more accurate than a Decision Tree of rules learned by a supervised algorithm but not as accurate as Support Vector Machine, Random Forest, Naïve Bayes, Ancestral Sequence, or Align-GVGD (Table 1 and Figure 1). It appears that it is difficult to correctly set thresholds and weigh the relative importance of empirically designed rules. For example, Rule-based decision tree classifies A1669S as deleterious, in disagreement with the functional assay (Figure 1 and Table S1). Rule-based decision tree uses a “mutation likelihood” rule that would classify A1669S as neutral because there is a serine in frog BRCA1 at this position. It also uses a rule that would classify A1669S as deleterious because the amino acid residue position is buried in the protein core and close to a buried charged residue. The order is such that the latter rule dominates the final decision. In contrast, the supervised learning approach makes no a priori decisions about thresholds or predictor ordering, but learns this information implicitly during the training process. Given an informative training set, such as the TP53 variants used in the present work, it is able to make highly accurate decisions about the BRCA1 variants.
The supervised learning algorithms described here make classification decisions based on nonlinear combinations of predictive features and fail to provide rationalizations that can be understood by humans. To address this issue, we apply a post-predictive step in which we analyze protein structure models and alignments. For example, the wild-type arginine amino-acid residue at position 1753 of BRCA1 forms a salt bridge with the glutamate at position 1735. The arginine is found in the linker region that connects the BRCT-N and BRCT-C domains and the glutamate is found in BRCT-N (Figure 4). This charge–charge interaction may be important for the stability of the BRCT homodimer. In our structural model of the threonine variant, the salt bridge is broken, potentially destabilizing the homodimer. Both R1753 and E1735 are completely conserved in our alignment of 13 BRCA1 orthologs (Figure 5), suggesting possible selective pressure to preserve their pairwise interaction. Such rationalizations increase our confidence in the predictions and suggest ways to test them experimentally—for example, by site-directed mutagenesis. Fifty-four percent of the variants can be explained by combining structural and evolutionary analysis in post-prediction (Tables S1 and S2).
Several studies of the BRCA1 BRCT domains have suggested that there may be surface patches that interact with protein partners [60–63]. In previous work, we predicted that a groove formed by both BRCT repeats (near nonfunctional variants L1657 and K1702) and the ridge that delimits the groove (near nonfunctional variant E1660) constitutes such a surface patch [7]. Our prediction was subsequently confirmed through X-ray crystallographic studies, which revealed a site where the phosphorylated peptides of BACH1 and CtIP have been found to bind [64,65]. Importantly, several missense variants were found to disrupt this interaction [64,65], suggesting that clustering of deleterious variants at solvent-exposed amino acid residue positions is indeed a useful indicator of binding site location. There is a large literature on the general topic of the relationship between deleterious mutations and binding sites [66–74].
Two surface variants found to be deleterious to BRCA1 transactivation activity in our companion paper (R1753T and T1685I) [14] lie on a highly conserved patch of amino acid residues, forming an exposed groove. The R1753T variant yields a changed electrostatic surface potential, which may be sufficient to disrupt the binding of BRCA1 to a protein partner or nucleic acid ligand important for transactivation. Following the logic that predicted the BACH1/ CtIP binding site, we suggest that this groove may be a previously uncharacterized binding site, whose disruption inactivates BRCA1 transactivation function. Accordingly, we are currently testing the binding of several candidate protein partners to the predicted binding site using site-directed mutagenesis and a yeast two-hybrid assay.
Biochemical assays can play an important role in identifying deleterious UCVs in cancer susceptibility genes [75], but the work is labor-intensive and time-consuming. We estimate that, on average, an assay for one BRCA1 UCV in a mammalian cell system costs US$125–US$150 and requires three weeks of personnel time, from ordering primers to final results. The time can be reduced by processing the variants in batches. To assay the 54 uncharacterized BRCT BRCA1 missense variants found in the BIC database (April 2006) would take approximately 18 months of personnel time. Accurate computational classification of UCVs can significantly reduce the required time by prioritizing UCVs most likely to be deleterious. Importantly, while computational classification and functional assays can contribute to medical decision making, other factors such as family history, co-occurrence with known deleterious mutations, and studies of patient tumor tissue will continue to be important in a clinical setting.
We have applied supervised learners trained on the TP53 variant set to prediction of UCVs in BRCA2, with promising results (unpublished data). We are currently exploring whether this training set and our current set of features can be used to evaluate UCVs in other genes associated with familial cancer syndromes: MLH1, MSH2, MSH6 (hereditary nonpolyposis colon cancer), APC (familial adenomatous polyposis), MYH (MYH adenomatous polyposis), and P16 (melanoma). We have applied a modified version of this method to classify all human amino-acid changing SNPs found in the dbSNP database [76] as deleterious or neutral [56]. The SNPs were classified with a support vector machine trained on amino acid residue substitutions from more than 1,500 human proteins. Because X-ray crystal structures are not available for most human proteins [77], we built homology models with an automated modeling pipeline MODPIPE that relies on the MODELLER package for fold assignment, sequence-structure alignment, model building, and model assessment [78]. A small number of these predictions have been validated by biochemical and epidemiological studies found in the literature.
We are exploring the extent to which a decision rule learned with a training set of variants from one protein, such as TP53, can be generalized to variants from other proteins. One possibility is that most deleterious missense mutants do not affect specific binding interactions, but are instead slightly destabilizing [79]. If this is true, a training set of missense variants from a protein with similar stability to the protein of interest may be the best choice. Other possiblities include training on a protein sharing GO terms, or from the same fold family (all-alpha, all-beta, alpha-beta, etc.) as the protein of interest. We are working on generating large variant datasets from selected proteins to test these hypotheses.
In summary, we have systematically and comprehensively evaluated structure- and sequence-based computational prediction methods applied to variants in the BRCA1 BRCT domains and developed detailed structural explanations for the measured and predicted impact of 49 BRCA1 variants. When combined with 16 carefully selected predictive features, the best-supervised learning algorithms are in greater agreement with experimental results than has been reported previously.
The increased use of sequencing methods to genotype individuals at risk for inherited cancers and the observation that sequence variation is greater in ethnic minorities than in Caucasians highlight the need for improved methods of UCV risk assessment. Bioinformatics approaches including supervised learning algorithms, protein structure modeling, and evolutionary sequence analysis can contribute to an integrated approach to risk assessment by increasing coverage of classified UCVs more rapidly than is possible by functional assays. In the future, when clinicians counsel patients about their cancer risk, they will be able to take advantage of these bioinformatics prediction methods. Finally, successful generalization of these methods to a large number of disease-associated genes will play an important role in reducing the growing number of loci, variants, and phenotypes that confound modern whole genome disease-association studies.
Accession numbers from the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov) are shown in Table 3. |
10.1371/journal.pgen.1006168 | Two-Step Regulation of a Meristematic Cell Population Acting in Shoot Branching in Arabidopsis | Shoot branching requires the establishment of new meristems harboring stem cells; this phenomenon raises questions about the precise regulation of meristematic fate. In seed plants, these new meristems initiate in leaf axils to enable lateral shoot branching. Using live-cell imaging of leaf axil cells, we show that the initiation of axillary meristems requires a meristematic cell population continuously expressing the meristem marker SHOOT MERISTEMLESS (STM). The maintenance of STM expression depends on the leaf axil auxin minimum. Ectopic expression of STM is insufficient to activate axillary buds formation from plants that have lost leaf axil STM expressing cells. This suggests that some cells undergo irreversible commitment to a developmental fate. In more mature leaves, REVOLUTA (REV) directly up-regulates STM expression in leaf axil meristematic cells, but not in differentiated cells, to establish axillary meristems. Cell type-specific binding of REV to the STM region correlates with epigenetic modifications. Our data favor a threshold model for axillary meristem initiation, in which low levels of STM maintain meristematic competence and high levels of STM lead to meristem initiation.
| In seed plants, branches arise from axillary meristems (AMs), which form in the crook between the leaf and the stem. How AMs initiate to produce branches remains unclear. In this study, we show that a group of meristematic cells maintain expression of the meristem marker SHOOT MERISTEMLESS (STM); the progeny of these cells form the axillary buds. Our results suggest that low-level STM expression is required (but not sufficient) for AM initiation, and that high-level STM expression induces initiation of the AM. The initial expression of STM requires the auxin minimum in the leaf axil and the transcription factor REVOLUTA directly up-regulates STM expression.
| In plants, many somatic cells can regenerate into complete plants; thus, many plant cells are considered totipotent, unlike most somatic cells in animals [1]. Plants also show well-defined developmental patterning, which leads to questions about how cell fates become established. Specialized cell lineages generate guard cells or pavement cells in the leaf epidermis [2], and produce callus during regeneration [3]; both of these cell types have similarities to animal stem cell lineages. Much less is known about cell fate determination in other aspects of plant development.
An iconic feature of plants is their branching growth habit, an innovation considered crucial for their conquest of land [4, 5]. Plants maintain meristems with undifferentiated stem cells, which are responsible for the life-long organogenesis of growing plants. Branching occurs by periodic initiation of new meristems. In the seed plants, secondary growth axes arise from axillary meristems (AMs, also termed lateral meristems) in or near the adaxial side of leaf axils [6–8]. During AM initiation, a morphologically detectable bump forms in the leaf axil and develops into a bud [9–11]. Two models have been proposed to explain AM initiation. The ‘detached meristem’ model proposes that a few pluripotent cells detach from the primary shoot apical meristem (SAM) and associate with the leaf axil as the leaf differentiates from the SAM [10, 12]. Histological analysis shows that leaf axil cells likely remain undifferentiated, providing support for the detached meristem theory [12, 13]. Analysis of the Arabidopsis thaliana phabulosa-1d mutant led to the alternative ‘de novo induction’ model [14], in which an AM initiates from differentiated leaf cells. A major difference between these models is whether AM initiation requires a meristematic cell lineage [10, 15, 16].
Although the origin of AMs is presently unclear, genetic studies in Arabidopsis have shown that AM initiation is regulated by several transcription factor-encoding genes, such as LATERAL SUPPRESSOR (LAS), REGULATOR OF AXILLARY MERISTEMS, CUP-SHAPED COTYLEDON (CUC), and REGULATOR OF AXILLARY MERISTEM FORMATION [11, 17–20]. Genetic and molecular studies revealed direct and indirect interactions among these genes in a regulatory network [18, 21]. Many of these genes have conserved functions in the regulation of AM initiation in dicots and monocots, such as tomato (Solanum lycopersicum), maize (Zea mays), and rice (Oryza sativa) [22–26]. Phytohormones also regulate AM initiation, which requires an auxin minimum and a subsequent cytokinin signaling pulse [9, 27, 28].
Here, we asked whether post-embryonic AM initiation requires meristematic cells with a fixed developmental fate [10, 11, 15], and how these cells are regulated. Our results show that initiation of branching meristems in the shoot requires a meristematic cell population embedded in differentiated cells. Examination of the fine-tuning of these cells led to a threshold model for AM initiation.
Previous in situ hybridization results showed STM expression in all stage leaf axils from examination of fixed samples, but it remains unclear if a continuous STM-expressing cell population exists during development [10, 11, 16]. The STM-expressing cells are closer to the meristem side of the boundary in young leaf primordia, and are closer to the leaf side of the boundary in older leaves. Thus, it has been proposed that the initial STM-expressing cells may create a separation while their neighboring cells re-differentiate as AM progenitor cells [10]. To better resolve the origin of STM expressing cells, we used live-cell imaging to determine if a continuous STM-expressing cell population exists in the leaf axil. To this end, we imaged axils of living leaf primordia that we isolated from the shoot apex and maintained in culture. As shown previously, cultured leaf primordia (P6 and older) efficiently initiated AMs in the absence of exogenous phytohormone [9], which is distinct from de novo organogenesis [29].
By live-imaging the expression of a functional pSTM::STM-Venus reporter in P6 and older leaves (Fig 1A), we found that cells with continuous STM-Venus expression are AM progenitors. A recent study has shown that this reporter line can fully complement stm mutation, and the enhanced boundary expression reflects the endogenous STM expression [30]. For the P8/9 leaf primordium, which has the fewest STM-expressing cells of all the stages (see below), we observed the STM-Venus signal only in a continuous cell mass close to the incision line (Fig 1B). The number of cells with STM-Venus signal initially decreased after 24 h of culture (Fig 1C), but partially recovered after 48 h of culture (Fig 1D). Starting from 72 h of culture, following a series of rapid cell divisions (Fig 1E), these cells organized into a meristem with new leaf primordia (Fig 1F and 1G). Occasionally, a few cells without initial STM-Venus signal at the first time point, but next to one or more STM-positive cells, showed detectable signal. These cells could have initially had low-level STM-signals below our detection threshold. Alternatively, their STM expression could be due to STM proteins trafficked from neighboring STM-expressing cells [31].
In early stage leaf axils, STM-Venus also persisted in the boundary region. In tissue sections, the number of STM-expressing cells gradually decreased during leaf primordia maturation from P3 (Fig 1H) to P9 (Fig 1I). Later, in P10 and older leaves, the number of STM-expressing cells and level of STM expression increased significantly (Fig 1J), which is consistent with the live-imaging results. Quantitative measurement of leaf axil STM-Venus fluorescence intensity confirmed this STM expression dynamic pattern during leaf maturation. In particular, P8 or P9 have the lowest intensity in Ler, which significantly increase in the next developmental stage (Fig 1K–1N and S1A–S1C Fig). There is a small variation between individual plants with either P8 or P9 having the lowest STM expression, which is in line with a previous morphological analysis (10).
In addition, we also used reporter lines to follow expression of the shoot meristem central zone stem cell marker CLAVATA3 (CLV3), the shoot meristem organizing center marker WUSCHEL (WUS), and the pericycle-like cell marker J0121, which marks progenitor cells for regeneration [3, 32, 33]. We did not detect CLV3 or WUS expression in young Ler leaf axils until the twelfth-youngest primordium (P12, S1D–S1F Fig) and J0121 was not expressed at all in the leaf axil during axillary bud formation (S1G Fig), suggesting that their corresponding cell identities were not maintained.
We next tested whether AM initiation required the STM-expressing cells. It has been reported that mild stm alleles have more active branching and show an ‘abort-retry’ mode of growth [34–36]. However, because SAM termination (in stm mutants) promotes outgrowth of axillary buds, branch growth may not reflect axillary bud formation. To show if STM functions in AM initiation, we analyzed the pattern of axillary bud formation in plants carrying the weak stm-bum1 allele, which can still form leaves from a partially functional SAM [37]. We found a dramatic reduction in the number of axillary buds in stm-bum1 plants, with 60% (242 out of 405) of leaves lacking axillary buds, which is distinct from Col-0 wild-type plants (Fig 2A). This reduction in axillary buds is more dramatic for rosette leaves (91% leaves lack buds). In contrast to the wild type (Figs 1I and 3A–3C), leaf axil cells in stm-bum1 plants are enlarged (Fig 3D–3F), suggesting that the leaf axil cells have undergone differentiation. On the other hand, stm-bum1 plants have reduced apical dominance, resulting in the reported enhanced branching phenotype [34–36].
To test if AM initiation requires STM-expressing cells, we applied laser ablation. When we ablated the cells adjacent to the STM-expressing cells, AMs initiated normally from the ablated leaf axil region (Fig 2B–2D, 16 out of 19), showing that ablation per se does not abolish AM initiation. However, after ablation of most cells within the STM-expressing cell mass in both the epidermis and internal cell layers, AMs could not initiate (Fig 2E–2G, 10 out of 10). Note that this result is in contrast to observations in shoot and root apical meristems [38–41], where neighboring cell fate can switch after ablation. Furthermore, we observed that AMs did not initiate from cultured leaves if we removed the proximal portion of the petiole containing the STM-expressing cells (S2A–S2C Fig). Taken together, our data strongly suggest that AM initiation requires the STM-expressing cells as AM progenitor cells.
We next asked what regulates the maintenance of STM expression in the leaf axil. We have recently shown that the AM progenitor cells also maintained a low auxin level [9, 27], suggesting that maintenance of STM expression may require the leaf axil auxin minimum. To test this hypothesis, we analyzed STM expression in pCUC2>>iaaM and pLAS::iaaM-en plants, which ectopically accumulate auxin in leaf axils and are deficient in AM initiation [9, 27]. We could not detect STM expression in leaf axils of pCUC2>>iaaM plants (Fig 3G–3J). In addition, leaf axil cells in pCUC2>>iaaM plants are enlarged (Fig 3K–3O), suggesting cell differentiation. Similarly, pLAS::iaaM-en plants also have substantially reduced or undetectable leaf axil STM expression and have enlarged leaf axil cells (Fig 3P–3S).
To test if STM expression alone is sufficient for AM initiation, we introduced p35S::STM-GR into pCUC2>>iaaM plants. In p35S::STM-GR plants [42], dexamethasone (Dex) can induce the nuclear translocation of a STM-glucocorticoid-receptor (GR) fusion protein. We aimed to test if leaf axil cells that have lost STM expression can respond to ectopic STM activity. Firstly, we detected a dramatically increase of STM expression by reverse transcription quantitative PCR (RT-qPCR) in leaf-removed leaf axil-enriched shoot apex tissues (Fig 3T). In mature leaf axils, we found that, following Dex induction, no axillary bud could form (Fig 3U, 3V and 3Y), highlighting the importance of the low level STM expression for subsequent AM initiation. Similarly, when we introduced p35S::STM-GR into stm-bum1 plants with compromised AM initiation, we found that Dex treatment did not induce axillary buds from mature leaf axils (Fig 3Y). Taken together, these results indicate that the recently identified leaf axil auxin minimum is required to maintain low level STM expression, which is then required for later axillary buds formation.
In contrast to pCUC2>>iaaM and pLAS::iaaM-en plants, we found that STM expression was maintained in the rev-6 mutant (Fig 4A–4E), which also lacks axillary buds [43]. In contrast to the wild type (Fig 1K–1M), the expression of STM does not increase in the rev-6 mutant (Fig 4A and 4B), as it does in wild-type leaf axils, during leaf maturation (compare Figs 1B–1G and 4C–4E). Subsequently, the STM-expressing cells did not undergo active cell division to form a meristem with well-organized structure (Fig 4C–4E). The change of leaf axil STM expression in rev-6 implies that up-regulation of STM expression in P10 and older leaves requires REV, but maintenance of STM expression does not require REV. Also in contrast to stm-bum1 and pCUC2>>iaaM, our genetic analysis indicates that over-expressing STM can suppress the AM initiation defect of rev-6 mutants (Fig 3W–3Y). Therefore, we conclude that STM expression must not only be maintained in meristematic cells, but also subsequently up-regulated for AM initiation.
To test if REV up-regulates STM expression in a cell-autonomous manner, we imaged REV distribution by using a functional pREV::REV-Venus reporter line [32]. REV-Venus is broadly expressed in the adaxial side of P8 and younger leaves (Fig 4F–4I), but it is restricted to the center of leaf axils, especially the epidermis (L1) layer, in P9 and older leaves (Fig 4J–4L). Furthermore, REV has stronger expression in P9 and older leaf axils than in younger leaf axils. The leaf axil enrichment of REV is consistent with up-regulated STM expression in P10 and older leaves, suggesting that REV up-regulates STM expression in a cell-autonomous manner.
Overexpressing alleles of REV and related HD-ZIPIII genes can induce ectopic AMs in the abaxial side leaf axils [14, 44, 45]. By using one such mutant, phavulota-1d (phv-1d), we observed ectopic STM expression in abaxial leaf axils prior to axillary bud initiation (S3B–S3E Fig). By using transgenic lines overexpressing microRNA-insensitive REV and PHABULOSA (PHB), another related HD-ZIPIII gene, we detected up-regulation of STM expression in leaf-removed shoot apex tissues, which are enriched with leaf axils (S3J Fig). Notably, we also detected ectopic auxin minima in abaxial leaf axils by using the auxin concentration sensor DII-Venus [46], whose strong abaxial axil signal indicates low auxin concentrations (S3F–S3I Fig). Taken together, REV and related HD-ZIPIII proteins can promote STM expression, which, together with auxin minima, promote ectopic AM initiation.
To test if REV directly up-regulates STM expression, we generated functional Dex-inducible pREV::REV-GR-HA rev-6 lines (S4A–S4C Fig). We measured the effect of REV activation on the expression of STM by RT-qPCR. REV activation resulted in rapid elevation of STM mRNA levels within 2 h of treatment, with or without the protein synthesis inhibitor cycloheximide (CHX) (Fig 5A and S4D Fig), strongly suggesting that induction of STM does not require de novo protein synthesis and that STM is likely a direct target of REV. REV activation also triggered in vivo accumulation of STM-Venus, as shown by live-cell imaging (S4E–S4H Fig). Consistent with this, our recent large-scale yeast one-hybrid assay identified REV and related HD-ZIPIII proteins as binding to the STM promoter region [21].
We next performed chromatin immunoprecipitation (ChIP) assays to examine whether REV directly binds to the STM promoter in vivo. We scanned the STM genomic sequence for ATGAT, the conserved binding site for REV [47], and designed primers near identified motifs and other regions (Fig 5B). In both shoot apex tissues enriched with leaf axils and inflorescence tissues, we found that REV-GR-HA strongly associated with the regions containing multiple ATGAT motifs, but only after Dex treatment, by using antibodies against GR or HA (Fig 5C and S4I Fig). In addition, REV-GR-HA weakly associated with seven other upstream ATGAT motif-containing regions. A transient transfection assay in protoplasts further confirmed that REV bound to multiple ATGAT motif-containing STM genomic regions, especially the ones close to the start codon, and up-regulated STM expression (Fig 5E). These newly transformed pSTM::LUC constructs would lack epigenetic modifications that might interfere with REV binding.
REV is widely expressed in young leaves, including in the adaxial domain and vascular tissues [43], but only up-regulates STM expression in boundary tissues-enriched samples (Fig 6A), as previously shown [11]. Furthermore, a recent ChIP-seq analysis did not identify STM as a REV-binding target in whole seedlings [47]. By using the same antibodies and protocol, we found that REV associated with the STM genomic region only in vegetative shoot apex and inflorescence tissues, but not in leaf axil region-removed mature leaves (compare Fig 5C and 5D, only data from vegetative shoot apex tissues was shown).
In animals, lineage-specific epigenetic modification of transcription factor genes leads to the fixation of stem cell fate [48]. Furthermore, the STM locus was epigenetically silenced in mature leaves containing only differentiated cells [49–51]. In mature leaves without STM-expressing leaf axil cells, the chromatin modification H3K27me3, which is associated with transcriptional repression, is highly enriched at the STM locus (Fig 6B and 6D). By contrast, H3K4me2 and/or H3K4me3, which are associated with transcriptional activation, are enriched at the STM locus in inflorescence tissues enriched with organ axils (Fig 6C and 6D). This histone modification pattern implies that epigenetic factors may regulate REV binding to the STM locus.
In both animals and plants, the Polycomb Repressive Complex 2 (PRC2) establishes the H3K27me3 mark, which provides a docking site for PRC1 to establish a repressive chromatin configuration [52]. Mutants affecting PRC1 and PRC2 have elevated STM expression [49, 50]. To test if the ectopic activation of STM expression requires REV, we introduced rev-6 into PRC mutants. We found that rev-6 mutation partly suppressed ectopic STM up-regulation (Fig 6E).
Plant cells, especially isolated cells, have amazing developmental plasticity, yet intact plant development follows defined patterning. Within meristems, clonal analysis and root regeneration studies suggest that meristem cells usually lack predictable destinies and that positional control is most important for plant cell fate determination [38–41]. However, distinct cell lineages emerge at later developmental stages [2, 3]. In this study, we show that AM initiation is accompanied by the maintenance of a meristematic cell population, and differentiation of surrounding cells. We traced this cell population in P6 and older leaves, and confirmed that STM-positive cells at the leaf axil are progenitors of axillary buds. Imaging results indicate that cells usually cannot acquire STM expression de novo (Fig 1B–1G), at least in P6 and older leaves, indicating the existence of a cell lineage. Further laser ablation results show that these STM-positive cells are necessary for formation of axillary buds, whereas neighboring STM-negative cells are differentiated (Fig 2B–2G).
This leaf axil meristematic cell population relies on positional cues. Our recent studies have shown that an auxin minimum, which is associated with the leaf axil position, is required for AM initiation [9, 27]. In the current work, we further demonstrate that the maintenance of the meristematic cells depends on low auxin (Fig 3G–3J and 3P–3S), which is likely determined by positional information. The observation of abaxial auxin minima and STM expression in phv-1d, which forms axillary buds at the abaxial side, also support the importance of positional cues for the maintenance of meristematic cells (S3B–S3I Fig).
Cell fate determination occurs gradually with cell cycle progression in animals [48]. Previous studies focused on cells within shoot meristems and root meristems, and found that cells from different meristematic domains can switch cell fate [38–41]. These results do not necessarily indicate that cell fate determination does not occur after additional rounds of cell cycle progression. In fact, root meristem regeneration does not occur if additional tissue beyond the meristematic zone has been removed [53]. If one assumes that cell fate determination takes place after more cell cycles in plants (than in animals), it would be conceivable that: i) cells within or close to meristems remain meristematic and can reverse cell fate, and ii) certain non-dividing or slow-dividing cell types in differentiated organs may maintain a meristematic status while their neighboring cells become fully differentiated and can no longer reverse to a meristematic status. Because boundary cells are non-dividing or slow-dividing cells [54], leaf axil cells can maintain an undifferentiated status while their neighboring differentiated cells cannot.
Previous studies have shown that overexpression of STM (or the related gene KNAT1) alone [55–57], or in combination with ectopic WUS [42, 58], induces ectopic meristems. The effect of ectopic STM is highly dependent on tissue stage. As shown in a previous works [57], leaf primordia older than P10 are not competent to ectopic STM activity (S5A Fig). For younger leaf primordia, ectopic meristems initiated only from leaf axils and the adaxial side of the proximal portion of leaf blades, especially in the sinus region between the blade and the petiole (55 out of 72 P7 to P9, i.e. 72%, S5B–S5G Fig). Thus, STM alone is not sufficient to induce meristems from most cells, but is sufficient in presumably undifferentiated cells.
Similarly, we found that ectopic STM activity was insufficient to rescue axillary bud formation defects in mature leaf axils of pCUC2>>iaaM, pLAS::iaaM-en, or stm-bum1 plants, which have lost low level STM expression in leaf axil cells. By contrast, ectopic STM activity was sufficient in rev-6 maintaining low level STM expressing cells (Fig 3W and 3X). Therefore, STM expression and proper cell fate are both required for AM initiation. In tomato, recent work showed that ectopic meristems may form at the base of leaflets, where KNOX genes express, and this requires the AM initiation pathway [59]. This finding again supports the model that cell competency is required for shoot meristem formation. Epigenetic regulation is involved in the maintenance of meristematic cell competency, and STM expression serves as a marker for cell competency.
Our data support the detached model for AM initiation, in which meristematic cells are detached from the SAM. When leaf primordia (P1) separate from the SAM, boundary cells keep STM expression as SAM cells (Fig 1H). From P1 to P5 stages, the number of STM-expressing cells continues to decrease (Fig 1H–1N). Although the exact clonal relationship of early STM-expressing cells remains unknown, our data suggest that many STM-expressing cells differentiate but some may maintain STM expression. In P6 and older leaf primordia, we used live-cell imaging to track the STM-expressing cell population (Fig 1B–1G). Notably, all cells in the enlarged STM-expressing domain are progeny of cells with previous STM expression. Our data also explain the ectopic axillary formation of phv-1d mutants. Ectopic axillary buds form in the abaxial side away from the SAM, providing key support to the de novo model [14]. PHV is highly similar to REV, and can bind to the STM promoter region in yeast. It is conceivable that ectopic PHV expression in phv-1d would result in ectopic STM expression (S3E Fig), resulting in ectopic meristematic cells in the abaxial leaf axil that initiate ectopic axillary buds.
Furthermore, our results support a ‘threshold model’ in which maintenance of low levels of STM expression is required but not sufficient for AM initiation, and subsequent elevated expression of STM would induce AM initiation (Fig 7). Leaf axil cells show low auxin-dependent low levels of STM expression starting at leaf primordium initiation. The early low level STM expression is required for later AM formation (Fig 3). In addition, cells lost STM expression are no longer sensitive to ectopic STM activities at a later stage. Before AM initiation, STM is up-regulated in the center of the leaf axil, triggered by REV activation, which in turn requires LAS activity [11]. We also show that this up-regulation is a local event (Fig 4I–4L), and it depends on prior, maintained STM expression (S4E–S4H Fig). We further show that REV binding to the STM promoter is tissue-specific (Fig 5), and that epigenetic regulation may underlie this cell type specificity (Fig 6), suggesting that the binding requires permissive chromatin statues. Our data favor the idea that the up-regulation of STM is causal for AM initiation, rather than a consequence of a newly formed AM, because the expression of WUS and CLV3 are still missing at the stage of initial STM up-regulation.
The Arabidopsis thaliana ecotypes Landsberg erecta (Ler) or Columbia (Col-0) were used as the wild type. The atring1a atring1b, emf2-11, clf-29, clf-29 swn-21, rev-6, stm-bum1, p35S::REVm-MYC, p35S::PHBm-MYC and pREV::REV-GR-HA lines are in the Col-0 background [37, 50]; the pCLV3::GFP-ER pWUS::DsRed-N7, pREV::REV-Venus, pSTM::STM-Venus, and p35S::STM-GR lines are in the Ler background [32, 42, 60], and the J0121 line is in the Ws-0 background. In this study, we confirmed that the pSTM::STM-Venus reporter can rescue the stm-11 mutant phenotype. Genotyping primers are listed in S1 Table. Plants were grown in the greenhouse on soil at 22°C under short-day conditions (8 h light/16 h dark) unless otherwise specified.
Leaf culture followed a previously described protocol [9]. Briefly, seedlings were grown in MS medium under short-day conditions for 15 d after seed stratification. Leaves between P5 and P11 were then detached from seedlings, laid flat on MS medium supplemented with 0.5 mg/L folic acid and 100 mg/L inositol, and grown for up to 30 d under the same conditions.
The pREV::REV-GR-HA construct was made by replacing the endogenous stop codon of TAC clone JAtY80N08 covering the REV genomic region with a GR-HA sequence using recombineering [61]. The construct was then transformed into rev-6 plants. Over 20 transgenic lines were obtained and lines with stringently inducible rescue phenotypes were used.
Confocal microscopy images were taken with a Nikon A1 confocal microscope. Samples were either live-imaged or fixed and sectioned as previously described [9]. Excitation and detection wavelengths for GFP, Venus, and DsRed were as previously described [9, 32]. To detect FM4-64 and PI staining, a 514 nm laser line was used for excitation and a 561 nm long-pass filter was used for detection. The modified pseudo-Schiff-PI (mPS-PI) staining was performed as described and a 488 nm laser line was used for excitation and emission was collected at 520–720 nm [62]. DAPI staining was excited at 405 nm and detected in the 425–475 nm. Autofluorescence was excited at 488 nm or 514 nm and detected in the 660–700 nm range.
Optical photographs were taken with a Nikon SMZ1000 stereoscopic microscope or an Olympus BX60 microscope equipped with a Nikon DS-Ri1 camera. Scanning electron microscopy was performed using a Hitachi S-3000N variable pressure scanning electron microscope after standard tissue preparation [9].
Laser ablations were performed on a Nikon A1 confocal microscope equipped with an Andor MicroPoint laser system consisting of a pulsed 440 nm nitrogen laser. We adjusted a variable neutral density filter to attenuate the output laser to limit damage to targeted cells, as assessed by confocal imaging. The observation of cell collapse was used to confirm successful ablation (Fig 2C and 2F).
Total RNA was extracted from leaves, shoot apex tissues, or inflorescences (~6 d after bolting) of 12 plants using the AxyPrep Multisource RNA Miniprep kit (Corning). For shoot apex tissues enriched for leaf axils, leaves were manually removed from 25 d plants grown under short day conditions. For induced meristems, total RNA was extracted from leaf sinus tissues using the RNAqueous-4PCR kit (Life Technologies). First-strand cDNA synthesis was performed with 2 μg total RNA using TransScript One-Step gDNA Removal and cDNA synthesis SuperMix (TransGen), or with 300 ng total RNA using SuperScript III reverse transcriptase (Life Technologies), and 22-mer oligo dT primers according to the manufacturer’s instructions. RT-PCR analysis was performed in a 20 μL reaction using Taq DNA polymerase (TianGen) and gene-specific primers (S1 Table). Reverse transcription quantitative PCR (RT-qPCR) was performed on a Bio-Rad CFX96 real-time PCR detection system with the KAPA SYBR FAST qPCR kit (KAPA Biosystems). Relative expression by RT-qPCR was normalized to TUB6 (At5g12250). Gene-specific primers (S1 Table) were used to amplify and detect each gene. Error bars of RT-qPCR experiments in Figures are derived from three independent biological experiments, each run in triplicate.
Chromatin immunoprecipitation (ChIP) experiments were performed according to published protocols [28, 49]. Leaf axil-enriched shoot apex tissues (under short day conditions) or inflorescences (under long day conditions) of approximately 4-week-old Col-0 wild-type or pREV::REV-GR-HA rev-6 plants were used. Plant material (800 mg) was harvested and fixed with 1% (v/v) formaldehyde under vacuum for 10 min. Nuclei were isolated and lysed, and chromatin was sheared to an average size of 1000 bp by sonication. The sonicated chromatin served as input or positive control. Immunoprecipitations were performed with a polyclonal antibody against GR (Affinity Bioreagents, PA1-516), a monoclonal antibody against HA (Beyotime, AH158), a polyclonal antibody against H3K27me3 (Millipore, 07–449), or a polyclonal antibody against H3K4me2/3 (Abcam, ab8580). The precipitated DNA was isolated, purified, and used as a template for PCR. RT-PCR was performed as described above (S1 Table). The data are presented as degree of enrichment of STM genomic fragments. The amount of precipitated DNA used in each assay was determined empirically such that an equal amount of ACT2 (At3g18780) was amplified. Two independent sets of biological samples were used.
To produce the effector constructs, full-length REV was amplified from Arabidopsis cDNA and inserted into the pBI221 vector to generate pBI221-AP1. To generate STM promoter-driven LUC reporter genes, STM promoter regions were amplified from Arabidopsis genomic DNA. PCR fragments were inserted into the corresponding sites of the YY96 vector to produce pSTM::LUC constructs (Fig 5E, and S1 Table for primers).
Isolation of Arabidopsis protoplasts and PEG-mediated transfection were performed as described previously [28]. The reporter construct, effector plasmid, and a p35S::GUS construct (internal control) were co-transformed into protoplasts. After transformation, the protoplasts were incubated at 23°C for 12–15 h. The protoplasts were pelleted and resuspended in 100 μL of 1 × CCLR buffer (Promega). For the GUS enzymatic assay, 5 μL of the extract was incubated with 50 μL of 4-methylumbelliferyl-β-d-glucuronide assay buffer (50 mM sodium phosphate pH 7.0, 1 mM β-d-glucuronide, 10 mM EDTA, 10 mM β-mercaptoethanol, 0.1% sarkosyl, 0.1% Triton X-100) at 37°C for 15 min, and the reaction was stopped by adding 945 μL of 0.2 M Na2CO3. For luciferase activity assays, 5 μL of the extract was mixed with 50 μL of luciferase assay substrate (Promega), and the activity was detected with a Modulus Luminometer/Fluometer with a luminescence kit. The reporter gene expression levels were expressed as relative LUC/GUS ratios. Error bars in Fig 5E are derived from three independent biological experiments, each run in triplicate.
|
10.1371/journal.pgen.1004331 | Retinoic Acid-Related Orphan Receptor γ (RORγ): A Novel Participant in the Diurnal Regulation of Hepatic Gluconeogenesis and Insulin Sensitivity | The hepatic circadian clock plays a key role in the daily regulation of glucose metabolism, but the precise molecular mechanisms that coordinate these two biological processes are not fully understood. In this study, we identify a novel connection between the regulation of RORγ by the clock machinery and the diurnal regulation of glucose metabolic networks. We demonstrate that particularly at daytime, mice deficient in RORγ exhibit improved insulin sensitivity and glucose tolerance due to reduced hepatic gluconeogenesis. This is associated with a reduced peak expression of several glucose metabolic genes critical in the control of gluconeogenesis and glycolysis. Genome-wide cistromic profiling, promoter and mutation analysis support the concept that RORγ regulates the transcription of several glucose metabolic genes directly by binding ROREs in their promoter regulatory region. Similar observations were made in liver-specific RORγ-deficient mice suggesting that the changes in glucose homeostasis were directly related to the loss of hepatic RORγ expression. Altogether, our study shows that RORγ regulates several glucose metabolic genes downstream of the hepatic clock and identifies a novel metabolic function for RORγ in the diurnal regulation of hepatic gluconeogenesis and insulin sensitivity. The inhibition of the activation of several metabolic gene promoters by an RORγ antagonist suggests that antagonists may provide a novel strategy in the management of metabolic diseases, including type 2 diabetes.
| The circadian clock plays a critical role in the regulation of many physiological processes, including metabolism and energy homeostasis. The retinoic acid-related orphan receptor γ (RORγ) functions as a ligand-dependent transcription factor that regulates transcription by binding as a monomer to ROR-responsive elements. In liver, RORγ exhibits a robust circadian pattern of expression that is under direct control of the hepatic circadian clock. However, the connection between the circadian regulation of RORγ and its control of downstream metabolic processes is not well understood. In this study, by using ubiquitous and liver-specific RORγ-deficient mice as models, we demonstrate that hepatic RORγ modulates daily insulin sensitivity and glucose tolerance by regulating hepatic gluconeogenesis. Genome-wide cistromic profiling, gene expression, and promoter analysis revealed that RORγ is targeting and regulating a number of novel metabolic genes critical in the control of glycolysis and gluconeogenesis pathways. We provide evidence for a model in which RORγ regulates the circadian expression of glucose metabolic genes in the liver downstream of the hepatic circadian clock, thereby enhancing gluconeogenesis and decreasing insulin sensitivity and glucose tolerance. This study suggests that attenuating RORγ activity by antagonists might be beneficial for the management of glucose metabolic diseases including type 2 diabetes.
| RORγ constitutes with RORα and RORβ, the retinoic acid-related orphan receptor (ROR; NR1F1–3) subfamily of the nuclear receptors, which regulate transcription by binding as monomers to ROR-responsive elements (ROREs) in the regulatory region of target genes [1], [2]. Through alternative promoter usage, the RORγ gene generates 2 isoforms, RORγ1 and RORγ2 (RORγt), that regulate different physiological functions. RORγt is restricted to several distinct immune cells and is essential for thymopoiesis, lymph node development, and Th17 cell differentiation [1], [3]–[5]. RORγ antagonists inhibit Th17 cell differentiation and may provide a novel therapeutic strategy in the management of several autoimmune diseases [4], [6].
In contrast to RORγt, relatively little is known about the physiological functions of RORγ1. The expression of RORγ1 is highly restricted to tissues that have major functions in metabolism and energy homeostasis, including liver and adipose tissue, and in contrast to RORα and RORβ, RORγ is not expressed in the central nervous system, including the hypothalamus and suprachiasmatic nucleus [1], [6]–[13]. In several peripheral tissues RORγ1 exhibits a robust rhythmic pattern of expression with a peak at zeitgeber time (ZT) 16–20 that is directly regulated by the clock proteins, brain and muscle ARNT-like (Bmal1) and circadian locomotor output cycles kaput (Clock), and the Rev-Erb nuclear receptors [1], [8]–[12], [14], [15]. Although RORγ is recruited to ROREs in the regulatory regions of several clock genes, including Bmal1, Clock, Rev-Erbα, and cryptochrome 1 (Cry1); the loss of RORγ has little influence on the expression of Bmal1 and Clock, and only modestly reduces the expression of Rev-Erbα and Cry1 [10], [12]; The robust oscillatory regulation of RORγ1 expression by the clock machinery raised the possibility that RORγ might regulate the expression of certain target genes in a ZT-dependent manner. Because the clock machinery plays a critical role in the circadian regulation of many metabolic pathways, including glucose metabolism [13], [16]–[19], RORγ may function as an intermediary between the clock machinery and the regulation of metabolic genes. Since recent studies indicated an association between the level of RORγ expression and obesity-associated insulin resistance in mice and humans [20], [21], these observations led us to propose that RORγ1 might be an important participant in the diurnal regulation of glucose metabolic pathways [10], [16], [18], [22].
To study this hypothesis further, we examined the effect of the loss of RORγ on the diurnal regulation of glucose metabolism in ubiquitous and the hepatocyte-specific RORγ knockout mice. This analysis showed that loss of RORγ enhances glucose tolerance and insulin sensitivity particularly during early daytime (ZT4–6) and reduces the peak expression of several glucose metabolic genes. RORγ cistrome and promoter analysis indicated that several of these metabolic genes were regulated directly by RORγ and involved ZT-dependent recruitment of RORγ to ROREs in their regulatory region. Together, our observations are consistent with the concept that RORγ directly regulates the diurnal expression of a number of glucose metabolic genes in the liver downstream of the hepatic clock machinery, thereby enhancing gluconeogenesis and decreasing insulin sensitivity and glucose tolerance. The inhibition of the activation of several glucose metabolic gene promoters by an RORγ antagonist suggests that such antagonists might provide a novel therapeutic strategy in the management of insulin resistance and type 2 diabetes.
Glucose tolerance and insulin sensitivity, as RORγ1 expression, have been reported to be under endogenous circadian control [23], [24]. Recently, we proposed that RORγ1 might be an important participant in the diurnal regulation of several glucose metabolic pathways downstream of the circadian clock [10], [22]. To study the potential role of RORγ in glucose homeostasis, we examined the effect of the loss of RORγ on insulin sensitivity, glucose tolerance and the rhythmic expression pattern of glucose metabolic genes in ubiquitous and hepatocyte-specific RORγ knockout mice. Our data revealed that the loss of RORγ expression had a significant effect on insulin tolerance (ITT) and glucose tolerance (GTT) in mice fed with a high-fat diet (HFD). Comparison of the insulin responsiveness at two different time periods, ZT4–6 (daytime) and ZT18–20 (nighttime) showed that in wild type mice fed a HFD (WT(HFD)) insulin was more effective in controlling glucose levels at ZT18–20 than at ZT4–6 indicating that insulin sensitivity was ZT dependent [23], [24] (Figure 1A). Interestingly, this ZT-dependent difference in insulin responsiveness was greatly diminished in RORγ−/−(HFD) mice. ITT analysis showed that at ZT4–6 blood glucose levels remained significantly lower in RORγ−/−(HFD) mice after insulin injection than in WT(HFD) mice particularly after reaching a trough at 60 min (Figure 1A and Table S1). ITT performed at CT4–6 under constant darkness similarly showed improved insulin sensitivity in RORγ−/−(HFD) mice (Figure S1A), suggesting that RORγ significantly affects insulin sensitivity also under a Zeitgeber-free condition. At ZT18–20 the difference in ITT response between WT(HFD) and RORγ−/−(HFD) mice was significantly smaller than at ZT4–6. Consistent with the improved insulin sensitivity, GTT analysis showed that RORγ−/−(HFD) mice were more glucose tolerant than WT(HFD) particularly at ZT4–6 (Figure 1C). Although the difference was smaller than in mice fed with a HFD, RORγ−/−(ND) mice fed with a normal diet (ND) were also significantly more insulin sensitive and glucose tolerant at ZT4–6 than WT(ND) mice (Figure S1C and S1D). Because of the larger difference in mice fed a HFD, we focused much of our further analysis particularly on these mice. Altogether our observations indicate that the loss of RORγ enhanced glucose tolerance and insulin sensitivity particularly at ZT4–6 and CT4–6. Analysis of the areas under the curves (AUC) for ITT and GTT was consistent with this conclusion (Figure 1B and 1D).
To obtain further insights into the improved insulin sensitivity in RORγ−/− mice, we compared the level of insulin-induced activation of Akt phosphorylation (P-Akt), one of the most sensitive phosphorylation targets in the insulin signaling pathway, in liver and several other metabolic tissues (Figure 1E). No significant difference in P-Akt was observed at ZT4–6 in liver, brown and white adipose tissue (BAT, WAT), skeletal muscle between WT(HFD) and RORγ−/−(HFD) mice after insulin stimulation. Moreover, no significant difference in P-Akt was observed between insulin-treated WT and RORγ−/− primary hepatocytes (Figure 1F). These results suggest that loss of RORγ does not alter insulin-dependent phosphorylation of Akt in several metabolic tissues.
Next, we examined insulin sensitivity and glucose fluxes at daytime by the hyperinsulinemic-euglycemic clamp test. Consistent with the results of ITT, the glucose infusion rate (GIR) required to maintain blood glucose level under constant insulin infusion was significantly higher in RORγ−/−(HFD) mice than in WT(HFD) mice at daytime (ZT2–9), while their glucose absorption rate estimated by whole-body glucose disappearance (Rd) was almost equal during the clamp (Figure 2A, S2A, S2B). Importantly, basal hepatic glucose production (HGP) and clamp HGP were significantly lowered in RORγ−/− mice. Insulin equally suppressed the HGP about 70% in both WT and RORγ−/− mice (Figure 2B), indicating that the insulin responsiveness was not changed in RORγ−/− mice, consistent with the observation in Figures 1E and 1F. Glucose turnover estimated from the steady-state infusion of 3H-glucose (Basal HGP and Rd) [25] was lower in RORγ−/− mice, indicating that the glucose absorption rate might also be reduced. These results suggest that the increased GIR required to maintain blood glucose level in RORγ−/− mice was due to reduced hepatic glucose production and not due to improved insulin responsiveness.
The clamp test suggested that the output of hepatic glucose produced by gluconeogenesis and glycogenolysis was reduced in RORγ−/− mice. Because hepatic gluconeogenesis is under close control of the circadian clock [18], [23], [26], we analyzed gluconeogenesis efficiency at 2 different ZTs in WT and RORγ−/− mice fed with either a ND or HFD. The pyruvate tolerance test (PTT) indicated that gluconeogenesis was significantly higher at ZT4–6 than at ZT18–20 in both WT mice RORγ−/− mice with fed either a HFD or ND (Figure S1E). However, gluconeogenesis was greatly reduced at ZT4–6 in RORγ−/− mice compared to WT mice independent of whether the mice were fed a ND or HFD, while little difference in pyruvate tolerance was observed at ZT18–20 between the two genotypes (Figure 2C, S1E). Analysis of the AUC for PTT supported this conclusion (Figure 2D, S1E). RORγ−/−(HFD) mice also showed a reduced gluconeogenesis at CT4–6, a subjective daytime, under constant darkness (Figure S1B). Together, these observations indicate that loss of RORγ affects pyruvate tolerance particularly at ZT4–6 and support a regulatory role for RORγ in the circadian control of hepatic gluconeogenesis.
To obtain additional evidence that RORγ enhances hepatic gluconeogenesis, we analyzed PTT in RORγ−/− mice in which RORγ was over-expressed in liver by adenovirus administration. As shown in Figure 2E, gluconeogenesis was significantly increased in mice injected with RORγ-expressing adenovirus compared to mice injected with empty adenovirus. Further support for a role of RORγ in gluconeogenesis was provided by data showing that over-expression of RORγ in RORγ−/− primary hepatocytes increased glucose production (Figure S2C). Together these results suggested that RORγ modulates insulin resistance and glucose tolerance by regulating hepatic gluconeogenesis.
Food intake during daytime and nighttime was not significantly changed in RORγ−/−(HFD) mice (Figure 3A) and although glucose levels tended to be somewhat lower during daytime, a period in which gluconeogenesis was reduced, serum glucose levels were largely maintained in RORγ−/−(HFD) mice (Figure 3B). Serum insulin levels in WT mice exhibited a circadian pattern reaching peak levels at ZT16, while insulin levels were significantly lower in both RORγ−/−(HFD) and RORγ−/−(ND) mice particularly during ZT12–20 (Figure 3B, S3A). Glucose-stimulated insulin secretion (GSIS) experiments indicated no difference in insulin secretion between WT and RORγ−/− mice fed with either a ND or HFD (Figure 3C). In addition, little difference was observed in the level of pancreatic insulin at ZT16, the time at which the difference in serum insulin levels was the greatest (Figure 3D). These results suggested that lower serum insulin levels in RORγ−/− mice were not due to impaired insulin secretion or reduced pancreatic β-cell mass. Moreover, the amount of insulin secretion in response to the same quantity of glucose injected was not changed, suggesting that the reduced insulin level in RORγ−/− mice is likely due to reduced glucose production.
Glyconeogenesis and glycogenolysis play an important part in glucose homeostasis; 10–20% of hepatic glucose production in mice fasting for 4 h depends on glycogenolysis [27]. Hepatic glycogen reached its highest level at ZT0 and its lowest between ZT8–12 in both WT(HFD) and RORγ−/−(HFD) mice; however, its peak level was significantly lower in RORγ−/−(HFD) mice (Figure 3E). After 16 h fasting, the level of hepatic glycogen was dramatically reduced in both WT(HFD) and RORγ−/−(HFD) mice, but levels remained significantly lower in RORγ−/−(HFD) mice (Figure 3F). The level of hepatic glycogen was also reduced in RORγ−/− mice fed with a ND (Figure S3B). Glycogen accumulation was increased in RORγ−/−(HFD) mice injected with RORγ-expressing adenovirus (Figure 3G), indicating that RORγ positively contributes to hepatic glycogen accumulation. Altogether, these results indicate that RORγ−/− mice are able to maintain blood glucose levels at lower insulin levels due to reduced hepatic glucose production and possibly reduced glucose uptake by the liver. The latter is consistent with the reduced glycogen accumulation and clamp test data showing that basal HGP/Rd was reduced in RORγ−/− mice (Figure 2A).
We next examined the behavior activity and energy homeostasis in WT(ND) and RORγ−/−(ND) mice in relationship to the effect of RORγ on circadian rhythm and hepatic glucose metabolism. No significant difference in total body weight was observed between WT and RORγ−/− mice fed a ND (Figure S3C). The wheel running test showed that the circadian phase of behavioral activity was not changed in RORγ−/−(ND) mice consistent with a previous report [12], but peak activity was lower than in WT mice (Figure S3D). Indirect calorimetry showed that oxygen consumption (VO2), CO2 production (VCO2), respiratory exchange ratio (RER), and heat production were significantly lower in RORγ−/−(ND) mice compared to WT(ND) mice particularly at nighttime (Figure 3H and Figure S3E). Lower RER particularly at nighttime might indicate a preference for fatty acid consumption over glucose for energy production. Plotting of these parameters as a ratio between RORγ−/−(ND) and WT(ND) mice showed that the largest difference between WT and RORγ−/− mice occurred around ZT20 (Figure 3I), which corresponds closely to the peak expression of RORγ [10]. These results indicate that the change in glucose metabolism in RORγ−/− mice is associated with reduced energy expenditure.
To obtain further insights into the mechanism underlying the regulation of hepatic glucose metabolism by RORγ, we performed ChIP-Seq analysis to determine the genome-wide map of cis-acting targets (cistrome) of RORγ in murine liver at ZT22, a few hours after the peak expression of RORγ (Figure S4A) [10]. This analysis identified 3,061 RORγ binding sites (P<0.001) that were localized within intergenic regions (40.5%), introns (34.5%), within a 5 kb region upstream of the transcription start site (TSS)(11.5%), and the 5′UTR (10.8%) (Figure 4A, 4B). Notably, RORγ-binding sites were enriched near the transcription start sites (Figure 4C). De novo motif analysis using MEME program identified a classic RORE motif, AGGTCA preceded by an AT-rich region (Figure 4D and 4E) as well as direct repeat 1 (DR1)-like nuclear receptor binding motif and a RORE variant motif. Interestingly, a similar DR1 and variant RORE motifs were recently found within the binding sites of Rev-Erbs [14], [28]. Gene ontology analysis of 1,443 RORγ candidate target genes, defined as those that have one or more detected RORγ binding site within 5 Kb upstream of the TSS and/or within the gene body, indicated that the RORγ cistrome was enriched for genes involved in fatty acid, amino acid, and carbohydrate metabolism (Table 1 and Table S2). Comparison of the ChIP-Seq data with those obtained from our previous microarray analysis [29] indicated that about 23% of the RORγ candidate target genes were differentially expressed between WT and RORγ−/− liver. CircaDB (http://bioinf.itmat.upenn.edu/circa/) database analysis indicated that about 25% of the RORγ target genes exhibited a rhythmic expression pattern.
Because RORα and RORγ bind similar DNA response elements, we examined the degree of functional redundancy between RORγ and RORα in regulating hepatic gene expression by comparing the RORα and RORγ binding sites identified by ChIP-Seq analyses. The specificity of each anti-ROR antibody was confirmed by WB and ChIP assays using chromatin of ROR-deficient mice as a negative control (Figure S4B and S4C). ChIP-Seq analysis identified 1,319 RORα binding sites (P<0.001) and 957 candidate target genes (Figure 4F). Comparison of the RORα and RORγ cistromes revealed that 288 sites, including the ROREs within several clock genes reported previously [10], recruited both RORα and RORγ (Figure 4G and Table S3). Thus, the relatively small overlap indicates that in liver RORα and RORγ exhibit a limited functional redundancy.
Our ChIP-Seq analysis indicated that RORγ is recruited to regulatory regions of several genes implicated in hepatic glucose metabolism, including G6pase, Pepck, Glut2, Pklr, Gck, Gckr, Gys2, Pparδ, Pcx and Klf15 (Figure 4G, S5). Loss of RORγ resulted in a ZT-dependent decrease in the hepatic expression of most of these genes (Figure 5A–5D) and are consistent with our ChIP-Seq data indicating that their transcription is directly regulated by RORγ. The expression of G6pase was repressed in RORγ−/− liver during most of the circadian cycle, while Pepck expression was reduced during ZT4–12; both genes play a key role in gluconeogenesis (Figure 5A). Peak expression of Gys2, encoding a rate-limiting enzyme for glycogenesis, and Pparδ, which regulates several genes involved in glucose and lipid metabolism [30], was decreased between ZT4–16 and ZT16-4, respectively. The expression of several other gluconeogenic genes, including Pcx and Klf15, the glucose transporter Glut2, and several genes important in the glycolysis pathway, including Plkr, Gck, and Gckr, was also diminished in RORγ−/− liver (Figure 5A–5D). Decreased expression of these genes was also observed in liver of RORγ−/− mice fed with a HFD (Figure 5C). Importantly, the loss of RORγ had very little effect on the expression of Bmal1 and Clock, and a limited influence on the expression of Cry1 and Rev-Erbα [10], which all play a critical role in the circadian regulation of lipid/glucose metabolic genes (Figure S6) [10], [12]. These results are consistent with the conclusion that the changes in the circadian pattern of expression of glucose metabolic genes are directly related the loss of RORγ rather than changes in the regulation of clock genes by RORγ.
We further showed that exogenous expression of RORγ in RORγ−/− liver tissue by adenovirus significantly increased the expression of G6pase, Pepck, Gck, Gckr, Pparδ, Pcx, and Klf15 as well as the RORγ-target gene, Avpr1a (Figure 5E) [10]. Similarly, exogenous expression of RORγ in RORγ−/− primary hepatocytes significantly activated the expression of several of these genes (Figure 5F). These data are consistent with the conclusion that these genes are positively regulated by RORγ.
To examine whether any of these changes in gene expression translated into alterations in corresponding protein, we analyzed the expression of Pklr, which plays a key role in catalyzing the formation of pyruvate from phosphoenolpyruvate. As shown in Figure 5A and 5B, the level of Pklr protein in WT and RORγ−/− liver correlated rather well with the level of RNA expression. The levels of Pklr protein and RNA were higher at ZT16 than at ZT4 and clearly repressed in RORγ−/− liver.
Our ChIP-Seq analysis indicated that in liver both RORα and RORγ are recruited to the proximal promoter of G6pase and to intron 2 of Pparδ (Figure 4G and Figure S5A). ChIP-QPCR analysis showed higher association of RORγ with these regulatory regions at ZT22 compared to ZT10, whereas relatively little recruitment was observed in RORγ−/− liver at either ZT10 or ZT22 (Figure S5D, S5E). Analysis of the G6pase proximal promoter (−500/+58) identified, in addition to a classical RORE (RORE1) [31], a RORE variant motif (RORE2), and a PPAR responsive-element (PPRE) (Figure 6A), which has been reported to mediate the transactivation of G6pase by PPARα [32]. Reporter gene analysis showed that both RORγ and RORα were able to highly activate the G6pase promoter (Figure 6A), while the RORγ-selective antagonist “A” [10] inhibited the activation by RORγ at concentrations as low as 100 nM (Figure 6B). Mutation of either the RORE1 or RORE2 greatly reduced the activation by RORs. Interestingly, these RORE mutations also inhibited the transcriptional activation of the G6pase promoter by PPARα. Inversely, a PPRE mutation significantly reduced the activation by RORs as well as by PPARα, while mutation of both ROREs and PPRE almost totally abolished G6pase transactivation (Figure 6A). These observations suggested that RORs and PPARα collectively regulate G6pase expression.
The ROR binding region in intron 2 of Pparδ contains three putative ROREs, including a variant sequence (Figure 6C). Reporter analysis showed that RORγ and RORα activated the Luc reporter gene driven by this regulatory region about 45- and 140-fold, respectively. Mutation of any of these 3 ROREs strongly reduced the activation of the reporter by RORγ, while the triple mutation almost totally abolished activation. The RORγ antagonist inhibited this activation in a dose-responsive manner (Figure 6D). These results support the conclusion that Pparδ transcription is directly regulated by RORγ through these response elements and suggest that the circadian regulation of certain metabolic outputs by RORγ may be in part due to its regulation of Pparδ expression.
Although RORα was recruited to the RORE-containing regions of G6pase and Pparδ (Figure S5D, S5E) and activated the G6pase and the Pparδ regulatory region in reporter assays, loss of RORα had little effect on the circadian expression of G6pase and Pparδ (Figure 6E). The expression of these genes in double knockout RORαsg/sgRORγ−/− liver was reduced to a similar degree as in RORγ−/− liver (Figure 6F). These results suggest that under the conditions tested RORγ rather than RORα, plays a significant role in the hepatic regulation of G6pase and Pparδ in vivo.
In addition to G6pase and Pparδ, RORγ was recruited to several other genes important in glucose homeostasis, including intron 1 of Gck, the proximal promoter (−685/+42) of Gckr (Figure 6G and 6H, Figure S5B), intron 2 of Glut2, the promoter of Gys2, and the promoter of Dlat (Figure S7A). RORγ was able to activate the Luc reporter gene driven by these regulatory regions. Mutation or deletion of the RORE(s) in the Gck and Gckr regulatory region as well as addition of the RORγ antagonist significantly reduced the activation by RORγ (Figure 6G, 6H, S7B). ChIP-Seq analysis showed that RORα was not associated with these genes, and except for Gys2, RORα-deficiency had little effect on the expression of these genes in vivo (Figure S7C, S7D). Together, these results support the conclusion that RORγ directly regulates the transcription of a series of genes important in glucose metabolism and homeostasis.
To determine whether the effects on hepatic glucose metabolism were based on the hepatocyte-specific loss of RORγ function rather than loss of RORγ in other metabolic tissues or immune cells, we analyzed liver-specific RORγ-deficient (RORγfx/fxAlb-Cre+) mice. Our data confirmed that RORγ expression was completely lost in the liver of RORγfx/fxAlb-Cre+ mice and was not changed in the kidney (Figure 7A). ITT, GTT, and PTT analysis showed that, as demonstrated for the RORγ ubiquitous knockout mice, RORγfx/fxAlb-Cre+(HFD) mice exhibited a greater glucose tolerance, were more responsive to insulin, and showed reduced gluconeogenesis, respectively (Figure 7B–7D). Moreover, as in RORγ−/− mice, the blood insulin concentration at ZT16 was significantly reduced in RORγfx/fxAlb-Cre+(HFD) mice and so was the peak accumulation of hepatic glycogen at ZT0 (Figure 7E). Moreover, gene expression analysis showed that the hepatic expression of a series of RORγ target genes important in glucose metabolism, including G6pase and Pparδ, were also decreased in RORγfx/fxAlb-Cre+ mice as seen in RORγ−/− mice (Figure 7F). Together, these observations suggest that the changes in hepatic glucose metabolism are related directly to the loss of RORγ function in the liver and support the conclusion that RORγ directly contributes to the regulation of hepatic gluconeogenesis and glucose metabolism.
In this study, we identify a novel function for RORγ in the daily regulation of hepatic glucose metabolism and insulin sensitivity. Our results demonstrate that at ZT4–6 RORγ−/− mice are significantly more insulin sensitive and glucose tolerant than WT mice, while there was a smaller difference between the two strains at ZT18–20. The euglycemic clamp test revealed that hepatic glucose production was considerably reduced in RORγ−/− mice (Figure 2A). This was supported by PTT data showing that the conversion of exogenously administered pyruvate to glucose was significantly lower in RORγ−/− mice particularly at ZT4–6 (Figure 2C). Inversely, ectopic RORγ expression in RORγ−/− liver tissue or primary hepatocytes increased glucose production (Figure 2E, S2C). Our ITT and PTT data indicate that the regulation of glucose metabolism by RORγ is also functional at subjective daytime, CT4–6, under constant darkness (Figure S1A, S1B). Together, these observations demonstrate that gluconeogenesis is less efficient in RORγ−/− liver and support the conclusion that RORγ is an important positive regulator of hepatic gluconeogenesis and insulin sensitivity particularly during early daytime.
The regulation of glucose metabolism is complex and not only depends on hepatic metabolism, but also involves control of metabolic pathways in other tissues in which RORγ is expressed, such as adipose and skeletal muscle. It also involves certain regions of the brain, including the SCN and the hypothalamus, which are implicated in the regulation of the central circadian clock and appetite, respectively [16]–[18]. However, in contrast to RORα and RORβ, RORγ is not or very poorly expressed in the SCN, hypothalamus or other parts of the brain [11], [33]. Therefore, it appears unlikely that the brain plays a major role in the phenotypic changes observed in RORγ−/− mice. In addition, many of the changes in RORγ−/− mice, including the reduction in glucose metabolic gene expression, were also observed in liver-specific RORγ-deficient mice, indicating that these effects are directly related to the loss of RORγ in hepatocytes and separate from the loss of RORγ in other metabolic tissues (Figure 7F).
Since RORγ functions as a transcription factor, the reduced gluconeogenesis in RORγ-deficient mice must involve alterations in the transcription of RORγ target genes. De novo motif analysis of the RORγ cistrome identified, in addition to the classic RORE, two variant RORE-like motifs. The variant ROREs appear to allow a greater diversity in ROR binding than expected from the in vitro binding assays [34], [35]. A greater promiscuity in in vivo DNA binding has also been observed for other nuclear receptors, and might be due to promoter context, chromatin structure, and histone modifications. Gene ontology analysis showed that many of the potential RORγ-target genes are linked to metabolic pathways (Table 1 and Table S2), including glucose homeostasis (e.g., G6pase, Pepck, Pklr, Pparδ, Gck, Gckr, Glut2, Gys2, Dlat, Pcx, and Klf15). Analysis of their rhythmic pattern of expression demonstrated that RORγ deletion reduced peak expression of most of these genes, without affecting their phase. Regulation of these genes by RORγ was supported by data showing that exogenous expression of RORγ in RORγ−/− liver and primary hepatocytes significantly enhanced their level of expression (Figure 5E, 5F). Promoter and mutation analysis demonstrated that RORγ was able to activate several of the RORE-containing promoters, while mutation of either the classic or variant ROREs significantly reduced this activation by RORγ indicating that these motifs are functional. The RORγ-mediated promoter activation was further supported by data showing that treatment with a RORγ-selective antagonist considerably inhibited this activation (Figure 6B, 6D, S7B). Our RORγ cistrome data together with the mRNA expression and promoter analysis support the model that in murine liver, RORγ positively regulates the expression of a series of glucose metabolic genes directly through RORE binding. The reduced peak expression of several key metabolic genes, including G6pase and Pepck, which are critical for gluconeogenesis, the glucose transporter Glut2, and several genes important in the glycolysis pathway, including Plkr, Gck, and Gckr, likely contribute to the reduced glucose uptake, the less efficient gluconeogenesis and the lower glycogen accumulation observed in RORγ deficient liver.
In addition to RORγ, glucose metabolism is under the control of a number of other transcription factors. Although loss of RORγ reduced peak expression of several glucose metabolic genes, most of these genes still exhibited a substantial rhythmic pattern of expression, indicating that additional factors are involved. For example, analysis of the G6pase promoter showed that in addition to the classic and variant RORE proximal promoter, it contained a PPRE (Figure 6A), which has been reported to mediate the transactivation of G6pase by PPARα [32]. Mutation of either the ROREs or PPRE caused a significant reduction in the activation of this promoter suggesting that RORγ and PPARα cooperatively regulate G6pase. Although comparison of the RORα and RORγ cistromes indicated that RORα and RORγ have largely distinct functions, there was a 10% overlap in target genes that included several glucose metabolic genes, such as G6pase and Pparδ (Figure S5). However, in contrast to RORγ−/− mice, loss of RORα did not affect the expression of G6pase or Pparδ (Figure 6E, 6F) suggesting that under the conditions tested these genes are regulated by RORγ rather than RORα.
Although several studies have demonstrated a role for Bmal1 and Clock in the regulation of several metabolic genes and shown that RORγ is recruited to ROREs in Clock and Bmal1, the loss of RORγ had little effect on the hepatic expression of Bmal1 and Clock (Figure S6) [8], [10]. These observations suggest that changes in glucose metabolic genes in RORγ−/− liver are not due to changes in Clock or Bmal1 expression and are consistent with the hypothesis that RORγ regulates these genes downstream of the clock machinery. However, cistrome analysis has shown that Bmal1 can also be recruited to certain glucose metabolic genes, such as G6pase, suggesting that Bmal1 in conjunction with RORγ positively regulates the expression of these genes. In addition, RORγ might cause changes in chromatin structure and as such influences the recruitment of Bmal1 or Clock to common target genes. The Rev-Erb nuclear receptors also play a critical regulatory role in the robust oscillation of circadian expression of a number genes [14]. RORs and Rev-Erb receptors can interfere with each other's activity by competing for RORE binding [10]. Despite the modest reduction in peak expression of Rev-Erbα in RORγ−/− liver (Figure S6), which should result in increased target gene expression, the loss of RORγ may reduce the competition with Rev-Erbα for RORE binding and as a consequence increase the repression of gene transcription by Rev-Erbα. A more comprehensive comparison between the cistrome of RORs, clock proteins, and Rev-Erbs is needed to provide further insights into the crosstalk between these transcription factors.
Although insulin levels were significantly lower in RORγ−/− mice, blood glucose levels were largely maintained (Figure 3B). At daytime, hepatic glucose production is less efficient in the knockout mice and consistent with this, blood insulin level was significantly reduced at ZT4. We hypothesize that insulin sensitivity in RORγ−/− mice is also improved during nighttime due to reduced hepatic glucose production, which as a consequence would require less insulin to maintain blood glucose level and explain the lower level of blood insulin in RORγ−/− mice. This is supported by AUC analysis for ITT, which indicates that also at nighttime insulin sensitivity was significantly better in RORγ−/− mice (Figure 1B). When mice eat during nighttime, more insulin is required to maintain blood glucose levels and this may account for the greater difference in blood insulin level compared to the difference at daytime. The observation that the PTT indicated little changed in gluconeogenesis efficiency at nighttime may be related to the fact that the PPT determines the efficiency of the gluconeogenesis pathway by measuring the formation of glucose from pyruvate after exogenous pyruvate injection, which is not a total reflection of all the pathways involved in the regulation of hepatic gluconeogenesis in vivo because pyruvate for gluconeogenesis can be supplied by other metabolic pathways.
A lower RER is considered to indicate that fat is increasingly preferred as a fuel source, whereas a higher RER is indicative for an increased use of carbohydrates. Thus, the lower RER observed at daytime in both WT and knockout mice indicates a greater preference for fatty acid consumption over glucose compared to nighttime (Figure 3H), while the lower nighttime RER levels in RORγ−/− mice compared to WT mice indicate a greater preference for fatty acid consumption over glucose. The latter is likely related to reduced glucose production and reduced glucose uptake in RORγ knockout liver. Our data show that hepatic glycogen accumulation was reduced in RORγ knockout mice during ZT16-0 clearly indicating that loss of RORγ also affects glucose homeostasis at nighttime. This reduction in glycogen is likely due a reduced glucose uptake, which correlate with the lower levels of blood insulin in RORγ knockout mice (Figure 3B and 3E). Further analyses will be needed to precisely understand the precise interrelationships between various transcription factors, their diurnal regulation of various metabolic pathways and glucose and energy homeostasis.
In summary, our study identifies a novel function for RORγ in the regulation of gluconeogenesis and insulin resistance. Our data are consistent with the model in which RORγ directly regulates the expression of glucose metabolic genes in the liver downstream of the hepatic circadian clock, thereby enhancing gluconeogenesis, and decreasing insulin sensitivity and glucose tolerance (Figure 7G). The temporal organization of tissue metabolism is coordinated by reciprocal crosstalk between the core clock machinery and key metabolic enzymes and transcription factors. Our study indicates that RORγ is a novel important participant in this crosstalk. The improved insulin sensitivity and glucose tolerance observed in RORγ-deficient mice suggest that the loss of RORγ might be beneficial in controlling glucose homeostasis and in the management of metabolic diseases. This is supported by recent studies showing that in human patients the level of RORγ expression positively correlates with insulin resistance [20], [21]. The inhibition of the activation of several glucose metabolic gene promoters by an RORγ-selective antagonist, thereby mimicking the effects in RORγ−/− liver, suggests that such antagonists might provide a novel therapeutic strategy in the management of insulin resistance and type 2 diabetes.
Heterozygous C57BL/6 staggerer (RORα+/sg) were obtained from the Jackson Laboratories (Bar Harbor, ME). RORγ−/− and RORαsg/sgRORγ−/− double knockout (DKO) mice were described previously [10], [36]. Liver-specific RORγ knockout mice, referred to as RORγfx/fxAlb-Cre+, were generated by crossing B6(Cg)-Rorctm3Litt/J (RORγfx/fx) with B6.Cg-Tg(Alb-cre)21Mgn/J transgenic mice (Jackson Laboratories). Mice were supplied ad libitum with NIH-A31 formula (normal diet, ND) and water, and maintained at 23°C on a constant 12 h light∶12 h dark cycle. Two month-old male mice were fed with a high fat diet (40% kcal fat) (HFD: D12079B Research Diets Inc., New Brunswick, NJ) for 6 weeks. Littermate wild type (WT) mice were used as controls. All animal protocols followed the guidelines outlined by the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at the NIEHS.
After 16 h fasting, WT and RORγ−/− mice (n = 8–10) fed a ND or HFD for 6 weeks were injected intraperitoneally with glucose (2 g/kg), insulin (0.75 U/kg) (Eli Lilly, Indianapolis, IN) or sodium pyruvate (2 g/kg) (Sigma-Aldrich) at ZT4 or ZT18. The blood glucose was measured every 20 min for up to 140 min with glucose test strips (Nova Biomedical, Waltham, MA). These tests were performed in the same way using RORγfx/fxAlb-Cre+ and RORγfx/fxAlb-Cre− mice (n = 11) fed a HFD. ITT and PTT were also performed under red light at CT4 after WT(HFD) and RORγ−/−(HFD) mice (n = 12) were kept for 1 day under constant darkness. Total AUC (Area under the curve) was calculated by the trapezoid rule. Two-way ANOVA was performed using GraphPad PRISM software.
To evaluate insulin signaling, liver, BAT, WAT, and skeletal muscle were isolated from fasting WT(HFD) and mice RORγ−/−(HFD) mice 30 min after injection with either 0.75 U/kg insulin or PBS. Protein from these tissues was extracted with lysis buffer (25 mM Tris-HCl pH 7.6, 150 mM NaCl, 1% Nonidet P-40, 1% sodium deoxycholate, 0.1% SDS). In a separate experiment, primary hepatocytes isolated from WT and RORγ−/− mice were treated with 20 nM insulin in serum-free 199 medium (Sigma-Aldrich) for 10 min. Phosphorylated Akt (Ser473) and whole Akt proteins were detected by Western blot analysis with antibodies 7408 and 7102 from Cell Signaling Technology. Pklr and Gapdh were detected in liver lysates from WT and RORγ−/− mice (n = 3) at ZT4 and ZT16 by Western blot analysis with anti-Pklr (22456-1-AP, Proteintech Group Inc., Chicago, IL, USA) and anti-Gapdh (Cell Signaling Technology) antibodies.
WT and RORγ−/− mice (n = 5) fed a HFD for 6 weeks underwent surgery under anesthesia to attach catheters to the jugular vein and carotid artery. Mice were left at least 2 days to recover. After a 3.5 h fasting, the basal rates of glucose turnover were measured by continuous infusion of HPLC-purified D-[3-3H] glucose (0.05 µCi/min) (Perkin Elmer, Boston, MA) for 90 minutes following a bolus of 1 µCi. Blood samples (about 40 µl) were taken from the carotid artery catheter at 75 and 85 min after the infusion to determine the plasma [3-3H] glucose concentration. Subsequently the hyperinsulinemic euglycemic clamp test was performed for 120 min in conscious, restrained mice. Human insulin (HumulinR, Eli Lilly) was infused at a constant rate (30 mU/kg/min) through the end of the experiment following a bolus of 90 mU/kg/min for 3 min. Glucose was measured every 10 min in blood from tail vein with glucose test strips. The glucose concentration was maintained at 110–130 mg/dl by a variable rate of 20% glucose infusion under a continuous infusion of [3-3H] glucose (0.1 µCi/min). Blood samples (about 40 µl) were taken from the carotid artery catheter every 10 min during the last 40 min. [3H]-glucose was used to trace hepatic glucose production and glucose turnover. The experiment was performed during daytime at ZT2–9.
For the determination of the plasma 3H-glucose concentration, plasma samples were deproteinized with 0.3 N Ba(OH)2 and ZnSO4 and dried to remove 3H2O before the radioactivity was measured in a liquid scintillation counter. Basal hepatic glucose production (Basal HGP) was calculated as the ratio of the preclamp [3H]-glucose infusion rate (GIR) (dpm/min) to the specific activity of plasma glucose. Clamp whole-body glucose disappearance (Rd) was calculated as the ratio of the clamp [3-3H] GIR (dpm/min) to the specific activity of plasma glucose. Clamp glucose production (Clamp HGP) was determined by subtracting the average GIR in the last 40 min from the Rd.
Recombinant adenoviruses were generated using the AdEasy adenoviral system (Agilent Technologies, Palo Alto, CA). Full-length RORγ1 cDNA was inserted to pShuttle-IRES-hrGFP-1 vector, and co-transformed with pAdEasy-1 in BJ5183-AD-1 bacteria by electroporation. The recombinant adenovirus plasmid was then transfected in AD-293 cells. The amplified adenoviruses were purified and concentrated by cesium chloride density gradient centrifugation. The empty control and RORγ expressing adenoviruses were injected into the retro-orbital sinus of RORγ−/−(HFD) mice (n = 6–7). Pyruvate tolerance test was performed 4 days later and after an additional four days, liver was collected at ZT8 to analyze glycogen accumulation and gene expression.
Hepatocytes from 2 month-old WT and RORγ−/− mouse were isolated with a Hepatocyte Isolation System (Worthington Biochemical Corporation, New Jersey, USA) according to the manufacturer's instructions. Primary hepatocytes were cultured in collagen-coated dishes with Medium 199 supplemented with 100 nM dexamethasone, 1 nM insulin, 10 nM triiodothyronine, 5% fetal bovine serum, and penicillin/streptomycin. After 8–12 h, cells were infected with empty lentivirus pLVX-mCherry-N1 or RORγ1-expressing lentivirus. 24 h later cells were washed twice in PBS and then incubated in serum-free medium 199 in the presence or absence of 100 nM insulin or 100 nM glucagon (Sigma-Aldrich) for 6 h before RNA was isolated. Glucose production was measured with a glucose production buffer (glucose/phenol red-free DMEM (Sigma-Aldrich), 1 mM lactose, 2 mM sodium pyruvate) in RORγ−/− hepatocyte infected with lentivirus for each empty and RORγ expression (n = 3). Glucose in the medium was measured with a Glucose assay kit (Sigma-Aldrich).
Serum and liver samples were collected from WT and RORγ−/− mice on a HFD (n≥5) every 4 h over a period of 24 h. Serum insulin was measured by a sandwich ELISA with a Rat/Mouse Insulin ELISA kit (EZRMI-13K, Millipore). Glucose stimulated insulin secretion (GSIS) was measured at ZT4 in WT and RORγ−/− mice on a HFD (n = 5–6) or ND (n = 2–3). Serum was collected at 2.5, 5, 15, and 30 min after intraperitoneal injection of glucose (2 g/kg). Pancreatic insulin was determined by rapidly removing the pancreas from WT and RORγ−/− mice (n = 10–14) on a HFD. Pancreas was then homogenized and extracted overnight with acid-ethanol at −20°C. Insulin in the extracts was measured with the insulin ELISA kit. Insulin was normalized by total pancreatic protein. Glycogen extracted from liver with 30% KOH at 100°C for 2 h followed by precipitation by ethanol, was measured with a Glycogen Assay Kit (BioVison Inc., Mountain View, CA).
To analyze metabolic parameters including oxygen consumption, CO2 production, respiratory exchange ratio, heat production, and food/water consumptions were measured in WT and RORγ−/− mice (n = 8) with a LabMaster system (TSE systems Inc., Chesterfield, MO) during 4 successive days.
The ChIP assay was performed using a ChIP assay kit from Millipore (Billerica, MA) according to the manufacturer's protocol with minor modifications as described previously [10]. Briefly, livers collected from WT, RORαsg/sg, and RORγ−/− mice at ZT10 and ZT22 were homogenized with a polytron PT 3000 (Brinkmann Instruments) and crosslinked by 1% formaldehyde for 10 min at room temperature. After a wash in PBS, an aliquot of the crosslinked chromatin was sonicated and incubated overnight with an anti-RORα or anti-RORγ antibody [10] generated against amino acids 129–231 and 121–213 in mouse RORγ1 and RORα4, respectively. After incubation with protein G agarose beads for 2 h, DNA-protein complexes were eluted. The crosslinks were reversed by overnight incubation at 65°C in the presence of 25 mM NaCl, digested with RNase A and proteinase K, and then the ChIPed-DNA was purified. The amount of ChIPed-DNA relative to each input DNA was determined by QPCR. All QPCR reactions were carried out in triplicate. Sequences of primers for ChIP-QPCR are listed in Table S4.
ChIPed-DNA and input DNA as a control were prepared using RORγ- and RORα-specific antibodies as described previously [10]. ChIP-Seq analysis was performed by the NIH Intramural Sequencing Center and data were analyzed as reported previously [37]. The sequencing reads were obtained from base-calling of Illumina Genome Analyzer. The wiggle-formatted alignment results were visualized on UCSC Genome Browser using mouse mm9 reference genome. SISSRs (Site Identification from Short Sequence Reads) were used for identification of significant RORγ and RORα binding sites (P<0.001) that have enriched reads in each ChIPed-DNA versus input control across the whole genome [38]. The distance from each ROR peak to the nearest transcriptional start sites was determined using custom scripts. De novo consensus motif search within ROR binding sites was performed using MEME. ChIP-Seq data was compared with gene expression data using Kolmogorov-Smirnov (KS) plot. Gene ontology analysis was performed using the NIH Database for Annotation, Visualization, and Integrated Discovery (DAVID) online web-server, and based on PANTHER Biological process definitions.
To quantify gene expression during circadian time, liver tissues were collected from WT, RORγ−/−, and RORαsg/sg mice every 4 h over a period of 24 h, processed overnight in RNAlater solution (Ambion, Austin, TX) at 4°C, and then stored at −80°C until use. Tissues were then homogenized with a Polytron PT-3000 (Brinkmann Instruments, Westbury, NY). Liver tissues were also collected from RORαsg/sgRORγ−/− DKO mice and littermate control WT mice, and RORγfx/fxAlb-Cre+ and RORγfx/fxAlb-Cre− mice at zeitgeber time (ZT) 8 and ZT20. RNA was then extracted using a QIAshredder column and RNeasy Mini kit (Qiagen, Valencia, CA) according to the manufacturer's instructions. The RNA was reverse-transcribed using a High-Capacity cDNA Archive Kit (Applied Biosystems). QPCR analysis was performed using SYBR Green I (Applied Biosystems, Foster City, CA). The reactions were carried out in triplicate using 20 ng of cDNA and the following conditions: 10 min at 95°C, followed by 40 cycles of 15 sec at 95°C and 60 sec at 60°C. The results were normalized by the amount of Gapdh mRNA. Primer sequences are listed in Table S4.
The promoter or intron region of mouse G6Pase (promoter; −500/+58), Pparδ (intron 2; +46417/+46987), Gck (intron 1; +29709/+30121), Gckr (promoter; −685/+42), Glut2 (intron 2; +16294/+16805), Gys2 (promoter; −256/+345), and Dlat (promoter; −1151/+22) genes was amplified using mouse genomic DNA (Promega, Madison, WI) and cloned into either the promoter-less reporter plasmid pGL4.10 or pGL4.27 containing a minimal promoter (Promega, Madison, WI). Point mutations in ROREs and PPREs were generated using a Quickchange Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA). Human hepatoma Huh-7 cells were co-transfected with the indicated pGL4 reporter plasmid, pCMV-β-Gal, and p3xFlag-CMV10-RORγ, –RORα, -Rev-Erbα, or -PPARα expression plasmids using lipofectamine 2000 (Invitrogen, Carlsbad, CA). After 24 h incubation, the luciferase and β-galactosidase activities were measured with a Luciferase Assay Substrate kit (Promega) and Luminescent β-galactosidase Detection Kit II (Clontech). All transfections were performed in triplicate and repeated at least twice. In certain experiments cells were treated for 24 h with a RORγ-selective antagonist “A”, (R)-N-(1-((4-methoxy-phenyl)sulfonyl)-4-methyl-1,2,3,4-tetrahydroquinolin-7-yl)-2,4,6-trimethylbenzene-sulfonamide provided by Dr. Veronique Birault (GlaxoSmithKline) [10] or with the selective PPARα antagonist, Wy14,643 (10 µM; Sigma-Aldrich) as indicated.
|
10.1371/journal.pgen.1003818 | Multiple Signaling Pathways Coordinate to Induce a Threshold Response in a Chordate Embryo | In animal development, secreted signaling molecules evoke all-or-none threshold responses of target gene transcription to specify cell fates. In the chordate Ciona intestinalis, the neural markers Otx and Nodal are induced at early embryonic stages by Fgf9/16/20 signaling. Here we show that three additional signaling molecules act negatively to generate a sharp expression boundary for neural genes. EphrinA signaling antagonizes FGF signaling by inhibiting ERK phosphorylation more strongly in epidermal cells than in neural cells, which accentuates differences in the strength of ERK activation. However, even weakly activated ERK activates Otx and Nodal transcription occasionally, probably because of the inherently stochastic nature of signal transduction processes and binding of transcription factors to target sequences. This occasional and undesirable activation of neural genes by weak residual ERK activity is directly repressed by Smad transcription factors activated by Admp and Gdf1/3-like signaling, further sharpening the differential responses of cells to FGF signaling. Thus, these signaling pathways coordinate to evoke a threshold response that delineates a sharp expression boundary.
| Graded signals often provide positional information to organize gene expression in animal embryos. In the simplest cases, graded signals are translated into all-or-none threshold responses. However, recent studies have shown that signal transduction processes and binding of transcription factors to target sequences are inherently stochastic. This means that even weak activating signaling might activate target genes stochastically. However, the precise mechanism, by which this stochastic undesirable activation is avoided, is still largely unknown. In the embryo of a simple chordate, Ciona intestinalis, FGF signaling is known to induce neural fate. In the present study, we demonstrate that three additional signaling molecules cooperate to evoke a threshold response for specification of neural fate. First, EphrinA signaling inhibits FGF signaling by attenuating ERK phosphorylation, accentuating differences in the strength of ERK activation. However, even weak ERK activity occasionally turns on the neural genes. This occasional undesirable activation of the neural genes is turned off by Admp and Gdf1/3 signaling through Smad transcription factors. Thus, these two qualitatively different negative regulatory mechanisms evoke an all-or-none threshold response to specify neural fate.
| In animal development, secreted signaling molecules often elicit the production of multiple cellular identities by controlling the activity of transcription factors. Molecular gradients can produce differential responses in identical cells [1], [2]. For example, in Drosophila syncytium embryos, a concentration gradient of the transcription factor Bicoid specifies the anterior-posterior axis [3], [4]. In the vertebrate neural tube, a gradient of the secreted signaling molecule Sonic Hedgehog is responsible for defining five distinct neural progenitor domains [5]–[7]. Translation of a graded distribution of molecules into sharp gene expression boundaries is central to many developmental processes, but apart from a few cases, the molecular mechanisms underlying this process are not yet fully understood. Especially, even a weak signal can potentially activate transcription of target genes due to the inherently stochastic nature of signal transduction processes and binding of transcription factors to target sequences [8]. How is such weak undesirable activation blocked in animal embryos?
Cells in the animal hemisphere of ascidian embryos (Ciona intestinalis) give rise to both epidermal and neural cells (Figure 1). At the 32-cell stage, an earliest neural marker gene, Otx, begins to be expressed in a pair of anterior animal cells (a6.5) and a pair of posterior animal cells (b6.5), and Nodal expression also begins in b6.5 (Figure S1A and S1B) [9]–[11]. Some embryos also express Otx in a6.7 [12], indicating that Otx expression in a6.7 is not tightly regulated. In the present study, we disregarded this cryptic expression unless otherwise noted. The remaining animal cells are all restricted to epidermal fate. In addition, Otx and Nodal are expressed in vegetal cells (Figure S1A and S1B). Otx is required for subsequent expression of neural genes [13], and ectopic Nodal expression in non-neural ectodermal cells results in embryonic patterning defects [14].
At the 16-cell stage, all ectodermal cells express the same set of regulatory genes, except for FoxA-a, which is expressed in anterior but not posterior cells [10] (Figure S1C). Even though FoxA-a activates the anterior fate, some other instructive mechanism likely functions to induce neural fate. However, no asymmetric localization of maternal mRNA has been detected in the animal hemisphere in spite of extensive efforts to identify such a molecule. In addition, a cell dissociation experiment indicated that cell-cell interactions are required for specification of the neural fate [9]. Therefore, it is likely that neural fate is specified primarily by cell-cell interactions.
It is possible that maternally provided signaling molecules and mRNAs encoding signaling molecules play a role in the specification of neural fate, even if they are distributed evenly within the embryo. However, it is more likely that signaling molecules expressed from the zygotic genome of specific cells play a more important role. Our previous study [10] showed that only five signaling ligand genes are zygotically expressed at the 16-cell stage, one stage earlier than the 32-cell stage when Otx and Nodal expression begins (Figure S1D–H). Fgf9/16/20 is expressed in all of the vegetal cells except for the most posterior ones [15], [16], EphrinA-d is expressed in the entire animal hemisphere, Wnt-NAe (a Wnt ligand gene whose phylogenetic position is unclear) and Admp are expressed in posterior vegetal cells (B5.1), and Gdf1/3-like (or orphan Tgfβ-1), is expressed in the entire animal hemisphere.
Among the ectodermal cells of the 32-cell embryo, cells with neural fate have a larger area of surface contact with FGF-expressing vegetal cells and are accordingly expected to be exposed to stronger FGF signaling [12]. This results in activation of maternal GATA and Ets transcription factors, which in turn directly activate Otx expression [16]. Nodal is similarly activated [17], but only in b6.5. However, because non-neural ectodermal cells also contact vegetal cells expressing Fgf9/16/20, it is very likely that these cells are exposed to weak FGF signaling. Due to the inherently stochastic nature [8], even weak FGF signaling might activate Otx and Nodal enhancers. In the present study, we show that weak FGF signaling indeed activates Otx and Nodal expression, and that EphrinA signaling amplifies the difference in ERK phosphorylation levels induced by differing strength of FGF signaling. Moreover, the occasional activation of Otx and Nodal by residual weak ERK activity is repressed by Admp/Gdf1/3-like signaling. Thus, FGF, Ephrin, and Admp/Gdf1/3-like signaling cooperate to evoke a threshold response to establish neural fate.
Our previous comprehensive screen [10] showed that FoxA-a is the only regulatory gene that are expressed differently between the a- and b-line cells. FoxA-a directly activates ectodermal genes in anterior cells at later stages [18] and represses Nodal at the early gastrula stage [13]. Therefore, we first examined whether FoxA-a similarly represses Nodal at the 32-cell stage. In embryos injected with an antisense morpholino oligonucleotide (MO) for FoxA-a, Nodal expression was indeed expressed ectopically in a6.5 at the 32-cell stage (Figure 2), indicating that FoxA-a normally suppresses Nodal expression in anterior cells.
As we described in the Introduction section, neural fate is probably specified primarily by cell-cell interactions. To understand the mechanisms that activate Otx and Nodal specifically in the neural lineage, we examined the functions of the five signaling ligand genes that are expressed at the 16-cell stage.
We first confirmed the effect of FGF signaling on neural marker expression. As previously shown [9], [16], [17], [19], responsiveness to FGF signaling, as indicated by activated ERK (dpERK), was observed in a6.5 and b6.5 in normal 32-cell embryos (Figure 3A), and expression of Otx and Nodal was absent from the animal hemisphere in Fgf9/16/20 morphants (Figure 3B–D; Tables S1 and S2). On the other hand, overexpression of Fgf9/16/20 by synthetic RNA microinjection into fertilized eggs and one posterior animal cell of 8-cell embryos resulted in ectopic expression of Otx (Figure S2A and S2B) [16]. Thus, FGF signaling activates Otx and Nodal expression via ERK activation.
As previously shown in later stage embryos [20]–[22] and in vertebrates [23], EphrinA-d attenuated ERK phosphorylation in 32-cell embryos, as indicated by the fact that dpERK immunostaining was observed ectopically in all of the animal blastomeres of EphrinA-d morphants (Figure 3E). Otx was expressed ectopically in animal cells in EphrinA-d morphants, and Nodal was expressed ectopically in posterior animal cells in these morphants (Figure 3F and 3G; Tables S1 and S2). Conversely, overexpression of EphrinA-d resulted in complete loss of Otx expression (Figure 3H). Thus, all of the animal cells indeed receive FGF signaling, and EphrinA-d appears to modulate FGF signaling by inhibiting ERK phosphorylation, generating clear differences in the strength of ERK activation.
In a previous study [12], “3D-virtual embryos” were reconstructed and the surface contacts of cells with their surrounding cells were calculated. This work showed that a6.5 and b6.5 have the greatest surface contacts with Fgf9/16/20-expressing cells and suggested that differences in the contact area of competent cells are important for Otx expression in neural cells [12]. Our calculation using this tool indicated that a6.5 and b6.5 have the least surface contact with EphrinA-d-expressing cells (Figure S3). Therefore, a6.5 and b6.5 are likely subject to the lowest levels of inhibitory signals repressing ERK activation, if cell contact areas represent the degree of EphrinA-d signaling as they do in the case of FGF9/16/20 signaling. Thus, inductive FGF signaling and inhibitory EphrinA signaling likely accentuate differences in the strength of ERK activation in animal cells.
In Wnt-NAe morphants, Otx and Nodal were expressed in both of the b5.3 daughter cells (b6.5 and b6.6) (Figure 3I and 3J), whereas overexpression of Wnt-NAe did not affect Otx expression (Figure 3K). This ectopic expression was likely due to the abnormal position of the b6.5 and b6.6 sister cells. In normal embryos, the b6.5 and b6.6 cells were both found in the periphery of the animal hemisphere (Figure 3L), while in the morphants one of them was located at a more interior position (Figure 3M). The boundary between these sister cells is significantly more oblique in the morphants. Because the positions of the rest of the blastomeres of the morphant embryos did not appear to be altered, we could identify these two cells as the daughter cells of b5.3. The mispositioning of the daughter cells of b5.3 likely changed the balance between FGF and EphrinA signaling, because dpERK signal was detected in both of the daughter cells of b5.3 in Wnt-NAe morphants (Figure 3N). Thus, this Wnt signaling was not directly involved in transcriptional regulation of Otx and Nodal.
In Admp or Gdf1/3-like morphants, the expression of Otx and Nodal was normal (Figure S4; Tables S1 and S2). Since these two molecules are both members of the TGFβ superfamily and might therefore work together, we knocked down these two genes simultaneously. In Admp and Gdf1/3-like double morphants (Admp/Gdf morphants hereafter), Otx and Nodal were ectopically expressed (Figure 4A and 4B; Tables S1 and S2), suggesting redundancy between Admp and Gdf1/3-like.
Admp signaling is transmitted through the BMP pathway, while GDF1 and GDF3 act through the Activin pathway [24]. A pharmacological inhibitor of BMP signaling, dorsomorphin, resulted in ectopic expression of Otx and Nodal (Figure 4C and 4D), but an inhibitor of Activin signaling, SB431542, did not (n = 70, 99% for Otx; n = 77, 99% for Nodal). Knockdown of Smad1/5, which encodes an effector transcription factor of the BMP pathway, resulted in ectopic expression of Otx and Nodal (Figure 4E and 4F). Knockdown of Smad2/3b, which encodes an effector of the Activin pathway, also resulted in ectopic expression of Otx and Nodal (Figure 4G and 4H), although the effect was weaker. These data indicate that the BMP and Activin pathways suppress Otx and Nodal expression, although the BMP signaling may contribute to this suppression more than Activin signaling.
The ectopic expression of Otx and Nodal in Admp/Gdf morphants was not due to elevated FGF/ERK signaling, as indicated by the facts that expression of Fgf9/16/20 and EphrinA-d was unaffected at the 16-cell stage (Figure 5A and 5B), and that no ectopic ERK activation was observed in Admp/Gdf morphants at the 32-cell stage (Figure 5C).
Nevertheless, FGF signaling was required for the ectopic expression of Otx and Nodal in Admp/Gdf morphants, because Otx and Nodal were not expressed in triple Fgf9/16/20/Admp/Gdf morphants (Figure 6A and 6B; Tables S1 and S2), Admp/Gdf morphants treated with an MEK inhibitor U0126 (Figure 6C and 6D), or triple Ets1/2/Admp/Gdf morphants (Figure 6E and 6F). These data suggest that even weak ERK activation that cannot be detected experimentally activates Otx and Nodal expression in non-neural ectodermal cells, if Admp/Gdf signaling is absent. However, this suppressing activity of Admp/Gdf signaling is limited and the distributions of these signaling molecules are probably unimportant, because overexpression of Admp and/or Gdf1/3-like rarely suppresses the endogenous expression of Otx (Figure S5).
As shown in Table S1, the ectopic expression of Otx was observed more frequently in a6.6 than in a6.8, and ectopic expression in a6.6 was observed in all embryos that ectopically expressed Otx in a6.8. In addition, expression in a6.7 was also observed in all embryos that ectopically expressed Otx in a6.6 and a6.8. Similarly, ectopic expression of Otx and Nodal in b6.7 was observed in all embryos that expressed these genes in b6.8 (Tables S1 and S2). The expression in b6.6 was observed in all embryos that expressed them in b6.7 and b6.8. These hierarchies within the a- and b-lines (a6.5<a6.7<a6.6<a6.8, b6.5<b6.6<b6.7<b6.8) closely accord with the order of the estimated strength of the EphrinA-d activity (a6.5<a6.7<a6.6<a6.8, b6.5<b6.7<b6.6<b6.8; Figure S3). The only exception is b6.6 and b6.7, and notably the contact area with FGF-expressing vegetal cells is estimated to be larger in b6.6 than in b6.7 [12]. Therefore, the above observation supports the estimation of EphrinA-d signaling strength by the 3D-virtual embryos.
Our data suggested that Admp/Gdf morphants are more sensitive to FGF signaling than normal embryos. Indeed, we found that Fgf9/16/20/Admp/Gdf morphants responded more sensitively to human bFGF added to the sea water than Fgf9/16/20 morphants; namely, Fgf9/16/20/Admp/Gdf morphants expressed Otx more frequently with increasing concentrations of bFGF (Figure 7A). On the other hand, there was no significant difference in the proportion of cells stained with the dpERK antibody (Figure 7B). At an intermediate concentration (5 ng/mL), 76% of the animal cells in Fgf9/16/20/Admp/Gdf morphants and 37% in Fgf9/16/20 morphants expressed Otx (Figure 7A), while dpERK signal was detected in 31% and 38% of cells in these morphants (Figure 7B). Thus, weak FGF signaling that is experimentally undetectable by dpERK immunostaining can activate Otx expression, and this weak signaling is inhibited by Admp/Gdf signaling. At the same time, the dose-dependent response of Otx activation indicates that differential FGF/ERK signaling strength alone cannot explain the threshold response.
Previous studies [16], [25] showed that an upstream region (a-element) of Otx is responsible for Otx expression in a6.5 blastomeres at the 32-cell stage. GATA and Ets transcription factors activated by the ERK signaling pathway bind to the a-element [16] (Figure S6A). Thus, we used a previously characterized reporter construct, in which the a-element and the minimal promoter region of the Brachyury gene were fused to the LacZ coding sequence [16] (Otx[a]>LacZ). This reporter construct was electroporated into fertilized eggs, and expression of LacZ mRNA was examined at the 32-cell stage. In addition to strong signal in a6.5 and b6.5 [16], we found weak signals in non-neural ectodermal cells in 10% of the experimental embryos (Figure 8A and 8D).
By examining the genomic sequence around the a-element of Otx, we identified two putative Smad-binding elements and one binding element for Smad4, a co-factor of regulatory Smad proteins [26], [27], within the 100-bp upstream region of the a-element (Figure S6A). When the region containing these Smad-binding elements (collectively called SBEs) was placed upstream of the a-element (Otx[SBE-a]>LacZ), LacZ was expressed specifically in the neural lineage, although the number of embryos expressing the reporter was reduced (Figure 8B and 8D). Treatment with dorsomorphin again induced ectopic expression of LacZ and enhanced overall expression, indicating that the SBEs work downstream of BMP signaling to weaken the activity of the enhancer (Figure 8C and 8D).
A Nodal cis-regulatory element responsible for expression in the neural lineage of cells (Nodal-a-element) was also identified previously [17] (Figure S6B).The Nodal-a-element induced the reporter gene expression in the anterior and posterior animal cells (Nodal[a]>LacZ), probably because it lacks FoxA-a binding sites. As in the case of Otx, this Nodal-a-element also induced non-neural expression (Figure S7A). We found one regulatory Smad binding site and one Smad4 binding site downstream of this enhancer. These SBEs suppressed LacZ reporter expression, when connected to the Nodal-a-element, and this suppression was abolished by dorsomorphin treatment (Figure S7B–D). Thus, Admp/Gdf signaling directly suppresses the activity of Otx and Nodal enhancers to evoke a robust threshold reaction.
Previous studies showed that differential FGF signaling from vegetal cells to animal cells plays a primary role in specific expression of Otx and Nodal [9], [12], [16], [17], [19]. However, it was unclear why non-neural ectodermal cells, which still receive FGF signals but at lower levels, fail to activate Otx and Nodal at all. Here, we showed that EphrinA-d, which antagonizes FGF signaling [20]–[22], amplifies the difference in ERK activity between ectodermal cells, as shown by dpERK immunostaining. Even below the detection limit, weak ERK activation occasionally activates Otx and Nodal expression, probably due to the inherently stochastic nature of signaling pathways and transcriptional activation [8]. The activity of Otx and Nodal transcriptional enhancers is weakened by Admp/Gdf signaling through the SBEs within the enhancers. The silencing activity of the SBEs is relatively weak and never overcomes fully activated enhancing activity. Thus, cooperation of multiple signaling pathways evokes a robust threshold reaction.
However, this cooperation cannot perfectly evoke a threshold response, because some embryos express Otx in a6.7 (6% in a previous assay [12] and 35% in our assay). As previously shown [12], FGF signaling is expected to be stronger in a6.7 than in a6.6 and a6.8. EphrinA-d signaling is expected to be stronger in a6.7 than in a6.5, and weaker in a6.7 than in a6.6 and a6.8, if cell contact areas with EphrinA-d-expressing cells simply represent the degree of EphrinA-d signaling. It is very likely that the sum of the positive and negative signaling activities in a6.7 is near the threshold, and consequently a6.7 occasionally activates Otx.
Admp is expressed in the posterior vegetal cells, and Gdf1/3-like is expressed in all of the cells in the animal hemisphere. Although these two factors are probably two major factors activating the BMP and Activin pathways, several members of the TGFβ-superfamily, including BMP2/4 and BMP3 are also expressed maternally [10]. In addition, TGFβ-superfamily molecules must be processed to become functional. Therefore, it is difficult to measure how these signaling molecules are distributed. However, because Admp and Gdf1/3-like cannot repress endogenous Otx and Nodal expression when overexpressed, the distributions of these two signaling molecules are probably unimportant for controlling Otx and Nodal expression.
Intriguingly, we found that knockdown of either of Smad1/5 or Smad2/3b causes ectopic expression of Otx and Nodal, while knockdown of either Admp or Gdf1/3-like does not produce an obvious phenotype. There are several possible explanations for this observation. Admp and Gdf1/3-like might not be fully knocked-down by the MOs we used, or other maternally expressed TGFβ-superfamily members might function redundantly. Additionally, there might be crosstalk between the BMP-signaling and Activin signaling pathways [28]. Nonetheless, the role of BMP/Activin-signaling we demonstrated in the regulation of Otx and Nodal expression is clear.
Transcriptional repressors play an important role in delineating sharp boundaries of gene expression [29]. For example, in Drosophila embryos, repressors that antagonize Bicoid activity are responsible for converting gradients into threshold responses [4]. Although reverse gradients can make a steep gradient, transcriptional repressors are often also required to repress residual activities, as in the case of neural cells of the Ciona embryo. A similar mechanism might function in a variety of developmental processes in which multiple signaling pathways are involved.
Similar to neural fate specification in the Ciona embryo, neural induction in Xenopus embryos involves BMP and FGF signaling. According to the most widely accepted “default model”, BMP inhibition is both necessary and sufficient for neural induction of vertebrate embryos [30], while FGF has an instructive role [31]–[34]. In addition, FGF signaling inhibits BMP signaling by phosphorylating Smad1, leading to the degradation of Smad1 [35], [36]. Inhibition of BMP signaling also induces FGF expression [33]. These mechanisms do not seem to be the principal mechanism of neural induction in Ciona embryos. However, it would be interesting to investigate whether the mechanism we describe here in Ciona embryos also functions in the vertebrate organizers.
Although it is not involved in evoking a threshold reaction, Wnt signaling is required for the proper spatial expression of Otx and Nodal. Our finding that Admp, Gdf1/3-like, and Wnt signaling regulate Otx and Nodal expression in the neural lineage is based on an unbiased and systematic analysis of signaling molecule genes expressed at the 16-cell stage in Ciona embryos. Because Admp/Gdf signaling and Wnt signaling do not play an instructive role in Otx and Nodal expression in the neural lineage, their involvement might have been difficult to uncover apart from such a comprehensive and unbiased approach.
C. intestinalis adults were obtained from the National Bio-Resource Project for Ciona. cDNA clones were obtained from our EST clone collection [37]. Inhibitors of BMP signaling (dorsomorphin; Wako), Activin signaling (SB431542, Sigma), and MEK signaling (U0126, Promega) were used at concentrations of 100 µM, 5 µM, and 10 µM, respectively. To examine responses to FGF, we used human recombinant bFGF (Sigma). SB431542 and U0126 were shown to work properly in the Ciona embryo in previous studies [11], [19]. As shown in Figure S8A, dorsomorphin treatment inhibits phosphorylation of Smad1/5; Western blotting with polyclonal antibodies against phosphorylated Smad5 (Abcam, ab92698) showed that treatment with human BMP4 (100 ng/mL; humanzyme) evoked hyper-phosphorylation of Smad1/5 in the 32-cell embryo and dorsomorphin (50 µM) inhibited this phosphorylation. After stripping the membrane, we performed Western blotting with antibodies for β-tubulin for a loading control (Sigma, T5293).
The morpholino oligonucleotides (MOs) (Gene Tools, LLC) for FoxA-a, Admp, Gdf1/3-like, Fgf9/16/20, Wnt-NAe, EphrinA-d and Ets1/2 were the same ones that we used in a previous study [13]. We designed an additional MO for Wnt-NAe (5′-TGTAAATGAAGACAACAGTTTAGAG-3′), which produced the same phenotype (ectopic Otx expression) as the original one, so only the results obtained with the second MO are shown. We also designed MOs for Smad1/5 (5′-AACAACTTCTCCACACAACAACCTG-3′) and Smad2/3b (5′-CATATTTACTCTCAATGTTCGATGT-3′) in the present study. All of these MOs were designed for blocking translation. The specificity of the Smad1/5 MO was confirmed by Western blotting. As described above, in embryos treated with human BMP4, phosphorylated Smad1/5 was detectable. When embryos injected with the Smad1/5 MO were treated with human BMP4, phosphorylated Smad1/5 was rarely detected (Figure S8B). The specificity of the Smad2/3b MO was confirmed by a rescue experiment: when we injected the Smad2/3b MO with a synthetic Smad2/3b mRNA that the MO cannot bind, ectopic expression of Otx, which is a phenotype of Smad2/3b morphants, was not observed (Figure S8C).
Synthetic overexpression transcripts were prepared from cDNA cloned into pBluescript RN3 vector [38] by in vitro transcription using a commercially available kit (mMESSAGE mMACHINE T3, Ambion), and injected into fertilized eggs at a concentration of 1 mg/mL. DIG-RNA probes for whole-mount in situ hybridization were synthesized by in vitro transcription with T7 RNA polymerase. The detailed procedure has been described previously [10].
To detect activation of the receptor-tyrosine kinase cascade, embryos were fixed with 3.7% formaldehyde and were treated with 3% H2O2 for 30 minutes to quench endogenous peroxidase activity, and then incubated overnight with mouse anti-dpERK (1∶1000, Sigma, M9692) in Can-Get-Signal-Immunostain Solution B (TOYOBO). The signal was visualized with a TSA Kit (Invitrogen) using HRP-conjugated goat anti-mouse IgG and Alexa Fluor 488 tyramide. To visualize cell morphology, F-actin was stained with Alexa Fluor 555–conjugated Phalloidin (Invitrogen).
The contact surfaces of individual animal blastomeres of the 32-cell embryo with cells expressing EphrinA-d were calculated using the 3D-virtual embryo [12]. Given the delay between gene expression and protein translation, we assumed that cells descended from EphrinA-d-expressing cells at the 16-cell stage express EphrinA-d protein at the 32-cell stage. Because EphrinA-d is GPI-anchored, we ruled out autoregulatory effects. The contact surfaces of individual animal blastomeres of the 32-cell embryo with anterior vegetal cells expressing Fgf9/16/20 were previously calculated [12]. However, because Fgf9/16/20 is also expressed in posterior vegetal cells, we included the posterior vegetal cells in our present calculations using the 3D-virtual embryo [12].
DNA constructs for examining regulatory elements were introduced by electroporation [39]. Cis-regulatory elements of Otx and Nodal were fused to the Brachyury and Fog basal promoters [17], [40], [41], respectively. LacZ was used as a reporter gene. The expression of LacZ was examined by in situ hybridization.
|
10.1371/journal.pcbi.1004103 | An Integrated Approach to Reconstructing Genome-Scale Transcriptional Regulatory Networks | Transcriptional regulatory networks (TRNs) program cells to dynamically alter their gene expression in response to changing internal or environmental conditions. In this study, we develop a novel workflow for generating large-scale TRN models that integrates comparative genomics data, global gene expression analyses, and intrinsic properties of transcription factors (TFs). An assessment of this workflow using benchmark datasets for the well-studied γ-proteobacterium Escherichia coli showed that it outperforms expression-based inference approaches, having a significantly larger area under the precision-recall curve. Further analysis indicated that this integrated workflow captures different aspects of the E. coli TRN than expression-based approaches, potentially making them highly complementary. We leveraged this new workflow and observations to build a large-scale TRN model for the α-Proteobacterium Rhodobacter sphaeroides that comprises 120 gene clusters, 1211 genes (including 93 TFs), 1858 predicted protein-DNA interactions and 76 DNA binding motifs. We found that ~67% of the predicted gene clusters in this TRN are enriched for functions ranging from photosynthesis or central carbon metabolism to environmental stress responses. We also found that members of many of the predicted gene clusters were consistent with prior knowledge in R. sphaeroides and/or other bacteria. Experimental validation of predictions from this R. sphaeroides TRN model showed that high precision and recall was also obtained for TFs involved in photosynthesis (PpsR), carbon metabolism (RSP_0489) and iron homeostasis (RSP_3341). In addition, this integrative approach enabled generation of TRNs with increased information content relative to R. sphaeroides TRN models built via other approaches. We also show how this approach can be used to simultaneously produce TRN models for each related organism used in the comparative genomics analysis. Our results highlight the advantages of integrating comparative genomics of closely related organisms with gene expression data to assemble large-scale TRN models with high-quality predictions.
| The ever growing amount of genomic data enables the assembly of large-scale network models that can provide important new insights into living systems. However, assembly and validation of such large-scale models can be challenging, since we often lack sufficient information to make accurate predictions. This work describes a new approach for constructing large-scale transcriptional regulatory networks of individual cells. We show that the reconstructed network captures a significantly larger fraction of cellular regulatory processes than networks generated by other existing approaches. We predict this approach, with appropriate refinements, will allow reconstruction of large-scale transcriptional network models for a variety of other organisms. As we work towards modeling the function of cells or complex ecosystems, individually reconstructed network models of signaling, information transfer and metabolism, can be integrated to provide high information predictions and insights not otherwise obtainable.
| Coordinating cellular behavior in response to internal or external signals requires dynamic regulation at several levels [1,2]. Our ability to understand cellular dynamics requires detailed knowledge of each regulatory network and will, in part, depend on our ability to reconstruct models that integrate the datasets that report on these processes. Of the various levels at which cellular activities are regulated, transcriptional regulatory networks (TRNs) represent a particularly active area for modeling, as high-throughput techniques to monitor RNA levels and protein-DNA interactions can be applied in a wide range of organisms [2,3]. Using such datasets, one can analyze, model, and reverse-engineer TRNs [3,4].
Many published approaches to TRN inference depend on gene expression datasets to make predictions about direct interactions between transcription factors (TFs) and their target genes, assuming that the expression profile of a gene or cluster of genes, is directly related to that of a cognate TF(s) [5–11]. However, predictions based on this premise alone can be compromised by well-known indirect effects (e.g., co-expressed but not co-regulated genes) and post-transcriptionally regulated TFs, whose cellular levels remain relatively constant under conditions where their activity is significantly altered. In attempts to improve the TRN inference process, sequence analysis of the promoter regions of target genes has been used to inform models on the likelihood of a TF directly regulating a set of target genes [5,6,12–16]. However, there is intrinsic statistical variability in the definition of gene clusters obtained from co-expression analyses. Consequently, identifying directly co-regulated genes (i.e., genes that are both co-expressed and share conserved upstream regulatory sequences) is particularly challenging, as de novo identification of functional DNA binding motifs from co-expression clusters is hampered by the fact that the functional sequences of interest are often underrepresented [17].
Comparative genomics analysis of closely related organisms can facilitate identification of functional regulatory motifs by increasing the signal to noise ratio in the input DNA sequences that are used for de novo motif detection [13–15]. The apparent conservation of TFs and regulatory interactions across species has been leveraged to build TRNs across related species [13–16]. However, computational prediction of the presence of a shared DNA motif that is associated with the promoter in a group of genes should not be the only criterion for determination of co-regulation, as co-regulated genes would also be expected to share similar expression profiles under some conditions.
While these individual approaches to TRN inference have their strengths and limitations, they can be complementary and could potentially be combined to construct TRNs of greater coverage and better predictive power [3,6]. However, no integrated workflow currently exists that systematically combines these potentially complementary concepts. Thus, we sought to develop an approach for reconstructing large-scale TRNs that would integrate these various ideas to generate TRN models with higher information content and greater depth.
To achieve this goal, we developed a workflow to construct TRNs, which integrates comparative genomics data, global gene expression analyses, and intrinsic properties of transcription factors (TFs). Intrinsic properties comprise several well-known characteristics of bacterial TFs such as the proximity of TF structural genes to their binding sites [12,14,18,19], the correlation of expression profiles of TFs and their target genes [3,6–8], the similarity in DNA motifs bound by TFs having similar DNA binding domains [19,20] and the co-occurrence of TFs and their binding sites across species [19]. While these properties are established features of many bacterial TFs, they have not been systematically leveraged in the large-scale inference of TRN models. We assessed the function of such an integrated workflow using benchmark datasets for the well-studied bacterium Escherichia coli and we show that it is able to capture a significant portion of the known E. coli TRN. Furthermore, we show this integrated network provides significantly improved predictive power over expression-based inference approaches. We also observed that the content of the TRN models derived from our integrated workflow and from expression-based approaches are complementary, providing an opportunity to combine the TRN models derived from these different approaches.
We also used this workflow to construct and evaluate a large-scale TRN model for the metabolically versatile α-Proteobacterium Rhodobacter sphaeroides. R. sphaeroides is a purple non-sulfur bacterium that has been studied for decades as a model system for photosynthetic growth, being used to understand photon capture, light-driven energy metabolism, and other aspects of the photosynthetic lifestyle [21,22]. In addition to anoxygenic photosynthetic growth, this facultative bacterium is capable of aerobic and anaerobic respiration [22]. R. sphaeroides can also fix CO2 and N2, and produce H2, polyhydroxybutyrate or other compounds of industrial importance [21–30]. Thus, gaining a detailed understanding of its TRN will be pivotal in extending our knowledge of how these various lifestyles and metabolic processes are regulated. Using our integrated workflow, we identified clusters of co-regulated genes in R. sphaeroides and made predictions on DNA binding proteins that are likely to regulate these gene clusters. By focusing on several major sub-networks, we show that predictions of our TRN are consistent with prior knowledge in R. sphaeroides and related bacteria. In addition, experimental analysis of select TFs using chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) and global gene expression analyses provided direct validation of the predictive power of this large-scale R. sphaeroides TRN model. Our analyses illustrate the utility of this integrated approach to assemble TRN models that provide new insights into important biological processes and highlight the role of large-scale TRN inference in driving scientific discovery.
We developed an integrated inference approach to reconstruct large-scale TRNs that uses both sequence information from closely related bacteria and gene expression data, while taking into consideration known properties of bacterial TFs (summarized in Fig. 1). Gene clusters generated by this integrated approach could conceptually be thought of as being co-regulated, as they would share similar expression profiles and evolutionarily conserved upstream DNA sequence motifs. Furthermore, the prediction of TFs that directly control expression of these co-regulated clusters would not depend solely on expression information, potentially enabling more accurate TF-cluster assignments, even for post-transcriptionally regulated TFs whose expression profiles might be unrelated to those of their target genes.
The key steps in our workflow are summarized in S1 Fig., with implementation details of each step provided in the Material and Methods section. Several of these steps involve the use of a variety of well-established public domain algorithms and software packages, which are systematically integrated with new algorithms to build an automated workflow. Below, we summarize the keys steps in this workflow.
Selecting organisms for phylogenetic footprinting. To incorporate comparative genomics into TRN inference, our workflow begins with the selection of appropriate organisms for phylogenetic footprinting. The selection of organisms is critical for this analysis, as organisms that are too closely related may be uninformative, while organisms that are too distantly related may not possess conserved regulatory modules to inform the construction of highly predictive models (see Materials and Methods). Our analysis indicates that as few as 6 appropriately selected organisms could be sufficient to conduct a robust analysis, with addition of more species only providing marginal benefit to the TRN predictions (S2 Fig.). However, as the most appropriate organisms to use are not always known a priori, using a larger selection of organisms may be beneficial.
Identification of orthologs. Prior to de novo motif detection, orthologous genes shared between the selected organisms have to be identified. Approaches for predicting orthologs such as bidirectional best BLAST hits can provide satisfactory results for ortholog predictions in prokaryotic genomes [31]. However, the orthoMCL algorithm [32], which builds on bidirectional best BLAST hits by implementing additional normalizations for protein lengths and uses the Markov cluster algorithm (MCL) [33] to group orthologous proteins from multiple species, provides an automated approach to ortholog identification across multiple organisms that can yield improved results. Thus, our workflow leverages orthoMCL analysis to identify orthologs shared among the organisms selected for the analysis, with all orthologs of a given gene forming an orthologous group.
De novo motif detection. After identification of orthologs shared across species, de novo motif detection analysis is conducted on the intergenic regions of all the genes coding for proteins within a given orthologous group (S2 Fig.). From our analysis, we found that MEME [34] enabled the identification of a wide variety of evolutionarily conserved motifs and performed better than a Gibb’s sampling based approach [35,36]. These evolutionarily conserved DNA motifs are then used to scan the entire genome for other candidate sites, which are clustered based on sequence similarity (see Materials and methods). This results in the generation of clusters of genes with conserved upstream DNA sequence motifs.
Integration of gene expression data. In addition to containing shared upstream regulatory motifs, co-regulated genes might also be expected to have common or similar expression patterns, at least under a subset of conditions. Thus, approaches for reconstruction of TRNs should use both types of information, when available, to build higher confidence networks. To integrate information captured in comparative genomics-based gene clusters with gene expression data, our workflow uses DISTILLER [37]. DISTILLER is a bi-clustering algorithm that identifies conditions or sub-conditions (biclusters) under which a group of genes share a strong co-expression pattern, as condition-dependent regulation of genes means they may not share strong co-expression profiles across the entire dataset. Thus given pre-specified a group of genes (in this instance based on the presence of a shared evolutionarily conserved motif), DISTILLER is used to identify sub-conditions under which these genes share a significant co-expression pattern. We use this approach to generate clusters of “co-regulated” genes having both shared DNA sequence motifs and gene expression patterns.
Linking TFs to clusters. The task of predicting the TF(s) that regulate genes or gene clusters is typically carried out by assessing the relationship between the expression profiles of TFs and their predicted targets [3,5,9,38]. While this approach has been successfully applied in bacterial systems, it is of limited use in eukaryotes [3,38]. However, even in bacteria many TFs are post-transcriptionally regulated, and therefore, their expression profiles are unlikely to share any relationship to those of their target genes. This can lead to spurious predictions when using gene expression data alone. Use of prior knowledge about the properties of TFs, beyond just correlated expression profiles, could facilitate prediction of target genes of such TFs. Thus, in order to link known or predicted TFs to the putative co-regulated gene clusters, our workflow takes advantage of four known characteristics of bacterial TFs: (i) correlation in expression profiles between a TF and its target genes [3,6–8]; (ii) proximity of a TF to the location of the closest binding site within a given cluster (since many bacterial TFs are either auto-regulatory or bind to locations in close proximity to their structural genes) [12,14,18,19]; (iii) similarity in DNA motifs bound by TFs having similar DNA binding domains (since TFs belonging to the same protein families often bind to similar DNA sequence motifs) [19,20]; and (iv) phylogenetic correlation of the occurrence of a TF and occurrence of a DNA sequence motif across species, (assuming that a DNA sequence motif is likely present in an organism if the TF which recognizes this site is also encoded in its genome) [19].
Given the set of known or predicted TFs in the organism of interest, T = {TF1, …, TFi} (where i is the total number of TFs in the organism), and the set of all predicted gene clusters, C = {Cluster1, …, Clusterj} (where j is the total number of predicted clusters), these four properties are integrated as follows:
Correlation: To use correlation to discriminate between potential transcriptional regulators of a cluster of putatively co-regulated genes, the average Pearson’s correlation coefficient (Corrmean) was determined for each TF per gene cluster (eqn. 1) (S3A Fig.). This was achieved by determining the correlation of the expression values between a given TF (TFx) and each gene (gk) within a given cluster (Clustery) containing n genes, across the subsets of conditions under which the gene is tightly co-expressed with others in the cluster. The absolute values of these TF-gene correlations are then averaged to obtain a TF-cluster Corrmean (eqn. 1). This is carried out for all TFs in the target organism to determine the average correlation of each TF in relation to each cluster. These average correlation scores are then converted into p-values (Pcorr) by random permutation. Briefly, 1000 TF-cluster Corrmean scores were randomly generated, then each previously calculated TF-cluster Corrmean was compared to the set of randomly generated values. The total number of randomly generated scores greater than or equal to a given TF-cluster Corrmean divided by 1000 was used as an estimate of the p-value (eqn. 2).
Proximity: To use the proximity of TFs to link them to their binding sites, we determined the minimum distance (in number of genes) between each TF’s location in the genome and the genes present in a given cluster (eqn. 3) (S3B Fig.). Here, the proximity score would have a value of 0 (if the TF is a member of a cluster for a given TF-cluster pair) or larger. This proximity score is determined for every TF-cluster pair where at least one member of the cluster is located on the same replicon as the TF. These minimum distance scores (Proxmin) were also converted into p-values (Pprox) by random permutation as described above (eqn. 4).
DNA binding domain: To incorporate information on DNA binding domain (DBD) similarity into TRN predictions, we begin by determining the DBD family to which each TF in the target organism belongs to using Pfam analysis [39]. All E. coli TFs from RegulonDB [40], which had binding motif information (81 at the time of this analysis), were retrieved and their DBD families also determined using Pfam. Position specific scoring matrices (PSSMs) for the DNA binding sites of the RegulonDB TFs and each of the evolutionarily conserved de novo detected motifs are then constructed. For each TF-cluster pair to be assessed, the PSSM for the de novo detected motif of the cluster under consideration was compared to the PSSM(s) from E. coli whose associated TF(s) belongs to the same DBD family as the TF under consideration. This TF was then assigned the most significant (smallest) q-value from this set of comparisons. For instance, if TFx is a Crp family TF, to assign a score to TFx in relation to Clustery, the PSSM for Clustery is compared to all available Crp family PSSMs from the RegulonDB data set and TFx is assigned a value equivalent to the most significant match to these PSSMs (S3C Fig.). These q-values were then—log10 transformed to generate the DBD_score for that TF-cluster pair. PSSM comparisons were made using Tomtom [41,42] and all possible TF-cluster pairs were assessed similarly. These DBD_scores were converted into p-values (Pdbd) by random permutation as previously described (eqn. 5).
Phylogenetic correlation: To compute a score for this property, we first determine the occurrence of a given motif across all the genomes used in the analysis. For each de novo detected motif, we use MAST [41] to search for all instances of that motif in the intergenic regions of each organism used for phylogenetic footprinting. These genome-wide p-values of MAST hits for a given motif were stored in separate vectors for each genome. The correlation was then calculated between the MAST hits p-value vector of the target organism and that for each species used for phylogenetic footprinting (target organism inclusive). These correlations were referred to as “motif occurrence correlations” (S3D Fig.). We then determined the occurrence of each TF in the target organism across all the species used for phylogenetic footprinting via orthoMCL analysis. Finally, the correlation between the “motif occurrence correlation” and TF occurrence was calculated to determine the phylogenetic correlation. These phylogenetic correlation scores were converted to p-values by random permutation as described above (eqn. 6).
Combining scores: To rank candidate TFs, the -log10 of the computed p-values for the 4 different criteria were summed together to generate a final score Rscore (eqn. 7), resulting in a ranked list of TFs most likely to regulate a given cluster.
Predicting regulatory interactions from global gene expression data. The integrative approach described above identifies conserved clusters of putatively co-regulated genes, but its utility can be limited by the evolutionary distance and the degree of conservation of the individual regulatory modules across the organisms used to generate the TRN. For example, it may be difficult to identify conserved regulatory sequences across closely related species if these sequences or regulatory mechanisms have undergone significant evolution. Furthermore, individual sub-networks that are specific to a lifestyle or response of an individual species, genus and/or clade might not be captured via a comparative genomics-based approach. Thus, to complement predictions from the comparative genomics-based analysis, we considered the consensus predictions of multiple high performing direct expression-based inference approaches [3,9–11] to make predictions for additional TFs not included in the comparative genomics-based TRN. In particular, we used the consensus predictions from 3 approaches: context likelihood of relatedness (CLR), which uses normalized mutual information-based scores, as an indication of the relatedness of expression profiles, to assess potential TF-target interactions [9]; GENIE3, which uses multiple regression and tree-based feature selection to identify TFs whose expression profiles are most predictive of a given target gene [10]; and an approach which uses analysis of variance (ANOVA) to score how dependent the expression profile of a target gene is to potential transcriptional regulators [11]. The predictions from these approaches were combined using methods similar to those previously used for generating consensus networks from approaches assessed in the DREAM challenges [3]. Details of this are provided in the Materials and methods (see “Inferring regulatory interactions solely from expression data”).
The networks predicted from comparative genomic-based integration and the gene expression-based consensus network were then combined. This was achieved by taking the integrated comparative genomic-based TRN as the core of the network, then augmenting it by including high-scoring predictions for TFs not already included in the integrated network. We chose this approach based on observations from analysis of the E. coli TRN (see below).
To assess the performance of the integrated workflow outlined above, we built a TRN for E. coli using sequence data from 14 enterobacteriales species (including E. coli) obtained from NCBI and curated expression data obtained from the many microbes microarray database [9]. The TRN built using the described integrative comparative genomics-based component of our workflow consisted of 225 motifs and clusters, 1660 genes, 126 TFs and a total of 2457 interactions (S1 Dataset). In addition, 156 of the 225 clusters were significantly enriched for at least one functional category. These predictions were compared to similar sized TRNs (2500 highest ranked interactions) generated by CLR and GENIE3 using the same gene expression dataset. These TRN models were then validated against an experimentally verified list of regulatory interactions from regulonDB [40].
A widely used metric for assessing the performance of TRN inference approaches is the area under a plot of precision against recall for inferred TF-target gene interactions [4,43]. Assessing this area under the precision-recall curve (AUPR), we observed that the integrated approach performed significantly better than CLR or GENIE3, both when all predicted interactions were considered (AUPR ~3 times larger) and when only interactions for TFs with experimental data were considered (AUPR ~1.5 times larger) (Fig. 2A). At a precision of 25%, CLR and GENIE3 and the integrated TRNs achieved a recall of 1.8%, 2.1% and 5.7%, respectively. These analyses indicate that our comparative genomics-based integrated approach is more accurate and able to capture a larger fraction of known regulatory interactions. It should be noted that for this analysis, only the highest scoring TF predicted for each cluster was used to build the final list of predicted interactions for the integrated TRN. In some instances other high scoring TFs may actually be the direct regulators, but these were not considered here.
Integrated and expression-based networks are complementary. While the above analysis highlights the improved performance of the integrated approach over the expression-only TRN inference, it may be more informative to examine the predicted interactions and assess where each approach excels or fails, to determine if there is any complementarity between these approaches. Of the 81 TFs for which experimentally verified interactions exist in the regulonDB dataset used in our analysis, CLR, GENIE3 and the integrated TRN models were able to make at least one correct prediction for 28, 23 and 30 of these TFs respectively (Table 1, Fig. 2B). While CLR and GENIE3 use different approaches to infer their TRNs, there is a large overlap in the TFs for which they make predictions (Fig. 2B, Table 1). This is consistent with previous observations from analysis of TF-target interactions conducted as part of a comprehensive assessment of expression-based inference approaches [3]. Overall 96% of the TFs for which GENIE3 made correct predictions could also be captured using CLR and 79% vice versa, though the precision and recall for each of these TFs varies between approaches (Table 1). Conversely, only 57% and 61% of the TFs for which CLR and GENIE3 made predictions for, respectively, also had predictions in the integrated TRN, while predictions for 43% of the TFs in integrated network were unique to this approach (Fig. 2B).
These observations indicate that there are specific subsets of TFs that are amenable to predictions using expression-based assumptions. However, many TFs that are not amenable to analysis solely by expression-based analyses can be correctly assigned in a TRN constructed using an integrative approach. This is potentially due to instances where the expression profile of a TF does not show any significant relationship to those of its target genes. This typically occurs for TFs that are known to be post-translationally regulated such as FNR, ArgR, Fur, Cra etc [44–47] (Table 1). On the other hand, for several of the TFs where expression-based approaches performed better, the integrated approach failed to make any prediction (Table 1). This could be the result of a number of factors including lack of conservation of TF binding sites, small regulon size, complex DNA binding motifs or limitations in the motif detection algorithm utilized. Importantly, for TFs for which predictions were made by all three approaches, the predictions from the integrated approach were in general on par with, or better than, those obtained with expression-based approaches (Table 1). These observations lend themselves to a straight-forward approach for combining these approaches wherein the integrated comparative genomics-based network serves as the core of the TRN, and is complemented with high scoring predictions from expression-based approaches for TFs not already captured in the core network.
Using the same workflow and leveraging the observed complementarity of integrative and expression-based approaches, we generated a large-scale TRN model for the metabolically versatile photosynthetic bacterium R. sphaeroides. In this case, we used sequence information from 8 closely related α-Proteobacteria, including R. sphaeroides (S4 Fig.) and gene expression data from 198 experiments (S1 Dataset). The resulting TRN model consists of 120 clusters, 93 TFs, 76 distinct evolutionarily conserved DNA sequence motifs and 1858 TF (or motif)-target interactions (S5 Fig., S1 Table). This model includes a total 1211 R. sphaeroides genes (about 28% of the open reading frames predicted in its genome [48,49]). Below, we provide an overview of some of the pertinent predicted sub-networks in the TRN, as well as experimental validation of some key TFs in the network.
Reconstructed TRN encompasses a wide variety of functions. The R. sphaeroides TRN model encompasses a wide variety of cellular functions ranging from central carbon metabolism and global stress responses, to processes more specific to R. sphaeroides, such as nitrogen fixation and photosynthesis (Fig. 3, S5 Fig.). Of the 120 identified gene clusters, 80 were significantly enriched for at least one gene ontology (GO) [50] category (S1 Table, Fig. 3), indicating this TRN model captures a high degree of functional information even though this type of functional data was not used in the network inference workflow.
Photosynthesis. Previous analyses of the photosynthetic lifestyle of R. sphaeroides have implicated 3 TFs in this process: PpsR [51,52], FnrL (a homolog of FNR) [53–55] and PrrA (the response regulator of the PrrAB two component system) [56–60] (Fig. 4). More recently a small non-coding RNA, PcrZ has been implicated in the regulation of photosynthesis in R. sphaeroides [61]. Despite extensive prior analysis, our TRN model predicts at least 2 additional regulators of photosynthesis: CrpK (RSP_2572) and RSP_2888 (Fig. 4). To illustrate the predictive ability of our TRN, below we provide details about the known or predicted TFs in the R. sphaeroides photosynthetic lifestyle.
Previous analysis of PpsR (RSP_0282) identified this TF as a repressor of photopigment production under aerobic conditions [51,52,63,64]. The activity of PpsR is regulated by its cognate anti-repressor, AppA, which is reported to respond to both oxygen and blue light [65–68]. To gain a more complete picture of the PpsR regulon, as well as assess the predictive performance of our inferred TRN for this TF, we determined the genome-wide binding of PpsR to its target sites by ChIP-seq using a 3X-myc tagged PpsR protein that complements a defined ΔppsR mutant. We identified a total of 19 PpsR binding sites in the genome that were located upstream of 15 operons, only 2 of which had been previously verified as direct targets for this TF [51] (Table 2, Fig. 5A). Consistent with its role in regulation of photopigment formation, the majority of PpsR target operons had known or predicted photosynthesis-related functions (Table 2). Interestingly, PpsR was bound upstream of the prrA gene, which encodes another transcriptional regulator of photosynthesis in R. sphaeroides [56–60], suggesting a previously unknown genetic interaction between these TFs.
In addition to photosynthesis-related targets, PpsR was bound upstream of RSP_2095 and RSP_3000, which encode proteins of unknown function. However, these genes were not found to be significantly differentially expressed (DE) in a pair-wise comparison of RNA levels between a ΔppsR mutant and its parental strain [51], nor did their expression profiles show significant correlation to other members of the PpsR regulon across the available microarray dataset compendium (Fig. 5B), suggesting these might represent non-functional binding sites in the genome, despite possessing strong PpsR motifs (Table 2). Consistent with the known role of PpsR as a transcriptional repressor, all DE PpsR targets we identified were predicted to be repressed by PpsR as RNA levels were increased in cells lacking this TF (Table 2).
Our TRN predicted a total of 13 PpsR target operons, 12 of which were verified via ChIP-seq analysis (S1 Table (cluster 60), Table 2), corresponding to a recall of 80% (i.e., 12 of 15 PpsR ChIP-seq identified sites were predicted) and a precision of 92.3% (i.e., 12 of 13 predicted target sites were accurate). The only predicted PpsR target site not verified by ChIP-seq analysis (RSP_4172—a hypothetical protein) was classified as a false-positive since enrichment for PpsR binding was not detected by subsequent ChIP-qPCR analysis under the growth conditions tested (S6 Fig.). On the other hand, 3 PpsR sites identified in our ChIP-seq assay were not predicted in our TRN (RSP_2095, RSP_3000 and hemE). However, given that putative targets such as RSP_2095 and RSP_3000 were not DE in the absence of PpsR (Table 2, Fig. 5B), these might represent non-functional or false positive binding events. Independent ChIP-qPCR validation of ChIP-seq identified sites suggest that RSP_2095 and RSP_3000 are likely bound by PpsR but not DE under the conditions tested (S6 Fig.). Overall, our inferred TRN provided an accurate and expanded picture of PpsR binding sites across the genome with a large coverage of true binding sites. Accordingly, the consensus DNA sequence motifs obtained for PpsR from ChIP-seq and phylogenetic footprinting analysis are very similar (Fig. 5C).
FnrL (RSP_0698) is an iron-sulfur cluster-containing Crp-family TF which previous studies have reported to be essential for anaerobic growth in R. sphaeroides [54,55]. Previous ChIP-chip analysis of genome-wide FnrL binding sites in vivo indicated the direct involvement of this TF in a host of processes including photosynthetic and anaerobic respiratory growth [53]. Our inferred TRN captured a significant portion of the known FnrL regulon, predicting a total of 59 FnrL target operons (S2 Table, S1 Table (cluster 11)) that included 24 of the 25 previously identified FnrL target operons, a recall of 96%. The only previously verified FnrL target operon not identified in our analysis was RSP_6116, which is not represented on the R. sphaeroides Affymetrix gene chip, and thus dropped out during the integration of gene expression data. In addition to previously identified sites, our large-scale TRN predicted an additional 35 FnrL target operons not previously known or predicted to be under the control of FnrL (S2 Table). Each of these new FnrL target operons have putative binding sites with strong similarity to the FnrL consensus and share a similar expression profile with other members of the FnrL cluster (S3 Table). Several of these newly predicted FnrL targets encode functions for which this TF has been previously implicated; including the regulation of Fe-S cluster biogenesis (e.g., RSP_1949) and Fe-S binding proteins (e.g., RSP_0692_89—rdxBHIS). However, several new functions for FnrL that are predicted in this data set need to be tested experimentally. If these predictions are correct, it would significantly broaden the functional role of FnrL in this species.
In addition to PpsR and FnrL, whose regulons were globally characterized in this or previous studies, our TRN model also made predictions for direct targets of less-well characterized TFs. For instance, our TRN model made several new predictions for targets of the photosynthesis regulator PrrA (RSP_1518). PrrA has previously been proposed to be major global regulator in R. sphaeroides and other bacteria [57]. PrrA is essential for photosynthetic growth in R. sphaeroides and direct control of photosynthesis related operons, tetrapyrolle biosynthesis (hemA) and the Calvin—Benson—Bassham (CBB) cycle genes has be shown in vitro [59,70]. Our TRN predicts that a total of 17 operons are directly regulated by PrrA (S1 Table (cluster 96)). Of these, 7 predicted PrrA target operons have a photosynthesis related role, including pufLMX (RSP_0255–7), pufA (RSP_0258), ppaA (RSP_0283), bchFNBHLM-puhA (RSP_0284–91), hemC (RSP_0679), hemA (RSP_2984) and appA (RSP_1565). However, only two of these operons (bchF and hemA) have previously been experimentally verified as PrrA-dependent in R. sphaeroides [70], so direct analysis of PrrA binding to these newly proposed targets is required.
CrpK (RSP_2572) is a Crp/Fnr-family TF, which possesses predicted cyclic nucleotide-binding and Crp-like helix-turn-helix domains. However, unlike FnrL, CrpK does not possess N-terminal cysteine residues required for coordination of iron-sulfur clusters, suggesting CrpK might not directly sense oxygen. Our TRN predicts that CrpK regulates overlapping targets to FnrL, including several photosynthesis related operons such as bchEJGP (RSP_0281–76) and hemA (RSP_2984) (S1 Table (cluster 105)), as well as several other known FnrL target genes including nuoA-N (RSP_0100–12) and ccoNOQP (RSP_0696–3), amongst others. These predictions suggest CrpK could substitute for FnrL under some conditions, providing added, previously unappreciated, robustness to the photosynthetic TRN of this bacterium and possibly others containing homologs of both FnrL and CrpK. The overlapping nature of the CrpK and FnrL regulons was recently demonstrated experimentally [71].
RSP_2888 (recently renamed MppG [71]) is a BadM/Rrf2 family TF predicted by our TRN to control photosynthesis gene expression in R. sphaeroides. Predictions from our TRN suggest a direct role of MppG in the regulation of a bacteriochlorophyll biosynthesis operon bchFNBHLM (RSP_0284–91), in addition to key photosynthesis related genes, such as appA (RSP_1565) (S1 Table (cluster 110)). MppG mRNA levels are increased under photosynthetic conditions in our expression datasets and this gene is predicted in our TRN to be under the control of PrrA. These observations are consistent with a role for MppG in the photosynthesis sub-network of the TRN, which has been experimentally verified [71].
Overall our TRN captures a significant portion of the known regulatory interactions in the photosynthesis sub-network (Fig. 4), while making a large number of novel predictions that should provide new insights into the complex combinatorial regulation of this lifestyle in PNB.
Central and alternative carbon metabolism. For cells to survive in nature, they must adapt to the types and quantities of nutrients present in their environment. For instance, E. coli uses the cAMP receptor protein (CRP), in part, to preferentially utilize glucose over other nutrient sources, if present in its environment [72]. On the other hand, the ArcAB two-component global regulator represses portions of E. coli’s central metabolic pathways under anaerobic respiratory conditions [73,74]. In addition to these global regulators, the Cra/FruR regulator specifically regulates carbon and energy metabolism in enteric bacteria [47].
R. sphaeroides is not predicted to possess proteins analogous to CRP or ArcAB. However, our TRN predicts that the regulation of central carbon metabolism in R. sphaeroides is controlled by a LacI family transcriptional regulator, RSP_1663. RSP_1663 is predicted to regulate transcription of genes encoding the central carbon metabolism enzymes Mdh (RSP_0968), PckA (RSP_1680), malic enzyme (RSP_1593), PdhAB (RSP_2968-RSP_4047-RSP_4050), succinate dehydrogenase (RSP_0974–6), as well as glycolytic enzymes Zwf (RSP_2734), Pgl (RSP_2735), Pgi (RSP_2736) and FbaB (RSP_4045), potentially making this TF a major regulator of carbon metabolism under many conditions (Fig. 6). This predicted RSP_1663 regulon might make it functionally analogous to the Cra/FruR regulator in enteric bacteria [47] and the RpiR family TF HexR in β- and γ-proteobacteria [75]. RSP_1663 is predicted to bind to an inverted repeat DNA motif with the sequence [A/G/T]GTT N6–8 AAC[A/C/T] (where N is any nucleotide) (Fig. 6). In addition, differences in spacer between the inverted repeats divides the genes predicted to be regulated by this TF into 2 clusters (S1 Table (clusters 15 and 36)). Further experimental analysis is needed to understand the functional role of RSP_1663.
In addition to RSP_1663, RSP_0981—a GntR family transcriptional regulator, is predicted to regulate transcription of genes encoding the succinyl-CoA synthetase (RSP_0967–6), succinate dehydrogenase (RSP_0974–6) and α-ketoglutarate dehydrogenase (RSP_0965–62) complexes of the tricarboxylic acid cycle (Fig. 6, S1 Table (cluster 48)), while NtrC (RSP_2838) is also predicted to be involved in the regulation of the succinate dehydrogenase complex (Fig. 6, S1 Table (cluster 1)). Cluster 62 in our TRN (Fig. 6, S1 Table) also contains a number of genes encoding enzymes involved in central carbon metabolism including Icd (RSP_0446 and RSP_1559), L-malyl-CoA lyase (RSP_1771), citrate synthase (RSP_1994) and NuoA-N (RSP_2512–23). The members of cluster 62 share the inverted repeat motif (Fig. 6), indicating that these central metabolism genes are under the joint control of an as yet unidentified TF.
Our TRN also made predictions about regulation of metabolism of several other carbon sources. For instance, RSP_0489—a GntR family transcriptional regulator, is predicted to regulate transcription of genes encoding enzymes that are involved in the metabolism of carboxylic acids including UxuA (RSP_0773), UxaC (RSP_0488), KduID-UxuB (RSP_0482–80) and carbohydrate kinase (RSP_0490), as well as substrate transport (RSP_0487–3 and RSP_3168–5) (Fig. 6, S1 Table (cluster 83)), making it functionally analogous to UxaR [76]. We tested these predictions by comparing RNA levels between wild type (WT) and ΔRSP_0489 cells, and conducting ChIP-seq analysis with a myc-tagged version of RSP_0489 (Fig. 7). A total of 55 genes were DE (1.5 fold change, pvalue < 0.05) between WT and ΔRSP_0489 cells, including predicted targets uxuA, kduID-uxuB, uxaC and RSP_0487–3, which were repressed in the presence of RSP_0489 by as much as 36-fold (Fig. 7A, S4 Table). Several other genes involved in substrate transport and metabolism were also DE in this data set (Fig. 7A, S4 Table). ChIP-seq analysis with a 3X myc tagged variant of RSP_0489 revealed that RSP_0489 binds at the promoters for uxuA (RSP_0773), the uxaC operon (RSP_0488–0), RSP_0489, RSP_0490 and within the coding regions of substrate transporter (RSP_3372–70 and RSP_2667–3) (Fig. 7B, Table 3), verifying several predictions from our TRN model. Overall 4 out of these 6 RSP_0489 target operons (~67%) were correctly predicted in our TRN. The conserved DNA sequence motif derived from sites bound by RSP_0489 also showed similarities to that obtained from phylogenetic footprinting analysis of the RSP_0489 promoter (Fig. 7C). Other genomic locations enriched for RSP_0489 but with no corresponding DE genes are listed in S5 Table.
Fe-S cluster biogenesis and iron homeostasis. Genes of the Fe-S biogenesis pathway (iscSUA-hscBA-fdx) are regulated by the Rrf2-family TF IscR, in E. coli and several other bacteria [77,78]. In E. coli, IscR is a global regulator that is able bind to two different DNA target sequences depending on whether it is ligated to a 2Fe-2S cluster [77,78]. The R. sphaeroides homolog of IscR, RSP_0443, differs from E. coli IscR as it does not possess cysteine residues required for the ligation to a 2Fe-2S cluster, suggesting that this protein is unable to ligate a Fe-S cluster. If this is true, then the upstream signaling pathway utilized and target genes regulated by RSP_0443 is likely to differ from that of E. coli IscR.
Consistent with observations in E. coli, RSP_0443 is predicted in our TRN model to regulate transcription of its own operon (RSP_0443–31). However, the RSP_0443 operon encodes homologs of the Suf Fe-S biogenesis pathway (sufABCDSE), which is also a direct IscR target in E. coli [79]. In addition, RSP_0443 is predicted in our TRN model to regulate transcription of catalase (RSP_2779), bacterioferritin-associated ferredoxin (RSP_1547), imelysin (RSP_1548), biopolymer transport protein TonB-ExbBD (RSP_0920–2), napEFDABC (RSP_4112–8), all gene products with predicted Fe-S cluster or heme-binding domains or predicted to be involved in iron uptake (S1 Table (cluster 82)). Thus, members of the predicted RSP_0443 regulon could play a significant role in maintaining cellular iron homeostasis, possibly to provide the metal needed for Fe-S centers. There is also a strong positive correlation between RSP_0443 RNA levels and transcription of its predicted target genes in R. sphaeroides, suggesting this TF functions as an activator.
In addition to RSP_0443, FnrL is directly involved in regulating transcription of genes encoding iron transporters such as feoABC, as well as a number of Fe-S and heme containing proteins in R. sphaeroides. Thus, our TRN predicts that RSP_0443 and FnrL both play an important role in regulation of cellular iron homeostasis. Furthermore, FnrL is also predicted in our TRN to directly activate RSP_3341, a putative iron binding RirA-like [80] protein in R. sphaeroides, which in turn is predicted to negatively regulate the putative 4Fe-4S binding nitrate reductase (napEFDABC). We tested this prediction by comparing RNA abundance levels between wild type (WT) and ΔRSP_3341 cells, and via ChIP-seq analysis using a myc-tagged version of RSP_3341 (Fig. 8A). We found a total of 69 genes were DE (2 fold change, pvalue <0.05) between WT and ΔRSP_3341 cells including several members of the nitrate reductase operon (napEFDABC), which were all repressed by RSP_3341 (Fig. 8A, S6 Table). In addition, transcription of genes encoding other iron dependent proteins (such as cytochromes and ferredoxins) were also repressed by RSP_3341 (Fig. 8A, S6 Table). The mRNA level RSP_0443 was 2 fold higher in WT relative to ΔRSP_3341 cells, suggesting there might be some cross talk between these TFs. We conducted ChIP-seq analysis with a 3X myc tagged version of RSP_3341 and confirmed the direct regulation of napEFDABC by this protein, consistent with our gene expression data and TRN model predictions (Fig. 8B). In addition, RSP_3341 binding was found near Hsp70 DnaK (RSP_1173) and cycJ (RSP_2945) (Fig. 8B, Table 4). These genes were also DE in our gene expression dataset, thus were considered as additional direct RSP_3341 targets (Table 4). Twenty-two other sites showing significant enrichment for RSP_3341 but for which no genes in those genomic locations were DE are provided in S7 Table. These data verify the prediction of our TRN model of the involvement of RSP_3341 in the direct and indirect regulation of iron-dependent genes in R. sphaeroides.
Another RirA-like protein in R. sphaeroides, MppG, predicted to be important in regulation of photosynthesis is also involved in the regulation of iron containing proteins such as AppA (RSP_1565) and those involved in bacteriochlorophyll biosynthesis. Thus, the maintenance of iron homeostasis and the transcriptional regulation of genes encoding iron-dependent enzymes appears to involve a complex gene regulatory network in R. sphaeroides (Fig. 8C).
Other major cellular sub-networks. In addition to the sub-networks described above, many others were predicted in our R. sphaeroides TRN model including networks involved in carbon metabolism, nitrogen metabolism, hydrogen production, DNA repair, flagella biosynthesis and chemotaxis, heat shock and oxidative stress responses, methionine biosynthesis, phosphate transporter and carotenoid biosynthesis (described in S1 Text).
Links between sub-networks in the R. sphaeroides TRN. In addition to the depth and variety of networks captured in our TRN, we also identified several new and interesting links between these predicted sub-networks. For instance, the TRN predicts a previously unrecognized connection between photosynthesis and iron homeostasis in R. sphaeroides. The photosynthesis regulators MppG, CrpK, and FnrL, are predicted to regulate several iron/heme-dependent and iron transport proteins. Furthermore, FnrL is also predicted to regulate RSP_3341, which we have shown in this work to be directly involved in regulation of other iron-dependent genes. These data suggest that regulation of photosynthesis, which employs several iron-dependent proteins, and iron homeostasis need to be coordinated in R. sphaeroides to achieve optimal growth under anaerobic photosynthetic conditions.
NtrC, which is predicted in our TRN to be involved in regulation of nitrogen metabolism (S1 Text), is also predicted to control transcription of genes for central carbon metabolism (Fig. 3 and S5 Fig.), suggesting a possible previously unrecognized link between carbon and nitrogen metabolism in R. sphaeroides. Similar links between carbon and nitrogen metabolism have been identified in B. subtilis via the global regulator of carbon metabolism CcpA [81]. Our TRN also captures previously known links between sub-networks controlling the response to heat shock, singlet oxygen stress and DNA repair (S1 Text).
While this description of sub-networks is by no means exhaustive, it provides a useful overview of the various functionalities and connections captured in the R. sphaeroides TRN model. Overall our TRN model captures a significant amount of known transcriptional regulatory interactions in R. sphaeroides, while predicting a large number of new interactions for this bacterium that are consistent with observations in other organisms. Furthermore, the TRN model also makes a large number of novel predictions unique to R. sphaeroides, which represent high-quality targets for future experimental verification. In sum, given the high predictive ability of our TRN model for characterized TFs, we propose that it provides an excellent roadmap for future analysis of the R. sphaeroides TRN and those of related bacteria.
The integrated TRN inference approach provided significant improvement in information content. We compared the integrated R. sphaeroides TRN model to others built from our gene expression compendium using the direct inference approaches CLR [9] and GENIE3 [10], and a module-base inference approach LeMoNe [82]. Selecting networks of similar size (i.e., the top ~1900 predicted TF-target predictions from each approach), we found that our integrated approach generated a TRN with significantly improved information content (Fig. 9). Of the 120 clusters identified in our TRN, 80 (~67%) were enriched for at least one GO functional category compared to 34, 35 and 53% for networks built with CLR, GENIE and LeMoNe, respectively. This comparison suggests our approach captures more functional information. Furthermore, the number of de novo detected DNA sequence motifs obtained in the integrated TRN (88 motifs corresponding to ~73% of the clusters), significantly supersedes that obtained by searching the intergenic regions of predicted TF targets obtained from CLR, GENIE and LeMoNe analyses (7, 13 and 11 motifs corresponding to 4, 10 and 17% of the clusters respectively) (Fig. 9). These data suggest that while these expression-based approaches can group potentially functionally related and co-expressed genes together, the resulting clusters likely do not include a sufficiently high percentage of co-regulated genes, so the ability to detect conserved promoter motifs from these predicted clusters/regulons is very low. Thus, it appears that initiating TRN inference with motif detection (via phylogenetic footprinting) prior to incorporating expression data significantly improved its information content and allowed us to overcome some of the limitations in gene expression datasets.
While the regulons of only a handful of TFs have been studied on a genome-scale in R. sphaeroides, assessing predictions made for some of these TFs highlights other advantages of an integrated approach. For instance, CLR, GENIE and LeMoNe were not able to accurately predict targets for PpsR or FnrL, likely due to the almost invariant expression profiles of these TFs (see Fig. 5B for ppsR expression), as their activities are regulated post-transcriptionally. However, by taking other features of bacterial TFs into consideration, we were able to accurately link PpsR and FnrL to their respective regulons, while making predictions across our network for other similarly regulated TFs. On the other hand, for the alternative sigma factor σE whose binding elements are separated by a variable length spacer region and whose regulon might differ considerably across the species used in our comparative genomics analysis, the expression based approaches performed better at identifying members of this regulon. Thus, incorporating consensus predictions for expression-based inference approaches allowed us to capture such predictions in our final R. sphaeroides TRN.
Overall, for the 7 TFs for which genome-wide TF-target interaction data exist for R. sphaeroides (including the 3 TFs analyzed in this study), the predictions from the integrated network outperformed that obtained from expression-based inference approaches, achieving an overall precision (and recall) of 75% (32%), compared to 52% (6%), 74% (12%) and 82% (13%) for CLR, GENIE3 and LeMoNe networks respectively (S8 Table).
Though our approach performed relatively well for many R. sphaeroides TFs, a large number of verified target genes were not identified for some regulators (e.g., RpoHI and RpoHII). This could possibly be due to difficulties in discriminating DNA binding motifs for these or other closely related σ-factors as well as limitations in available gene expression data. Alternatively, it could be the result of constraints used in de novo motif detection or limitation of the motif finding algorithm itself. While these constraints performed well at identifying likely binding sites of many traditional TFs, they might be too prohibitive for identification of σ-type motifs.
In addition, of the 81 characterized E. coli TFs used for performance assessment, accurate predictions could be made for 42 (~52%) of them when considering the predictions from both integrative and expression-based approaches. This leaves a relatively large category of TFs for which available datasets do not provide sufficient information or resolution to make predictions at a reasonable level of precision. Thus, advances in algorithmic and experimental methodologies are still required to bridge this gap.
Preliminary TRNs for closely related organisms. An additional benefit of using comparative genomics for TRN inference is that preliminary TRNs can also be built for the other organisms used in the comparative analysis. For instance, the inference of the TRN model for R. sphaeroides served as the basis for the construction of preliminary sequence-based TRNs for R. sphaeroides ATCC 17025, Rhodobacter capsulatus SB 1003, Roseobacter denitrificans Och 114, Dinoroseobacter shibae DFL 12, Rhodopseudomonas palustris CGA009, Bradyrhizobium japonicum USDA 110 and Paracoccus denitrificans PD1222 (S9 Table). We expect that these preliminary TRN models will provide insights into the peculiarities of the TRNs of these α-Proteobacteria. They can also serve as starting points for construction of more detailed global TRNs for these and other related bacteria.
In this study, we developed a new workflow to generate genome-scale TRNs, which integrates genome sequence information and gene expression data, as well as taking into consideration properties of bacterial TFs. Validation of this workflow using benchmark datasets for E. coli showed that it provides significantly improved predictive capability compared to high-performing expression-based approaches. Further analysis of the predicted TRN models showed that the predictions from this workflow and expression-based inference approaches are highly complementary—a feature that could be exploited to build TRN models with greater coverage. We further demonstrated the utility of this workflow by building a large-scale TRN model for R. sphaeroides. The R. sphaeroides TRN model consists of 120 gene clusters and 1858 regulatory interactions encompassing ~28% of the genes for this organism. Several observations indicated that this approach generated a large-scale TRN with high predictive power. The majority of the predicted gene clusters were enriched for specific functions and the genes found in many of these clusters were consistent with prior knowledge in R. sphaeroides or other bacteria. In addition, experimental validation of select R. sphaeroides TFs showed that the TRN assembled via this integrated approach makes accurate predictions for several of these regulators. Our analysis also illustrates the ability of this workflow to generate of large-scale TRN models with increased information content relative to those built via other approaches. An additional benefit of our approach is that it simultaneously enables construction TRN models for other organisms used in the comparative genomics analysis. Overall, the workflow presented here represents a powerful approach by which to reconstruct TRNs for bacteria for which similar data types are available. It has also provided a large amount of new insight into transcriptional regulation in a phototroph, correctly capturing many aspects of the diverse lifestyles of R. sphaeroides, while providing novel predictions into regulatory networks that await experimental validation. Thus, this large-scale TRN model should serve as an indispensable data source for those interested in R. sphaeroides and related bacteria.
To build large-scale TRN models for E. coli and R. sphaeroides, we utilized an approach that combined comparative genomics, gene expression analysis and intrinsic properties of bacterial TFs. The workflow used for our reconstructions is detailed below in a stepwise fashion and summarized in S1 Fig.
Selecting genomes for phylogenetic footprinting. Our TRN reconstruction workflow begins with exploiting the sequence information from closely related bacteria [13–15]. In order to identify evolutionarily conserved sequences upstream of homologous genes across multiple species (i.e., phylogenetic footprinting), it is important that relatively closely related species are used, as regulatory mechanisms are more likely to be conserved across these organisms [83]. However, if species are too closely related analysis of upstream sequences becomes uninformative, as large stretches of identical or highly similar sequences prevent the identification of relevant regulatory sequences. Thus, species selected for phylogenetic footprinting analysis were carefully chosen to increase the utility of this approach [12]. To select organisms for our analyses, we used a combination of orthology, phylogeny and physiological information. We considered 3 factors in organism selection: (i) the number of orthologs shared between a given organism and our target organisms, E. coli and R. sphaeroides (as a larger number of shared orthologs would enable identification of a potentially larger set of regulatory motifs); (ii) phylogenetic distance (as more closely related species would be more likely to have conserved regulatory mechanisms); and (iii) metabolic diversity (in addition to general cellular processes, we considered the regulation of processes peculiar to these metabolically diverse organisms). Based on these criteria, we restricted the organisms selected for phylogenetic footprinting to those belonging to the orders Rhodobacterales and Rhizobiales for R. sphaeroides, as these organisms share a larger number of orthologs with R. sphaeroides (S4 Fig.), are close phylogenetic relatives to R. sphaeroides (S4 Fig.) and are more metabolically diverse than many members of other α-Proteobacterial orders. From these two orders we selected 8 organisms for our phylogenetic footprinting analysis: R. sphaeroides 2.4.1, R. sphaeroides ATCC 17025, R. capsulatus SB 1003, R. denitrificans Och 114, D. shibae DFL 12, R. palustris CGA009, B. japonicum USDA 110 and P. denitrificans PD1222. The criteria used for limiting our analysis to 8 organisms are discussed in the section “Identifying phylogenetically conserved motifs”. For the E. coli analysis we selected 14 organisms from the Enterobacteriales order based on the same rules: Escherichia coli str. K-12 substr. MG1655, Citrobacter rodentium ICC168, Cronobacter sakazakii ATCC BAA-894, Dickeya dadantii 3937, Escherichia fergusonii ATCC 35469, Enterobacter aerogenes EA1509E, Erwinia pyrifoliae DSM 12163, Klebsiella pneumoniae CG43, Pantoea ananatis AJ13355, Pectobacterium wasabiae WPP163, Salmonella enterica subsp. enterica serovar Typhi str. CT18, Shigella dysenteriae Sd197, Vibrio cholerae O1 biovar El Tor str. N16961, Yersinia pestis KIM10+. Sequence information for the selected organisms was downloaded from NCBI.
Identifying orthologous genes between species. To identify orthologs shared between the selected organisms, we used orthoMCL version 2.0.2 [32]. The blastall function was run with the following parameters:-v 100000-b 100000-F ‘m S’-m 8-e 1e-5. All other functions were run with their default settings. Each of the identified orthologous groups (i.e., all orthologs of a given gene across species) was required to have an ortholog from the target species (E. coli or R. sphaeroides). For each orthologous group, the intergenic regions (IGRs) greater than 40bp in length, upstream of each gene in the group were then extracted from the appropriate organism, if they existed (genes within operons would generally not contain IGRs of sufficient length). As subsequent motif finding steps would require a sufficient number of sequences to identify meaningful motifs shared by the orthologs, we restricted the orthologous groups carried over to the motif finding step to those having at least 4 IGR sequences. A total of 2162 and 1326 groups of sequences met these criteria and were used for subsequent de novo motif detection, for E. coli and R. sphaeroides, respectively.
Identifying phylogenetically conserved motifs. These groups of intergenic sequences upstream of orthologous genes were used as input for de novo motif detection. Motif detection was conducted using MEME [41] with the following parameters:-dna-mod zoops-evt 0.01-nmotifs 3-maxw 30. A third order background distribution file was generated using all the intergenic sequences from all the organisms selected for each analysis and was used to aid subsequent motif detection. A total of 5144 and 914 phylogenetically conserved (PC) de novo motifs were detected from these sequences, for E. coli and R. sphaeroides, respectively. These were represented PSSMs (S2A Fig.). It should be noted that increasing the number of organism used in our phylogenetic footprinting analysis did not significantly increase the number of identified PC motifs for R. sphaeroides (S2B Fig.). This analysis also indicated as few as 6 organisms could be sufficient to carry out this analysis, if they possess the appropriate characteristics with respect to the target organism.
Clustering of identified motifs. These PC motifs identified in the phylogenetic footprinting step will contain a significant amount of redundancy, as multiple instances of essentially the same motif, corresponding to different binding sites of a specific transcription factor (TF), exist in this set. To eliminate this redundancy from the data set, we grouped identical or very similar motifs into clusters based on their similarity. To achieve this, we first conducted a pair-wise comparison of all identified PC motifs using Tomtom [41,42], generating q-values as measures of the similarity of these motifs to one another. Only motif pairs with q-values <0.01 were considered as potentially identical motifs and retained for subsequent clustering analysis. We then used MAST [41] to identify all the instances of each of the PC motifs (represented as PSSMs) across the target genome. The set of instances identified for a given motif were called “motif groups”. We then conducted a pair-wise comparison of all these motif groups to one another. Motif group pairs showing a high degree of overlap (based on identification of the same motif instances across the genome—threshold set to 33%) and for which the parent motif pairs had a q-value <0.01 were clustered into one group. These clustered motif groups, theoretically contain all the targets for a putative TF within the target genome. The identified target sequences were then used to generate species specific PSSMs based on all instances of each motif identified (see S1 Dataset and S3 Table for E. coli and R. sphaeroides, respectively). Based on these analyses, the 5144 and 914 PC motifs were clustered into 225 and 76 unique motifs for E. coli and R. sphaeroides respectively, based on their similarity.
Processing of gene expression data. Collecting all of the publicly available microarray datasets from R. sphaeroides from the gene expression omnibus (GEO platform GPL162) (totaling 174 microarrays) and combining these with unpublished microarray experiments conducted in our lab (totaling 24 microarrays), we generated a compendium of 198 microarrays encompassing experiments conducted under a variety of conditions (S1 Dataset), as well as a variety of gene deletion strains (ΔRpoHI, ΔRpoHII, ΔRpoHI/ ΔRpoHII, ΔFnrL, ΔPpsR, ΔPrrA, ΔRSP_4157 and ΔAppA) all constructed in the R. sphaeroides strain 2.4.1 background. All microarray analyses were conducted on the same Affymetrix platform, circumventing some of the data consistency and normalization issues that can arise when working with heterogeneous data from multiple platforms. All microarrays were normalized together using Robust Multichip Average (RMA) to log2 scale with background adjustment and quantile normalization [84]. The RMA normalized data were standardized by row normalization. The normalized R. sphaeroides gene expression dataset is provided in S1 Dataset. Normalized E. coli gene expression data was obtained from M3D (http://m3d.mssm.edu/).
Identifying clusters of co-regulated genes. Based on the phylogenetic footprinting analysis described above, we identified a total of 225 and 76 clusters of putatively co-regulated genes that shared conserved motifs for E. coli and R. sphaeroides, respectively. While these sequence-based networks were rich in information content about co-regulated genes and their putative shared cis-acting regulatory sequences, the information content of such networks could be improved by integration of gene expression data, as genes regulated by the same TFs are likely to have similar transcriptional profiles at least under a subset of conditions [8,12,37]. Hence, a gene which has a predicted shared motif with the other genes in a cluster but does not share a similar transcriptional profile with any other genes in that cluster, over at least a subset of experimental conditions, could be a false positive prediction, which could potentially be filtered out by using expression data. Furthermore, by utilizing bi-clustering algorithms that allow identification of subsets of conditions under which genes are co-expressed, one can potentially determine under what experimental conditions the genes of different clusters are active, providing an indication of their functional roles and/or signals to which they are responsive [12,37].
To integrate the data generated from phylogenetic footprinting with expression datasets, we utilized the data integration frame work DISTILLER [37]. DISTILLER takes in motif information as a binary file indicating whether a particular de novo detected motif is present or not. It also takes an expression matrix of normalized expression data across conditions. It then uses an itemset data mining approach to predict what conditions genes sharing a common motif, show correlated expression patterns. We ran DISTILLER on our data sets using the following parameters: binary supports: 1, box supports: 30, box p-values: 0.001, number of randomizations: 100000, size of random modules: 4, minimal module size: 3, number of greedy modules: 400.
The integration of expression data into our predictions resulted in the removal of a subset of genes from the original sequence-based clusters due to an inability to identify sub-conditions under which they are co-expressed with other members of the cluster. For instance, in the case of target genes predicted for cluster 60 in the R. sphaeroides dataset, eighteen genes were predicted to be members of this cluster based on our phylogenetic footprinting analysis (S7 Fig.), while only 13 of these genes showed strong co-expression with other members of the operon, under at least a subset of conditions. The genes not showing strong co-expression were thus removed from the cluster. Subsequent experimental analysis of the predicted transcriptional regulator for this bi-cluster (PpsR, see Results) verified that these excluded genes were likely false positive predictions from our phylogenetic footprinting step. Thus, at least in subset of instances, integration of our gene expression data sets using DISTILLER appeared to improve the overall accuracy of our TRN.
Operon extension. While phylogenetic footprinting analysis enabled us to identify putative binding sites for TFs and thus identify the closest gene(s) to the binding site, other genes within close proximity of this target gene, and potentially in an operon with it, were not captured in the initial analysis. To incorporate operon structure into our predictions, we combined distance-based operon predictions from microbes online [85], with correlation data from the microarray datasets. Genes predicted to be in an operon based on distance and had a Pearson’s correlation coefficient of at least 0.8 across the entirety of our microarray compendium, were considered to be in an operon. This information was used to extend to predictions in our TRN to take into account genes that might be in an operon with targets identified via our sequence-based analysis.
Prediction of transcriptional regulators for clusters. Having identified and refined our clusters of co-regulated genes using sequence and gene expression information, we then predicted the most likely of the known or predicted TFs in our target organisms to regulate each of these clusters. To achieve this we used a combination of 4 criteria based on known properties of bacterial TRNs (S3 Fig.). They consisted of:
Correlation between a TF and its target genes [3,6–8]
Proximity of a TF to the location of its closest binding site in the genome [12,14,18,19].
Similarity in DNA motifs bounds by TFs having similar DNA binding domains (DBD) [19,20].
Phylogenetic correlation of the occurrence of a TF and occurrence of a motif across species [19].
Implementation details of these analyses are provided in the Result and Discussion. The top 3 highest scoring TFs for the R. sphaeroides network presented in S1 Table are provided in S10 Table, while those for the E. coli network are provided as part of S1 Dataset.
Inferring regulatory interactions solely from expression data. Recent analysis has shown that combining the predictions from a small number of high performing expression-based TRN inference approaches can result in significantly improved prediction accuracy [3]. Thus, to make predictions for TFs not captured in the comparative genomics-based TRN model for R. sphaeroides, we employed a combination of expression-based TRN inference approaches to try to identify regulatory interactions using only our microarray datasets. For this analysis, we combined the predictions from 3 well-established, high performing direct inference approaches: context likelihood of relatedness (CLR) [9], GENIE3 [10] and ANOVA-based approach [11]. As these approaches have previously been described [3], thus we do not provide any details of implementation or assumptions peculiar to each approach.
Our RMA normalized and row standardized gene expression data from 198 microarray experiments for R. sphaeroides were used as input data for these 3 inference approaches. A list of 216 R. sphaeroides TFs was also provided as potential transcriptional regulators (S1 Dataset). In addition, information on specific deleted or over-expressed genes was provided as additional input for ANOVA. The top 50,000 predicted TF-target interactions from each approach were selected. For each inference approach, the scores of TF-target predictions were converted to p-values by random permutation, generating 10000 random TF-target scores for each approach and comparing actual TF-target scores to this set. To determine the likelihood that TF i regulates target gene j, the predictions from each of the 3 approaches for that specific interaction were then combined by averaging the—log10 of the p-values from each approach (eqn. 8). In instances where no prediction was made for a particular TF-target interaction by any one approach, but predicted by at least one of the other 2 approaches, a score of 0 was assigned for that approach. Potential TF-target interactions not in the top 50,000 of any of the 3 prediction lists were not considered.
Predicted targets for each TF were then extended to include all genes in a potential operon, as described above, to generate the expression-based TRN.
To determine a score threshold to use as a cut-off for interactions to be retained in the R. sphaeroides expression-based network, we collated all previously generated genome-wide protein-DNA interaction (ChIP) datasets for R. sphaeroides and used this to generate a precision-recall curve for the network (S8 Fig.). Genome-wide protein-DNA interaction data for FnrL [53], σE [86], RpoHI and RpoHII [87], corresponding to a total of 467 TF-target interactions, were used for this analysis. We used these interactions as our set of true positives (TP). Precision-recall curves were generated for ranked lists of predictions from CLR, ANOVA, GENIE3 and the combined network (S8 Fig.), with precision and recall calculated at intervals of 100 predictions. Typically precision is calculated as:
TPPTPP+FPP=True positive predictionsTrue positive predictions + False positive predictions=True positive predictionsAll predictions made
(9)
where TPP and FPP are assessed based on the number of interactions considered and a gold standard of true positives interactions (TP) [43]. However, due to the small number of TP available for assessment, for each interval of 100 predictions from the R. sphaeroides expression-based TRNs, we only considered predictions for TFs for which we had ChIP data and thus we redefined precision as follows:
TPPTPP+FPP=TPP for TFs with ChIP daTaTPP for TFs with ChIP daTa+FPP for TFs with ChIP daTa=TPP for TFs with ChIP daTaAll predictions made for TFs with ChIP daTa
(10)
Recall was calculated as previously described [43]:
TPPTP=True positive predictionsAll known true positives
(11)
We selected a precision cut off of 95%, which corresponded to a recall of 8% and a score cut-off of 1.3 (S8 Fig.). Using this cut off for the entire set of predicted interactions resulted in a total of 1100 predicted TF-target interactions. In this analysis the best performing of the individual approach selected was GENIE3, whose performance was very close to the final composite TRN, though the predictions retained the final composite network differed significantly from the predictions of any one of the individual networks (S9 Fig.), as predictions supported by at least 2 of the 3 approaches received higher scores.
Combining sequence-based and expression-based networks. To leverage the potential complementarity between the reconstructed R. sphaeroides sequence- and expression-based networks, we merged predictions from the 2 networks giving precedence to predictions made using the comparative genomics-based integrative approach, as these predictions were supported by both sequence and expression data. Thus, for TFs for which predictions were made in both our comparative genomics- and expression-based networks, only the predictions from the comparative genomics-base network were retained in our final combined network. Based on this a total of 641 TF-target interactions from our expression-based analysis were retained in the final combined network. This included a total of 44 TFs. The final set of interactions predicted using expression-based approaches is provided in S11 Table.
Experimental analysis. Details of growth conditions, construction of mutants, microarray and ChIP-seq analyses are provided in S2 Text and S12 Table. All microarray and ChIP-seq datasets generated in this study were deposited in GEO under the accession GSE58658.
Data and software. The software required to run the workflow is written in JAVA and provided as part of S1 Dataset. In addition, the code together with files for an example run are available at http://dx.doi.org/10.6084/m9.figshare.1249869.
|
10.1371/journal.pcbi.1002674 | Molecular Dynamics Simulations Reveal Proton Transfer Pathways in Cytochrome C-Dependent Nitric Oxide Reductase | Nitric oxide reductases (NORs) are membrane proteins that catalyze the reduction of nitric oxide (NO) to nitrous oxide (N2O), which is a critical step of the nitrate respiration process in denitrifying bacteria. Using the recently determined first crystal structure of the cytochrome c-dependent NOR (cNOR) [Hino T, Matsumoto Y, Nagano S, Sugimoto H, Fukumori Y, et al. (2010) Structural basis of biological N2O generation by bacterial nitric oxide reductase. Science 330: 1666–70.], we performed extensive all-atom molecular dynamics (MD) simulations of cNOR within an explicit membrane/solvent environment to fully characterize water distribution and dynamics as well as hydrogen-bonded networks inside the protein, yielding the atomic details of functionally important proton channels. Simulations reveal two possible proton transfer pathways leading from the periplasm to the active site, while no pathways from the cytoplasmic side were found, consistently with the experimental observations that cNOR is not a proton pump. One of the pathways, which was newly identified in the MD simulation, is blocked in the crystal structure and requires small structural rearrangements to allow for water channel formation. That pathway is equivalent to the functional periplasmic cavity postulated in cbb3 oxidase, which illustrates that the two enzymes share some elements of the proton transfer mechanisms and confirms a close evolutionary relation between NORs and C-type oxidases. Several mechanisms of the critical proton transfer steps near the catalytic center are proposed.
| Denitrification is an anaerobic process performed by several bacteria as an alternative to aerobic respiration. A key intermediate step is catalyzed by the nitric oxide reductase (NOR) enzyme, which is situated in the cytoplasmic membrane. Proton delivery to the catalytic site inside NOR is an important part of its functioning. In this work we use molecular dynamics simulations to describe water distribution and to identify proton transfer pathways in cNOR. Our results reveal two channels from the periplasmic side of the membrane and none from the cytoplasmic side, indicating that cNOR is not a proton pump. It is our hope that these results will provide a basis for further experimental and computational studies aimed to understand details of the NOR mechanism. Furthermore, this work sheds light on the molecular evolution of respiratory enzymes.
| Bacterial denitrification is one of the examples of anaerobic respiration in which nitrate (NO3−) is stepwisely reduced to dinitrogen (N2) [1]–[3]. During denitrification, the key intermediate step of the reduction of nitric oxide (NO) to nitrous oxide (N2O) is catalyzed by a membrane-bound enzyme nitric oxide reductase (NOR) according to the following scheme: 2NO+2e−+2H+→N2O+H2O. Bacterial NORs perform fundamental chemistry and are the largest source of N2O, a greenhouse gas and an ozone-depleting substance, released into the atmosphere [1]. This enzyme also has an important role in the evolution of the respiratory system. NOR belongs to the superfamily of O2-reducing heme-copper oxidases (HCOs) and is believed to be evolutionary linked to a proton pump cytochrome c oxidase (CcO). Both enzymes may have evolved from a common ancestor [2]. The ancestral oxidase was probably involved in NO reduction, but later switched to oxygen reduction and additionally acquired the ability of proton pumping, although this issue is still open to debate [3]–[9].
After the structure of CcO was solved more than a decade ago [10], [11], that system became the focus of numerous experimental studies, which produced a number of X-ray structures from different organisms and a wealth of mutational, biochemical and spectroscopic data, as well as theoretical and simulation ones (for recent reviews, see refs. [12]–[14]). In contrast, the information about NORs was limited, but the first structure of cytochrome c-dependent NOR (cNOR) from Ps. aeruginosa has been recently determined by Shiro and co-workers [15], and that provides a basis for studies aimed at describing the mechanism of NO reduction at the atomic level. cNOR consists of two subunits, NorB and NorC, and contains four redox active metal centers, namely hemes b, b3 and c and a non-heme iron (FeB). The latter and the iron of heme b3 form the binuclear (BN) center, a site of the NO reduction. The crystal structure revealed that FeB has three His and one Glu ligands and that a tightly bound Ca2+ ion is bridging hemes b and b3. Although the function of Ca2+ is not yet fully clear, it is interesting to note that it has the same binding position as in a recently determined structure of the microaerobic respiratory enzyme cbb3 oxidase [16], which is a C-type HCO able to reduce NO to N2O in low-oxygen conditions [17], [18].
For the NO reduction reaction protons have to be delivered to the BN center, which is buried inside membrane. Previous experiments with the whole-cell [19] and liposome-reconstituted [20], [21] cNORs demonstrated that protons utilized in the catalytic reaction are taken up (on a ms timescale) from the periplasmic side (i.e. the same side as electrons), which suggests that the NO reduction reaction is non-electrogenic and therefore cNOR is not a proton pump. In order to explain the functioning of cNOR, it is necessary to understand the detailed mechanism of the proton delivery to the BN center. Since proton transfer (PT) can occur efficiently only when the donor and acceptor groups are immediately close to each other, the long-distance proton translocations in proteins (e.g. proton pumping across the membrane or proton delivery from the bulk to the buried active site) require specialized proton-conducting pathways, which involve protein ionizable groups and intermediate water molecules as proton-binding sites (see e.g. refs. [22]–[28]).
Analysis of the cNOR crystal structure yielded two independent H-bonded networks, designated as Channels 1 and 2, which are formed by the resolved water molecules and the charged/polar residues [15]. These channels were proposed as potential PT pathways. However, since X-ray crystallography provides only static snapshots of the protein structure, which are averaged over many unit cells, in general such structures even at high resolution show only a few water molecules (at the most stable positions) inside the protein but miss many dynamic ones. The proposed proton pathways did not provide a continuous connection from the surface to the active site (i.e. protonic “gaps” were present), and in particular the pathways near the catalytic center, where no water molecules were resolved, remained elusive. As we mentioned, in such situations the connectivity is expected to come from the intervening water molecules. Thus, water in cNOR could play a very important role in the enzyme function and has to be fully characterized.
Molecular dynamics (MD) simulations of membrane proteins within an explicit membrane/solvent environment (for some recent works and reviews see refs. [28]–[35]) can provide important information about the water dynamics, such as a “real” level of hydration and specific water positions inside the protein, and are most valuable in the cases when the water/PT channels are not yet described at the atomic level. For example, MD simulations have been recently used to explore the water dynamics in different regions in CcO and greatly contributed to the understanding of the details of PT channels in that enzyme [36]–[42]. We note that a study with an explicit membrane/solvent by Olkhova et al. [39] suggested a large number of water molecules within the PT channels in CcO, in contrast to simulations which utilized different kinds of reduced models or truncated systems.
In this work we performed MD simulations of cNOR. We focused on the water dynamics, with the aim to identify the water channels and H-bonded networks that could serve as pathways for the proton delivery to the active site. The obtained information will be important for further elucidation of the mechanisms of the proton translocation and NO reduction in cNOR.
We performed an all-atom MD simulation of cNOR in the explicit lipid/water environment (Figure 1a). The details of the system setup and simulation and analysis protocols are provided in the Text S1. Briefly, the initial system was prepared from a 2.7 Å resolution crystal structure of the cNOR from Ps. aeruginosa (PDB ID 3O0R) [15]. A simulation system is shown on Figure 1a: cNOR was embedded into the pre-equilibrated POPE (palmitoyl-oleoyl-phosphatidylethanolamine) lipid bilayer membrane and a solvent box of water molecules. The total size of the simulation system was ∼110,000 atoms. The main purpose of introducing the lipid bilayer in MD simulation is to model cNOR in situ, i.e. in the environment as close to its natural as possible. POPE is the major lipid component of bacterial membranes [43]. Explicit membrane provides additional stability to the protein in MD simulations and allows a correct description of the protein-solvent and protein-lipids interactions.
MD simulations were carried out in NAMD [44] with the CHARMM force field [45], [46]. After minimization and equilibration parts, production runs were performed at a constant temperature, pressure, and surface area (NPAT ensemble) for 300 ns, providing reasonable conformational sampling of the protein. Stability of the simulated protein-membrane complex was assessed from the analysis of several parameters along the MD trajectory (Figure S1). The root-mean-square deviation (RMSD) of the helical Cα atoms is below 2 Å while RMSD of Cα atoms in a transmembrane (TM) region is ∼1.2 Å. The RMS fluctuations (RMSFs) calculated for each residue also illustrate that the TM region is very stable while the outer and inner domains exhibit, as expected, larger motions. Finally, the area/lipid, which was calculated using the Voronoi analysis tool [47], remains close to the experimental value for the POPE lipids [48], also indicating a stable simulation of the protein-membrane complex.
One of the proposed PT pathways (Channel 1 in Figure 3 in ref. [15]) goes through a large hydrophilic region, which is located on the periplasmic side of the enzyme at the interface of a TM region (NorB subunit) and an outer soluble domain (NorC subunit) (Figure 1). Four residues, namely Glu135, Asp198, Lys53c, and Glu57c, were designated as Channel 1 [15]. [Subscript “c” indicates residues of the NorC subunit, while residues of the NorB subunit are numbered without additional subscripts.] In MD simulations we observed that Channel 1 indeed connects the protein surface to propionates of heme b3 (the distance ∼16 Å) via a number of ionizable residues and water molecules and supports formation of the H-bonded networks (see below). Our analysis provides important additional details about Channel 1 (Figure 2a). MD results indicate that the following three residues participate in the HB networks in that region: Arg134, Lys199, and Glu70c, and therefore they have to be included in Channel 1. All seven ionizable residues are highly conserved in cNORs. Together, they line up a large hydrophilic channel, and their sidechains assist in the formation of the H-bonded water chains.
Channel 1 has a connection to the bulk solvent between two helices, TM VI of NorB (with Asp198 and Lys199) and α2 of NorC (with Lys53c and Glu57c). The entrance site formed by the amino acids Glu57c, Lys53c, and Arg134 (Figures 2a and S2) remains rigid due to three stable salt bridges: Glu57c-Lys53c, Arg134-Asp198, and Lys53c-Asp198 (Figure S3). These residues partially block water influx. However, water molecules still occasionally cross through that site, and thus can serve as intermediate proton sites (Figure S4). Also, the dynamic HB networks involving sidechains of Glu57c and Asp198 and waters at both sides of the entrance, i.e. in the bulk and inside the Channel 1 cavity, are observed at any time of MD trajectory. Therefore it is possible that one of these residues could be directly involved in PT by picking up protons from the bulk and releasing them to the water chain inside Channel 1. The mutagenesis experiments with P. denitrificans cNOR [Pia Ädelroth, unpublished data] showed the importance of Asp185 (equivalent to Asp198 in Ps. aeuginosa cNOR) for the enzymatic activity and proton uptake, and provide partial support to this proposal.
From the entrance region the proton pathway proceeds further through the dynamic water chains. Water channels in cNOR have “irregular” shape and lack simple symmetry (like, e.g. straight TM channels in aquaporins or ion channels). Therefore to perform meaningful statistical analysis in each region we selected water molecules within a reasonable distance cutoff (typically 4.5 Å) near the sidechains of the pathway's residues. We verified that with such definition water molecules “inside the pathway” were not skipped. Our calculations show that Channel 1 is very well hydrated: in MD simulation ∼20 water molecules are observed in this hydrophilic region (Figure 2c), which is higher than ∼12 molecules resolved in the X-ray structure. This result can be explained by the presence of mobile water molecules, which were not resolved in X-ray crystallography.
To provide some quantitative representation, we have calculated volume occupied by water molecules during simulation (“water density”, see e.g. refs. [49], [50]) in different regions of cNOR. Figure 2 illustrates water spatial distribution in Channel 1 as obtained from MD simulation, showing both a 3D water volume map (Figure 2a, isosurface at 25% occupancy) and a 2D contour plot (Figure 2b, an XY-plane projection of the water density; see figure caption for details). A few observations can be made from these figures. (i) Water density representation shows the extent of the hydrophilic regions and confirms a stable connection from the bulk to the active site heme. (ii) Water molecules form an extensive water cluster between propionates of heme b3 (PropA and PropD). [Please note that compared to the previously reported cytochrome c oxidase structures the active site heme in cNOR, i.e. heme b3, is flipped and the order of the propionate groups A and D is different.] (iii) After the entrance region, the channel goes into a water-filled cavity. An important finding is that further the pathway splits into two branches: one path leads via a water chain (5–6 water molecules) directly to PropA, while another – via Glu70c and a short chain (2–3 water molecules) at the other side of that residue – to PropD. The terminal region of both paths is the water cluster near heme b3 propionates. The existence of two branches in Channel 1 could be observed in the MD simulation, but was not evident from the static X-ray structure. This feature provides a possibility of PT over different pathways and probably adds to the robustness of the proton uptake via Channel 1. (iv) When water density is plotted at a higher occupancy level, one obtains positions of the water sites that are occupied almost permanently during the simulation. One example is a crystallographic water molecule, Wat65, which remains bound near Ca2+ for the entire MD trajectory. Such “permanent” water sites in general superimpose well with the positions of waters resolved in the X-ray structure (indicated by purple spheres on Figure 2b).
Figure 2a also presents a typical configuration of the H-bonded networks forming in Channel 1, while Video S1 and multiple MD snapshots on Figure S5 illustrate their time-dependent dynamics. The average lifetime of a hydrogen bond (HB) in the water chains is in the ps range due to rotating and/or moving water molecules. It can be seen that water molecules in Channel 1 have high mobility and exchange rates and, as a result, the forming H-bonded networks are constantly “fluctuating” (similar to the H-bonded networks in CcO [39], [40]). Continuous HB paths between the bulk and heme b3 propionates do form, and their consistency is limited by the intervening water chains, namely a chain from Asp198 to Glu135 (probability ∼25–35%) and a chain from Asp198 to Glu70c (probability >60%). But it is important that such connections are forming at all times, and thus can assist efficient proton translocation [51]. Participation of the Channel 1 residues in the H-bonded networks can be assessed quantitatively by calculating the number of surrounding water molecules and formed HBs (Table S1). In particular, these results, in addition to visual analysis, suggest that Glu70c could play an important role in the proton uptake process. It is desirable to verify its involvement in the PT pathway by site-directed mutagenesis experiments.
From the inspection of the X-ray structure, Hino et al. identified another cavity, which contains many crystallographic waters, and proposed it as a second possible proton-conducting pathway (Channel 2 in Figure 3 in ref. [15]). In MD simulation we observed a large hydrophilic region formed by the residues Arg416, Thr66c, Glu77c, Gln411, and Gln415, with an exit to the bulk beyond the latter (Figure 3a). On average, there are 10 to 12 water molecules in the cavity. However, we found that water molecules from this cavity cannot pass to the water cluster near heme b3, and further to the active site. Two loops, and more specifically two glycine residues Gly340 and Gly69c, are in close contact en route to heme b3 and, together with a ring of Tyr73c, disrupt a possible water chain. A close steric contact between two loops remains for the entire length of the MD trajectory, as evidenced by the Gly-Gly distance (Figure 3c), which stays around 3.5–4 Å (i.e. similar to the distance in the crystal structure). Water densities (Figures 3a and b) clearly show a wide gap with no substantial density between the upper hydrophilic cavity and the water cluster. We do not completely rule out a possibility that mobile water molecules can occasionally cross the gap region; however, no such crossings or continuous HB networks were observed in 300 ns. Moreover, the proton translocation through the region with no polar/charged residues or water molecules would encounter high activation barriers. Thus, our results do not support the previously proposed Channel 2 as an alternative pathway for proton delivery to the active site. The exact functional role of this hydrophilic cavity in cNOR is not clear.
A careful analysis of MD trajectories revealed another plausible proton pathway, which we designated as the (periplasmic) Channel 3 (see Figures 1 and 4). This pathway involves the residues Glu135, Glu138, Arg57, Asn54, and Asn60c. The first three are highly conserved in all NORs and oxidases, while Asn54 is conserved in cNORs. In the X-ray structure, three water molecules are resolved in a cavity formed by these residues. During the initial part of the MD simulation this region has no connection to the periplasmic surface (Figures 4a, b). The calculated water density clearly shows that the cavity is completely separated from the bulk solvent and that two asparagine residues, Asn54 and Asn60c, effectively work as a gate, blocking water access from the outside.
However, after ∼165 ns in the MD simulation the Asn54-Asn60c gate opens and a new water channel is formed (Figures 4c, d). A continuous water density then extends up to two important residues, Glu135 and Glu138, and the H-bonded networks involving mobile water molecules and amino acid sidechains readily form. The number of waters in the hydrophilic region, and in particular around the sidechain of Glu138, significantly increases with the gate opening and remains high even after the gate closes back (Figure 5c). Figure 5 also shows minimal distances between the Asn54-Asn60c and Glu138-Asn60c pairs in the MD simulation, along with the representative snapshots of the gate region. Clearly, the gate is closed when two Asn are H-bonded. Sidechains of Glu138 and Asn60c exhibit large-amplitude rotations (Figure S6), in particular Glu138 can take several conformations, and the initial event leading to the gate opening seems to be a rotation of Glu138 to the “up” position after ∼135 ns and the formation of a HB to Asn60c. Soon after that a strong HB between two Asn is broken. As a result, a helix TM II (with Asn54 on a top) slightly tilts away, and that opens water access to the internal cavity.
An overlay of the open and closed configurations (Figure 5f) shows that the required structural changes are rather small: two Asn move away only by a few Å, but that is enough to break a HB between them and to open access to the internal cavity for water molecules from the outside. The gate is open for ∼60 ns, after which the HB between Asn60c and Asn54 is re-formed; the HB between Glu138 and Asn60c breaks prior to that. The explicit gate opening/closing process and formation of the dynamic water chains in Channel 3 are illustrated by Video S2. We would like to emphasize that similar events were also observed in the extended simulation as well as in independent runs (Figure S7), indicating that such events can occur in cNOR on a 100-ns timescale, which is much shorter than the experimentally measured rate of the proton uptake (∼25 ms) [20], [21]. This suggests that such structural reorganizations due to protein fluctuations are feasible during catalysis in cNOR and that Channel 3, in principle, can provide a pathway for a water-mediated proton uptake. We propose to examine the role of Asn54 and Asn60c in the Channel 3 gating by the mutagenesis experiments.
A newly found channel is consistent with previous experimental data. Two key residues, Glu135 and Glu138 (Glu122 and Glu125 in P. denitrificans cNOR), were shown by site-directed mutagenesis to be essential for the enzymatic activity and were proposed to be a part of the proton input pathway [52]–[54], though their exact positions predicted with the homology-based model (namely, on a protein outer surface) [21] turned out to be incorrect. With the cNOR structure available now, it is known that Glu135 is a ligand to Ca2+. That explains why its substitution with Asp still showed a level of activity close to the wild type (i.e. Ca2+ coordination was kept) while a substitution with Ala or Gln resulted in a loss of activity (most likely caused by a Ca2+ dissociation). The structural function of Glu135 also makes its direct participation in PT problematic: it is unlikely that Glu135 can get protonated or that the protons coming from the periplasm can be transferred through a densely packed region occupied by the Ca2+ ion and its ligands.
The substitution of Glu138 with Ala and Asp resulted in a loss of activity, while a mutation to Gln showed some, though significantly reduced activity [53], [54]. These results could indicate that the length of the sidechain is more important than retaining a negative carboxylic group at that position. The observation fits into the above suggested mechanism of the Channel 3 opening and “activation” of the proton pathway, which includes a Glu138 sidechain rotation to the “up” position to form a HB to Asn60c, thus helping to break a HB between two Asn.
In contrast to Glu135, Glu138 can actively participate in the PT process. A Glu122Asp mutation in P. denitrificans caused a significant pKa shift of a presumed nearby proton donor group [54], and Glu138 seems to be the best candidate for that role. The proton pathway beyond Glu138 is also offered by our MD results. After the gate opening, Glu138 is well hydrated, with typically 5 to 8 water molecules near its sidechain (Figures 4c,d and 5c). We observed the H-bonded water chains leading from this site toward the water molecules bound near BN, thus avoiding the Ca2+ site (see the corresponding discussion below).
A key finding is that the suggested novel channel in cNOR is equivalent to the putative PT pathway (the “periplasmic cavity”) in a recently determined structure of cbb3 oxidase [16]: a comparison of two regions shows that their positions are identical (Figure 6). Moreover, the important residues which form this hydrophilic cavity, namely Glu135, Glu138 and Arg57 in cNOR and Glu122, Glu125, Arg57 in cbb3, are conserved. The periplasmic cavity in cbb3 oxidase was suggested to be an exit pathway of the pumped protons or a pathway for proton uptake from periplasm when the enzyme is involved in NO reduction [16]. The fact that for NO reduction cbb3 uses protons from the periplasmic side of the membrane has been recently confirmed by the experimental work of Lee et al. [18]. We note that such cavity is not found in other structurally known HCOs and that aa3 oxidases (A-type HCOs) are incapable of NO reduction, while ba3 oxidases (B-type HCOs) can reduce NO but much slower than cbb3 [3], [5], [9]. The presence of a plausible PT pathway in the equivalent region in cbb3 oxidase is an additional argument for the functional importance of Channel 3 in cNOR. The finding that two enzymes likely have common elements of the PT mechanism, along with other common structural factors, such as the identical position of Ca2+, fits nicely into the phylogenetic pictures that draw C-type HCOs as the closest evolutionary relatives of NORs.
We have also analyzed the region equivalent to Channel 1 in the cbb3 structure [16]. It seems that the corresponding region cannot provide a pathway for proton translocation in cbb3 because: (i) some of the charged residues present in Channel 1 in cNOR, namely Lys199, Lys53c, Glu57c, and Glu70c, are either missing or located far away in cbb3, (ii) a coil with several hydrophobic residues is located in the central part of that region and splits it into two parts; the water distribution is disconnected too [to be published], (iii) a second Ca2+ site is located at the position equivalent to the entrance to Channel 1 in cNOR and most likely blocks proton transfer.
We have shown that Channel 1 and Channel 3 can connect the periplasmic surface to the region near heme b3. Its propionates together with a nearby water cluster and Glu138 are the likely intermediate proton acceptor groups. (It is less likely that PropA can get protonated since it serves as a ligand to Ca2+.) It is worth mentioning that in CcO one of the active site heme propionates is thought to be the likely proton loading site for the pumped protons [55]–[57]. The idea about the functional importance of protonated water clusters inside proteins is also not new. For example, in CcO a protonated water cluster was suggested as a proton storage site in the D-channel [58], while in bacteriorhodopsin a protonated water cluster is a presumed proton release group [59], [60].
An important question is how protons are delivered to the catalytic center when they are needed for the NO reduction, i.e. what are the structural elements critical for the final PT steps? The distance (>8 Å) is still long for direct PT, but no water molecules were resolved in the vicinity of the BN center. So the further proton path was not clear from the X-ray structure, and intermediate water molecules are expected to play important role. In a working enzyme, water will be produced at the active site as a byproduct of the catalytic NO reduction. In contrast to the crystal structure, the MD simulation reveals the presence of water molecules near the BN center (Figure S8) and describes their distribution (Figure 7). The exchange rate of waters is much lower compared to the channels discussed above. Water molecules are found persistently at several positions and keep these positions for 20–50 ns or longer (Figure S9); such water molecules might serve as intermediate proton sites. Figure 7 depicts a representative configuration of water molecules in that region, along with the calculated water density (see also Figure S10). It can be seen that one permanent water site is located between two irons of the BN center (i.e. where NO ligands will bind during the enzymatic cycle), another corresponds to the water molecule bound between FeB and Glu280, and two more water sites are located between FeB and PropA. It is interesting that in a recent high-resolution structure of Th. thermophilus ba3 oxidase [61] two water molecules were resolved at the identical positions.
Analysis of the water dynamics and distribution offers several possible paths for the final PT steps to the BN center (Figure 7):
High-resolution crystal structures of CcO and subsequent mutational studies identified a number of critical residues in the proton pathways from the cytoplasm to the active site (K and D channels). However, in cNOR most of these residues are replaced by hydrophobic residues. The crystal structure of cNOR neither provides an obvious water channel from the cytoplasmic side of the membrane nor a H-bonded network in the regions that correspond to the K and D proton channels in CcO (see Figure 4 in ref. [15]). Similarly, our MD simulation shows no water in those regions (Figure 8), with the exception of a hydrophilic cavity below the active site with three glutamates, Glu211, Glu280, and Glu215. Thus, in cNOR there is no proton pathway from the cytoplasmic side. This is consistent with the experimental observations that cNOR is not electrogenic and has no proton-pumping activity, and that the electrons and protons for the catalytic reaction are supplied from the periplasmic side. The position of the above-mentioned small hydrophilic region overlaps with the terminal part of the K-pathway in cytochrome oxidases. That could indicate a beginning of the K-channel formation in the evolutionary steps leading to the appearance of proton pathways from the cytoplasm and eventually to the proton pumping in other HCOs. A very recent structural characterization of a single-subunit quinol-dependent NOR (qNOR) from G. stearothermophilus [64] surprisingly revealed the existence of the water channel from the cytoplasmic side at the position equivalent to the canonical K-pathway and absence of the periplasmic pathways found in cNOR. It will be interesting to test by calculations if a similar cytoplasmic channel can be formed in cNOR as the result of selective mutations.
We have performed a 300 ns MD simulation of cNOR, based on its first crystal structure, and fully characterized water inside the protein. Our simulations have revealed two potential PT pathways from the periplasmic side, Channels 1 and 3. Both pathways are supported by the continuous distribution of water molecules and formation of the dynamic H-bonded networks within the channels, as well as by the highly conserved nature of the participating residues and previous experiments, which had shown functional importance of some of these residues. Since cNOR is not involved in a vectorial proton translocation (pumping against the gradient), a robust gating mechanism, as those suggested in CcO [56], [57], [62], [65], [66], is not required, and chemical protons have to arrive at the active site in one way or another. So, in principle, both pathways may be used. From our MD results we cannot unambiguously establish what the exact role of each channel is or how they are synchronized. In our opinion, Channel 1 is probably the main pathway for the proton uptake since both static and dynamic structures clearly show extensive H-bonded networks and water chains, and the path toward the catalytic site seems to be more straightforward. Meanwhile, Channel 3 is revealed only by the dynamic simulations (and the water channel is formed only for a part of the simulation), some protein structural rearrangements are required there to allow for channel formation, and the path from Glu138 to the active site goes through an intermediate hydrophobic region.
A further discussion about the details of the proton uptake mechanism in cNOR should be based on additional experimental evidences and explicit PT calculations. We would like to emphasize that MD simulations provide important information about the dynamics of water molecules and H-bonded networks and, as a result, about locations of potential proton pathways. However, classical MD simulations alone cannot describe explicit proton translocation, which is an intrinsically quantum mechanical process. The energetics of PT along different pathways has to be addressed by mixed QM/MM methods [25], [56], [67]–[69], and this will tell whether each pathway is feasible. The key issues in such calculations are the energies of charge formation at different sites along the translocation path and activation barriers of individual PT steps.
In our calculations we observed a fairly high number of mobile water molecules (which could not be resolved in the X-ray structure) in the cNOR hydrophilic cavities. Similar results were previously reported in analogous MD studies (with explicit membrane/solvent, at ambient temperatures) of systems like proton pumps cytochrome c oxidase [39], bacteriorhodopsin [28], [70], bc1 [71], voltage-gated proton channel Hv1 [27], [72] and calcium pump [73], [74], whose function relies on the water-assisted proton translocation. Therefore such simulations, although they are computationally expensive, can be used for the detailed characterization of water inside membrane proteins and for the identification of potential proton pathways, which in many cases are critical for protein function.
Finally, several common structural features, namely the position of the Ca2+ binding site and similarity of Channel 3 in cNOR and the periplasmic cavity in cbb3 oxidase, indicate the evolutionary relationship between the two enzymes. The likely loss of Channel 1 in cbb3 oxidase might be the key step during the molecular evolution leading to the establishment of the PT pathway from the cytoplasm, while a less effective Channel 3 was probably kept as a proton exit pathway for proton pumping. Our results have implications on the development of PT pathways in HCOs and the evolution of respiratory enzymes in general – a topic which remains a subject of intense debate.
|
10.1371/journal.pntd.0007075 | Community knowledge, attitude, and perceived stigma of leprosy amongst community members living in Dhanusha and Parsa districts of Southern Central Nepal | Though Nepal declared leprosy elimination in 2010, its burden is constantly rising in Terai communities for the past 2 years with 3000 new leprosy cases being diagnosed annually. Community’s perception is important for prevention and control of leprosy and enhancing quality of life of leprosy patients. Poor knowledge, unfavorable attitude and stigma create a hindrance to leprosy control. The main objective of this study was to assess the knowledge, attitude and stigma of leprosy amongst the community members living in Dhanusha and Parsa districts of Southern Central Nepal.
A total of 423 individuals were interviewed using a structured questionnaire in Dhanusha and Parsa districts. Data was analyzed using both descriptive (frequency, percentage, median) and statistical inferences (Chi-square test, Kruskal Wallis H test, Mann Whitney U test, binary logistic regression) using SPSSvs20.
All respondents had heard about leprosy. Source of information on leprosy was mainly found to be health workers/hospitals (33.1%). Only 62.6% reported bacteria being its cause followed by other myths such as bad blood/curse/heredity/bad deeds (36%). Only 43.8% responded that leprosy is transmitted by prolonged close contact with leprosy patients and 25.7% reported religious rituals as the treatment. Only 42.1% had good knowledge and 40.9% had favorable attitude. Good knowledge of leprosy was highly associated with favorable attitude towards leprosy (P<0.001). The outcome variables- knowledge, attitude and EMIC score were found to have highly significant association with age, sex, ethnicity, religion, education and occupation of the respondents (P<0.001). Having knowledge on leprosy transmission was positively associated with favorable attitude towards leprosy (P<0.001).
Strategizing the awareness programmes according to socio-demographic characteristics for enhancing the knowledge regarding leprosy cause, symptoms, transmission, prevention and treatment, can foster the positive community attitude towards leprosy affected persons. Enhancing positive attitude towards leprosy affected persons can reduce the community stigma, thus may increase their participation in the community. Positive attitude may further increase their early health seeking behaviour including their quality of life.
| Though Nepal declared leprosy to be no more a public health problem in 2010, its burden is constantly rising in Terai communities for the past 2 years with 3000 new leprosy cases being identified annually. With the fact that community’s knowledge and perception is important for prevention and control of leprosy this study aimed at assessing the community knowledge, attitude and stigma of leprosy amongst the community members living in Dhanusha and Parsa districts of Southern Central Nepal. The study was conducted in the communities of Dhanusha and Parsa by interviewing 423 individuals using structured questionnaire. All study respondents had heard about leprosy with main source of information to be health workers/hospital. A good proportion had myths such as bad blood/curse/heredity/bad deeds as the cause of leprosy and reported religious rituals as its treatment. Although more than half had good knowledge, only 2/5th had favorable attitude. The attitude was found to be influenced by knowledge. Also, knowledge, attitude and stigma score were found to be influenced by age, sex, ethnicity, religion, education and occupation. Strategizing the awareness programmes according to socio-demographic characteristics for enhancing the knowledge regarding leprosy cause, symptoms, transmission, prevention and treatment, could change the attitude to make it more favorable and thereby would help in reducing leprosy burden and enhancing the quality of life of leprosy patients.
| Leprosy, also known as Hansen’s disease, is a chronic infectious disease caused by bacteria Mycobacterium leprae [1]. It generally affects epidermis and peripheral nerves of the affected ones [2]. The disease is basically transmitted via prolonged close contact with untreated multibacillary leprosy patients through inhalation of bacilli [2; 3]. However, it is still an unequivocal issue regarding transmission of leprosy from one person to another [1]. Leprosy is more than a biological disease and is featured by stigma in the society leading to treating the affected ones with negative attitude [4]. Higher the associated stigma, lesser will be the chance to detect the new cases of leprosy early.
Despite being curable, each year globally around 200000 new cases of leprosy are detected [5]. Leprosy remains to be one of the neglected tropical diseases of developing countries in Africa and Asia with its burden being concentrated in Indonesia, Brazil and India. These three countries respectively contributed to 8%, 13% and 60% of the global new cases burden in 2015 while Nepal contributed to 1.3% [5; 6]. According to WHO factsheet, globally 210,758 new leprosy cases were detected in 2015 with prevalence rate of 0.29/10000 population [6]. However, the prevalence rate of leprosy in South-East Asian Region was 0.61/10000population [6]. Many countries have some areas of high endemicity showing high notification rates for new cases and witnessing continued transmission of leprosy [7]. Moreover, the open border between Nepal and India allows free migration of the population including leprosy affected persons. This may impede the early case detection and treatment.
Unfortunately, disability, disfigurement and the stigma associated with the leprosy have sustained and enhanced the stigma towards leprosy which in turn leads the affected ones to isolation, status concealment, delayed diagnosis and treatment. Leprosy affected ones in the early phase of the disease are generally suspicious of the diagnosis but fearing social isolation, leading to hesitancy towards seeking the advice and health care services [8]. A study done in Lalitpur Nepal in 1993 to 1995 showed that 6% (10/166) of leprosy affected persons reported of not seeking treatment earlier due to fear and social consequences including isolation [9]. Similarly, a quantitative study conducted in western Nepal in 2013 revealed that 66% of 135 leprosy affected persons intended to conceal their disease [10]. Moreover, in-depth interviews showed that 70% of 20 leprosy affected individuals intended to conceal their disease status with the major reasons being the fear of transmission, fear of exclusion, separation and rejection from the society [11]. Additionally, it has been observed that the social integration of people diagnosed of having leprosy is threatened when other people in the community come to know about it which results into applying the principle of silence and concealment of the disease status [12]. However, it is a serious public health problem since the multidrug therapy treatment should be initiated as soon as possible to prevent the disease progression resulting grade-2 disability which can pose further burden and severity condition in the lives of affected individuals [8; 13]. In order to control the cycle of concealment that leads to delayed health seeking and may lead to development of impairment and disability, early identification and treatment is critical.
Under leprosy control programme, Nepal declared that leprosy is no longer a public health problem in 2010 with the achievement of leprosy elimination in 2009 [14; 15]. However, there remain the challenges of sustaining this achievement and reducing the disease burden through quality services including early detection and prompt treatment [14]. On the contrary to decreasing incidence and prevalence of the disease, it has increased from 0.77 to 0.79, 0.84, 0.82, and 0.83 respectively during 2010 to 2014. Moreover, the country has been detecting more than 3000 new cases of leprosy annually [14]. Additionally, 18 districts of the nation have still prevalence above elimination level (prevalence rate of <1 case per 10000 population) and these districts contribute to 75% of the total incidence and accounts for around 3200 new cases of leprosy each year [15]. Most of these high prevalent districts are located in terai region of Nepal.
Despite the World Health Organization’s (WHO) target to eradicate leprosy by 2020, in the fiscal year 2016/2017, 19.77 leprosy affected individuals were diagnosed in every 10,000 population in the high prevalent districts of Terai regions of Nepal that are worst-hit by the burden of leprosy [16]. Towards increasing burden of leprosy in these regions, it has been argued that lack of awareness, poor personal hygiene, poor sanitation and low economic status of the people may be the reasons [17]. Additionally, people visit hospitals when the disease conditions get too worse; may it be due to inadequate awareness or lack of awareness towards skin-related diseases and its mode of transmission [17]. So, in order to address leprosy, better understanding about its cause, means of transmission and nature, and associated stigma is required.
In addition, to better understand leprosy and its social consequences it is important to study it in context, may it be the socio-cultural factors, belief systems, geography, economy, available resources or services [12]. The study done in Eastern Nepal revealed that leprosy affected individuals still encounter many constraints and restraints in their social life making them left out [4]. This fear of getting isolated may result in delayed in health care seeking [4]. Furthermore, a study in Nepal showed that majority of the respondent did not understand the cause of leprosy and were not aware of the duration of its treatment [18]. The study also emphasized the need of strengthening public/community awareness program towards leprosy [18]. According to a study done in Myanmar, it was found that community members were not sure about the cause of leprosy [19]. A study conducted in Pakistan revealed that more than one-fifth of the doctors did not have good knowledge regarding leprosy [20]. One of the major contributing factors towards the late diagnosis of leprosy is communities’ lack of knowledge regarding leprosy ultimately leading to increased likelihood of physical disability [2]. The available literatures have indicated that though leprosy is an old disease in terms of human civilization, it remains to be misunderstood and stigmatized. The knowledge and attitude of community towards leprosy remains poor which is mirrored by the study done in Cameroon which showed that less than one-fifth of the respondents knew the cause of leprosy and only about two-fifth of the respondents would shake hands with someone who is affected with leprosy [21].
A recent increase in new cases of leprosy from the Terai districts of Nepal has implications for the community where they live. How community members perceive leprosy affected persons and their attitude can affect their disease confession and health seeking at the hospital. Although studies in past have explored factors affecting community stigma towards leprosy in western and eastern Nepal, none in our knowledge has explored the community’s knowledge and attitude towards leprosy in Central southern Nepal. The main objective of this study was to assess the knowledge, attitude and stigma of leprosy amongst the community members living in Dhanusha and Parsa districts of Southern Central Nepal.
This was a cross-sectional study carried out amongst the community members living around teaching hospital (National Medical College and Teaching Hospital) of Parsa and Lalgadh Leprosy Hospital of Dhanusha district of state 2 of Nepal.
Respondents of both sexes aged between 18 years and 60 years were involved in the survey. However, individuals who were having hearing impairment and mental illness were excluded from the survey.
The sample size was calculated by using StatCalc Epi-Info. With prevalence 50% and margin of error 5%, the sample size was 384. Assuming the non-response of 10%, the final sample size calculated was 423.
The data was collected in two priority districts of state 2 of Nepal viz Parsa and Dhanusha. From both districts the communities (Bhediyahi and Mithila) surrounding the teaching hospital and leprosy hospital of Parsa and Dhanusha respectively were chosen for data collection. From Bhediyahi 212 and from Mithila 211 households were taken systematically. From each selected household one eligible respondent of age in between 18 and 60 years who gave his/her consent to participate in the study were selected for interview.
Structured questionnaire consisting of four parts was prepared after review of relevant literatures. The first part of the questionnaire was related to socio-demographic characteristics of the study participants, the second part was related to assessment of community member’s knowledge regarding leprosy, the third part was related to assessment of community member’s attitude towards leprosy and leprosy patients, and the fourth part was related to assessment of stigma attached in community towards leprosy and leprosy patients. The fourth part, EMIC scale (The Explanatory Model Interview Catalogue) is a reliable and validated tool to assess the community stigma towards leprosy (Rensen et al) [24]. The Nepali version of EMIC scale has been used in the past by one study conducted in western Nepal [25]. The other parts of the structured questionnaire in English were translated into Nepali language so that it can be relevantly used in the Nepalese context. A back-translation was then done to English language. The back translation of the tool from Nepali to English language was blind to the original questionnaire. Then the translated questionnaire in Nepali and back-translated questionnaire in English and the original questionnaire were all reviewed by assessing the meaning of each word to ensure the accuracy of the translation and the final questionnaire in Nepali was prepared.
Data was collected by the researchers themselves through face to face interview. The data was collected from the members of each selected household whose age was greater than 18 years old and less than 60 years and who gave his/her voluntary consent. The purpose of data collection was explained first to respondents to increase their awareness about the study before the start of the interview. They were informed regarding their voluntary participation in the study and their right to not answer any questions they did not want to. They were also ensured about regarding maintaining confidentiality of the information they provided as the researchers neither asked their name nor recorded any kind of respondent personal identity which could identify their name.
Knowledge of leprosy- Based on reported response each correct response towards each item of the knowledge questionnaire; the level of knowledge towards leprosy was assessed. Altogether sixteen self-reported items were considered for assessing level of knowledge which included questions with one correct answer as well as questions of dichotomous response (Yes/No) like hearing about leprosy, knowing its cause, sign and symptoms, leprosy as very infectious disease, its transmission, is it treatable, and its treatment. The level of knowledge was categorized as having good knowledge or poor knowledge. Good knowledge of leprosy- Respondents who were able to answer 75% or more of knowledge questions correctly were regarded as having good knowledge of leprosy. Poor knowledge of leprosy- Respondents who were able to answer less than 75% of knowledge questions correctly were regarded as having poor knowledge of leprosy. Additionally, the source of information on leprosy, knowledge regarding leprosy being a severe disease, and knowledge regarding first sign and symptoms of leprosy were also assessed. Attitude towards leprosy- It referred to community member’s perception towards leprosy and/or leprosy affected individuals. Attitude was assessed through 13 statements (10 positive statements towards leprosy and 3 negative statements towards leprosy) with response either ‘Yes’ or ‘No’. A response with ‘Yes’ towards each positive statement was given a score of 1 and a response with ‘No’ towards each positive statement was given a score of 0. Similarly, a response with ‘No’ towards each negative statement was given a score of 1 and a response with ‘Yes’ towards each negative statement was given a score of 0. Attitude was categorized as either having favorable attitude towards leprosy or unfavorable attitude towards leprosy based on individual respondent’s attitude score. Favorable attitude towards leprosy- Respondents who scored attitude score 7 or more (>50% of maximum attitude score) were regarded as having favorable attitude towards leprosy. Unfavorable attitude towards leprosy- Respondents who scored attitude score less than 7(<50% of maximum attitude score) were regarded as having unfavorable attitude towards leprosy. Level of stigma towards leprosy- The level of stigma towards leprosy was assessed based on respondent’s individual EMIC score. The EMIC score was calculated based on individual responses towards 15 items of the EMIC scale which is a standard tool to assess stigma towards leprosy. Each item/question in the scale was scored on the basis of response as “Yes = 2, Possibly = 1, No or Don’t know = 0”. Further, the level of stigma was assessed based on calculated individual respondent’s EMIC score and was categorized as high level of stigma, moderate level of stigma and low level of stigma. The category of the level of stigma towards leprosy was adapted. Respondents who scored EMIC score greater than 20 were regarded as having high level of stigma towards leprosy, respondents who scored EMIC score in the range of 10–20 were regarded as having moderate level of stigma towards leprosy, and respondents who scored EMIC score in the range of 0–10 were regarded as having low level of stigma towards leprosy.
The collected data were checked daily for completeness and consistency before data processing and analysis. The collected data was cleared, checked and analyzed by using tally sheet and computer. Data was entered and analyzed in SPSS version 20. Both descriptive and statistical inferences were used to analyze the data. Descriptive statistics like frequency, percentage, and median were used to describe the socio-demographic characteristics, level of knowledge, level of attitude and level of stigma of the study participants. Proportions were calculated, and the Chi-square test was used to examine relationship between socio-demographic characteristics and level of knowledge; socio-demographic characteristics and level of attitude; and level of knowledge and level of attitude. Mann Whitney U test and Kruskal Wallis H test were used to analyze the difference in total perceived stigma score using EMIC between different socio-demographic characteristics of the community. Further, binary logistic regression analysis was carried out to determine predictors of unfavorable attitude towards leprosy.
Ethical clearance was obtained from Institutional Review Committee (IRC) of National Medical College (FNMC-310-074-075). Further, for each study participants, the purpose of the study was stated by the researchers prior to data collection. In addition, participants were informed that they have full right to refuse participating in the study and can interrupt the interview if not comfortable with it. However, they were informed that their participation in the study is very important. Participation of each respondent in this study was voluntary and data was collected from each participant once they gave an informed consent. Confidentiality of the information was maintained, and anonymity of the study participants was respected during data processing and analysis.
Four hundred and twenty-three (423) individuals were contacted and interviewed in the survey with age ranging between 18 years and 60 years, with around 36% above 40 years and 29% below 24 years. They were 58.6% males, 34.8% Brahmin/Chhetri (considered to be higher class as per ethnicity in Nepal) and 69.7% married. Almost half of the respondents (49.9%) were from nuclear family. More than one-third of the participants (40.9%) had bachelors level or higher degree of education. Most of them were service holder (45.6%) followed by farmer (24.3) and with monthly income more than Nepalese Rupees twenty thousand (49.2%).
All the study participants had heard about leprosy. More than 4/5th of the participants (88.4%) reported of knowing the cause of leprosy. More than 3/4th of the study participants (79.4%) believed leprosy to be highly infectious disease. Similarly, 69% of them reported of knowing how leprosy is transmitted. Also, 81.1% of the study participants reported of knowing the signs and symptoms of leprosy. With regards to the first sign and symptom of the disease, 46.8% of the participants reported skin involvement, 3.1% reported nerve involvement and 31.2% reported both skin and nerve involvement as the first sign of the disease.
Although 88.4% responded of knowing the cause of leprosy, only 62.6% of them reported of bacteria being the cause of leprosy. Surprisingly, 21.1% of them reported bad blood as the cause of leprosy, followed by curse by god (8.8%), heredity (3.2%), bad deeds (2.7%), and unclean environment (1.6%). However, only 43.8% responded that leprosy is transmitted by prolonged close contact with leprosy affected individuals. Majority of the participants (87.7%) thought leprosy to be a curable disease. But, 25.7% of them reported religious rituals as the treatment for leprosy. In addition, most of the participants (65.2%) thought leprosy to be a severe disease. Nevertheless, only 62.2%, 48.9%, 30%, and 66.7% of them reported skin patches, loss of sensation, deformity and ulcer respectively to be the signs and symptoms of the disease. Surprisingly, 28.4% and 31.2% of the participants also responded tingling and skin irritation respectively to be the signs and symptoms of leprosy. Around 2/5th of the study respondents (38.8%) said that they would not go to hospital or doctor if they get to know of having leprosy. Based on correct response towards the questions related to knowledge, it was found that 57.9% of the study participants had poor knowledge of leprosy and remaining (42.1%) of them had good knowledge of leprosy.
The major source of information about leprosy for the community people was found to be hospital and health worker comprising (33.1%) followed by media (30.7%).
The knowledge of leprosy among community people were influenced by socio-demographic characteristics (Table 1). It was found that there was highly significant association between the level of knowledge of leprosy among study participants with age, sex, ethnicity, religion, educational status, occupation and monthly income (P< 0.001). Most of the study participants from older adult age group (>45 years) had good knowledge of leprosy while majority of female had poor knowledge of leprosy as compared to male. Most of the respondents from Dalit or Janajati background had poor knowledge of leprosy. Majority of the non-Hindu respondents had good knowledge of leprosy. Respondents with higher education including bachelor or master or higher degree were having good knowledge of leprosy. Respondents with occupation service had good knowledge of leprosy. Most of the individuals with monthly income more than Nepali Rupees 20000 (Nepali Rupees 20000 approximately equivalent to USD 178 as of 29 November 2018) had good knowledge of leprosy. The level of knowledge of leprosy was also significantly associated with type of family of the study participants where individuals from joint family had good knowledge of leprosy (P = 0.034). However, marital status was not found to influence the level of knowledge of leprosy (P = 0.101).
Around 3/5th of the study participants (59.1%) had unfavorable attitude towards leprosy and 40.9% had favorable attitude. Most of the participants (51.8%) responded they would sit together with leprosy affected individuals in public conveyance, 51.3% said that they would not avoid having food or other activities with leprosy patients, would agree to work in the same environment with leprosy affected ones (52.5%), would not feel shame to share the status to others if anyone in the family had leprosy (53.4%). More than 4/5th (96.2%) reported that they would support leprosy affected ones if they would need it. However, only 12.5% reported that they would share foods with leprosy patients, only 32.6% would take cooked foods by the leprosy affected individuals, and majority of them reported that they would not marry individuals from family with history of leprosy (82%). Similarly, majority of them (84%) reported that it is difficult for anyone with leprosy to get married.
The level of attitude towards leprosy among community members were found to be influenced by socio-demographic variables (Table 2). It was found that there was highly significant association between level of attitude towards leprosy and age, sex, ethnicity, religion, marital status, educational status and occupation (P<0.001). Most of the study participants from older adult age group (>45 years) had favorable attitude towards leprosy while majority of female had unfavorable attitude towards leprosy as compared to male. Most of the respondents from Dalit or Janajati background had unfavorable attitude towards leprosy. All of the non-Hindu respondents had favorable attitude towards leprosy. Majority of the married respondents had unfavorable attitude towards leprosy. Respondents with higher education including bachelor or master or higher degree were having favorable attitude towards leprosy. Majority of the housewives were found to have had unfavorable attitude towards leprosy.
The binary logistic regression showed that the individuals who knew how leprosy is transmitted are likely to have 3.35 times favorable attitude of sitting together with leprosy affected individuals in the public conveyance (Table 3). Also, those who think leprosy to be very infectious would have 2.1 times higher chance of staying far away from leprosy patients. In addition, those who know it is transmitted by prolonged close contact only would have 13 times higher chance of letting own child to play with children of leprosy affected individuals.
The Chi-square test showed that the attitude was highly influenced by the knowledge of leprosy among the community members (Table 4). There was highly significant association between level of attitude and level of knowledge of leprosy among the study participants (P<0.001). The finding revealed that better the knowledge of leprosy among individuals, more the chance of having positive attitude towards leprosy and leprosy patients.
The EMIC profile of the study participants revealed that 44% were having high stigma, 33.3% were having moderate level of stigma and 22.7% were having low stigma towards leprosy and leprosy patients.
The assessment of EMIC score was done to measure the perceived stigma towards leprosy and leprosy patients in community members. The median score was calculated to analyze the difference of stigma between various groups. It was found that 43.7% of the study participants would keep others from knowing leprosy status if possible, 32.2% would think less of self due to leprosy affected individual in the family, 24.8% think that leprosy has caused shame or embarrassment in the community, 29.3% feel others think less of a person with leprosy, 39.7% think that there would be adverse effect on others if they know someone’s status of leprosy, and 29.6% think that others would avoid a person with leprosy (Fig 1).
It was found that there was highly significant association between EMIC score and age, ethnicity, marital status, educational status, occupation, monthly income, knowledge of leprosy transmission, knowledge of cause of leprosy, sex, religion, and income sufficiency for living, knowledge regarding leprosy is treatable, Knowledge of sign and symptoms of leprosy (P<0.001) (Table 5, Table 6, Table 7). The finding showed that perceived stigma towards leprosy (EMIC score) was lower at increasing age > 40 years. EMIC score was high among Dalit/Janajati, unmarried, female, Hindu participants, and participants with insufficient income for living. Similarly, study participants who did not have knowledge of leprosy transmission and who did not know the cause of leprosy had high EMIC score. Further, the study participants who thought leprosy as not curable disease were found to have had high EMIC score. Furthermore, respondents who knew loss of sensation, deformity and ulcer as sign of leprosy had low EMIC score while respondents who knew skin patch as sign of leprosy had high EMIC score. Respondents who though skin itchiness as sign and symptom of leprosy had high EMIC score. Additionally, the study participants who correctly knew bacteria as the cause of leprosy and prolonged close contact being the means to transmit leprosy had low EMIC score. Similarly, respondents with higher educational status (Bachelor’s degree and above), engaged in service and with monthly income of >20000 NRs were found to have had low EMIC score. However, there was no association between EMIC score and type of family (P = 0.177), residence or district (P = 0.56), knowledge regarding leprosy being an infectious disease (P = 0.551), knowledge regarding leprosy being a severe disease (P = 0.51) and knowing tingling as a sign of leprosy (P = 0.133). These factors (type of family, residence, knowledge regarding leprosy being a severe disease and knowing tingling as a sign of leprosy) were found to have no influence on perceived stigma towards leprosy in community people.
The overall findings of the study revealed that only 42.1% of the community people had good knowledge of leprosy with major source of information being local health worker and media (63.8%). Similarly, only 40.9% of the study respondents were found to have favorable attitude towards leprosy. Additionally, it was also found that the community-based stigma towards leprosy and leprosy affected persons is still prevalent among study participants living in the study districts- Dhanusha and Parsa.
The findings of this community-based study showed that still around 3/5th of the study participants had poor knowledge of leprosy. This result is supported by the study done in eastern Nepal [4]. Similarly, the study done in western Nepal also showed similar result with almost half having some kind of knowledge on leprosy cause, transmission and clinical manifestation [25]. This finding is also congruent to the study done in community members of Andhra Pradesh and Orissa which showed that 35–50% of the respondents had high level knowledge of leprosy [26]. However, a study done in Indian rural community to assess knowledge and attitude towards leprosy after post elimination phase showed that 78.94% of the respondents had good knowledge and 69% had positive attitude towards leprosy [27]. Similarly, a study done in Ethiopia also revealed worse scenario with around 80% of the respondents having low level of knowledge of leprosy [28]. The reason for this difference in knowledge and attitude may be due to the different socio-cultural context in relation to Terai districts of Nepal.
This study and Ethiopian study resembles in relation to the response that 100% of the participants had heard about leprosy [28]. However, the study done in Cameroon showed that only 82.4% of the respondents had heard about leprosy [21]. Apart from responding bacteria as the cause of leprosy, participants also responded wrongly citing god’s sin, bad deeds, bad blood, and heredity as the causes of leprosy which is similar to the findings of the study done in Cameroon and Ethiopia [21; 28]. This finding has also been supported by the study done in Uttar Pradesh, India [29]. These myths are rooted in the socio-cultural context of the communities in Asia, Africa and South America as evident in the literature written by Wong and Subramaniam [30].
The current study revealed that around 3/5th of the respondents had unfavorable attitude towards leprosy. The findings related to prevalence of unfavorable attitude such as eating limitation and negative behavior in the community are consistent to the study done in eastern Nepal [4]. One of the reasons behind this unfavorable attitude may be overall poor literacy rate and more specifically poor literacy rate of female of the study region. The level of attitude among the community members towards leprosy is also similar in the study done in Ethiopia [28].
The study findings showed that knowledge and attitude of leprosy among community are influenced by various socio-demographic characteristics of the community members. This finding is supported by the study done in community members in western region of Nepal [25]. This result is also congruent with the studies done in Cameroon and Ethiopia [21; 28].
This study showed that stigma is still prevalent in communities of Terai districts of Nepal. The study identified various kinds of stigma/myths such as participants preferring to hide their leprosy status, thinking less of self if any of the family member is affected by leprosy, thinking that leprosy has caused shame, feeling others think less of a person with leprosy, thinking that others would avoid a person with leprosy, and reporting that it would cause problem in marriage. This indicated that preference to concealment towards leprosy status is still prevalent due to grounded stigma of leprosy in communities of Nepal. This finding is similar to study done in Western region of Nepal [25]. Such finding relating to prevalence of preference towards concealment of the disease and feeling of shame towards leprosy is also congruent to studies done in eastern Nepal [4; 31]. The finding regarding stigma related to leprosy causing problem in marriage was similar to the qualitative study done in South East Nepal [32]. The results regarding the nature of perceived stigma towards leprosy and leprosy patients among community people was similar in the study conducted in Thailand [33]. The finding is also supported by the study done in Indonesia [34]. The level of stigma is high among high proportion of study participants in this study and this finding is supported by the studies done in western Nepal [25] and rural India [34], example around 47% and 22% of the respondents with response of not preferring to buying foods from leprosy affected individuals in western Nepal and rural India respectively which is quite similar to this study (29%). In this study, socio-demographic variables like age, ethnicity, marital status, education, occupation, etc. and knowledge were found to influence stigma and EMIC score among community members. The finding is supported by the findings of the study done in Pokhara [25]. Nevertheless, unlike the finding of the study done in Pokhara, there was no association between perceived stigma of leprosy and residence (districts) of the study participants [25]. The reason for this dissimilarity may be the effect of teaching hospital and leprosy hospital raising similar kind of consciousness towards leprosy among the people. Knowledge and beliefs about leprosy has been found to be associated with stigma in leprosy in many studies conducted in China [35] and Nepal [4]. Myths such as not allowing child to play with children of leprosy affected individuals, not sitting together with leprosy affected ones, not preferring to marry with one with family history of leprosy, and not sharing foods with leprosy affected individuals suggest how deep-rooted misconceptions of leprosy are prevalent in the communities. It was also found that study participants reported that leprosy affected ones would get difficulty in getting job.
There were various unfavorable attitudes towards leprosy prevalent in the communities of Terai districts of Nepal. Similarly, knowledge regarding leprosy causes, transmission, sign and symptoms and treatment were also not adequate for breaking the transmission of the disease and early identification for prompt treatment. Also, different myths and misconceptions are still present in the communities in different socio-demographic groups of the population. Progress towards leprosy eradication is only possible by making people to better understand its transmission. So, to make the leprosy control programme a success by eliminating and eradicating this old disease, the first and foremost thing to do is to strategize the programmes as per strata according to socio-demographic characteristics of the population for enhancing their knowledge regarding leprosy, its cause, symptoms, transmission, prevention and treatment and thereby changing the attitude to make it more favorable towards leprosy.
Furthermore, advocacy programmes should be developed engaging people affected with leprosy, local health workers, and deep-rooted traditional healers of the rural communities and local media to provide information about leprosy to the community people. Also, empowerment workshops should be organized for the leprosy affected individuals including unaffected females of the community who can further help to aware other people once empowered. Additionally, more information, education and communication materials need to be developed and made accessible to the general public in both the least and the most affected communities. Further studies are needed to develop new diagnostic and screening tools which can identify leprosy at earlier and hidden stage at the community level. As state 2 of Nepal lies adjacent to communities of India which is one of the countries with concentrated burden of leprosy, further studies on communities of state 2 of Nepal bordering to Bihar state of India is recommended which might show the issue of inadequate knowledge, negative attitude and high stigma at more worse scenario. These communities need to be addressed in terms of strengthening their capacity to prevent and control this growing burden of leprosy with sufficient supporting evidence. Further study on the issues of this neglected tropical disease in a larger scale both in rural and urban areas of Nepal is recommended to bring forth clearer picture.
|
10.1371/journal.ppat.1003473 | A Type IV Pilus Mediates DNA Binding during Natural Transformation in Streptococcus pneumoniae | Natural genetic transformation is widely distributed in bacteria and generally occurs during a genetically programmed differentiated state called competence. This process promotes genome plasticity and adaptability in Gram-negative and Gram-positive bacteria. Transformation requires the binding and internalization of exogenous DNA, the mechanisms of which are unclear. Here, we report the discovery of a transformation pilus at the surface of competent Streptococcus pneumoniae cells. This Type IV-like pilus, which is primarily composed of the ComGC pilin, is required for transformation. We provide evidence that it directly binds DNA and propose that the transformation pilus is the primary DNA receptor on the bacterial cell during transformation in S. pneumoniae. Being a central component of the transformation apparatus, the transformation pilus enables S. pneumoniae, a major Gram-positive human pathogen, to acquire resistance to antibiotics and to escape vaccines through the binding and incorporation of new genetic material.
| Natural genetic transformation, first discovered in Streptococcus pneumoniae by Griffith in 1928, is observed in many Gram-negative and Gram-positive bacteria. This process promotes genome plasticity and adaptability. In particular, it enables many human pathogens such as Streptococcus pneumoniae, Staphylococcus aureus or Neisseria gonorrhoeae to acquire resistance to antibiotics and/or to escape vaccines through the binding and incorporation of new genetic material. While it is well established that this process requires the binding and internalization of external DNA, the molecular details of these steps are unknown. In this study, we discovered a new appendage at the surface of S. pneumoniae cells. We show that this appendage is similar in morphology and composition to appendages called Type IV pili commonly found in Gram-negative bacteria. We demonstrate that this new pneumococcal pilus is essential for transformation and that it directly binds DNA. We propose that the transformation pilus is an essential piece of the transformation apparatus by capturing exogenous DNA at the bacterial cell surface.
| Natural transformation, first discovered in Streptococcus pneumoniae [1], is observed in many Gram-negative and Gram-positive bacteria [2]. It increases bacterial adaptability by promoting genome plasticity through intra- and inter-species genetic exchange [3]. In S. pneumoniae, a major human pathogen responsible for severe diseases such as pneumonia, meningitis and septicemia, transformation is presumably responsible for capsular serotype switching and could therefore reduce the efficiency of capsule-based vaccines after a short period [4]. In this species, it occurs during a genetically programmed and differentiated state called competence that is briefly induced at the beginning of exponential growth. During this competent state, pneumococci secrete a peptide pheromone called Competence-Stimulating-Peptide (CSP) [5], which spreads competence in the pneumococcal population. Interestingly, in S. pneumoniae, some antibiotics and DNA-damaging agents induce competence, which would act as an alternative SOS response and ultimately increases bacterial resistance to external stresses [6].
During transformation, environmental DNA is bound at the surface of competent cells and transported through the cell envelope to the cytosolic compartment. This process has been mostly studied in the Gram-positive bacterium Bacillus subtilis with additional information coming from studies in S. pneumoniae [7], [8]. In both species, a DNA translocation apparatus mediates the transfer of DNA through the cellular membrane. In S. pneumoniae, it is composed of ComEA, EndA, ComEC and ComFA. Incoming double-stranded DNA would bind the membrane receptor ComEA. One DNA strand crosses the membrane through ComEC while the endonuclease EndA degrades the other strand. On the cytoplasmic side, ComFA, an ATPase that contains a helicase-like domain, would facilitate DNA internalization through ComEC. Once inside the bacterium, single-stranded DNA is either integrated into the chromosome by RecA-mediated homologous recombination or entirely degraded.
Strikingly, all transformable Gram-positive bacteria also carry a comG operon that resembles operons encoding Type IV pili and Type II secretion pseudopili in Gram-negative bacteria, as well as a gene encoding a prepilin peptidase homolog, pilD [7]. In B. subtilis and S. pneumoniae, comG and pilD genes are exclusively expressed in competent cells and are essential for transformation [9], [10], [11]. In S. pneumoniae, the comG operon encodes a putative ATPase (ComGA), a polytopic membrane protein (ComGB) and five prepilin candidates named ComGC, ComGD, ComGE, ComGF and ComGG (Figure 1A and B and table S1). By homology with Type IV pili, it is generally proposed that these proteins could be involved in the assembly of a transformation pseudo-pilus at the surface of competent cells [7], [8], [12]. So far, two studies show that a large macromolecular complex containing ComGC can be found at the surface of competent B. subtilis cells [9], [12]. In this complex, ComGC subunits appear to be linked together by disulfide bridges [9]. All the other ComG proteins and the PilD homolog, ComC, are necessary for the formation of this complex [9]. It was proposed that this complex could correspond to the transformation pseudo-pilus.
Despite these first clues, no transformation appendage could be directly visualized at the surface of any competent Gram-positive bacterium. In addition, the function of the ComG proteins during transformation remains unclear. Mutations in the cytosolic ComGA protein abolish DNA binding at the surface of both B. subtilis and S. pneumoniae [13], [14], [15]. This strongly suggests that the ComGC-containing macromolecular complex detected at the surface of competent B. subtilis cells could bind DNA. However, it was recently shown that ComGA is the only ComG protein essential to the initial DNA binding at the surface of competent B. subtilis cells [14]. This protein would interact with an unknown DNA receptor at the surface of competent cells while the other ComG proteins would only be required at a later stage during transformation.
In this study, we provide the first direct evidence for the existence of a transformation pilus in a Gram-positive bacterium. We discovered a new appendage at the surface of competent pneumococci that we could visualize using immuno-fluorescence and electron microscopy. Competent cells harbor one or a few appendages that are morphologically similar to Type IV pili found in Gram-negative bacteria. We were able to purify this pilus and showed that it is essentially composed of the ComGC pilin. We also demonstrate that pilus assembly is required for transformation. As we provide direct evidence that the transformation pilus binds extracellular DNA, we propose it is the primary DNA receptor at the surface of competent pneumococci.
Mechanical shearing is frequently used to release bacterial surface appendages and to study their protein composition [16]. To see if ComGC was part of a macromolecular complex at the surface of competent S. pneumoniae, we adapted the method to this bacterium and raised antibodies against the purified soluble domain of ComGC. Using this antibody, we showed that ComGC could be detected by immunoblotting in the sheared fraction of competent bacteria (Figure 2A). While ComGC level in the cell fraction was not affected, no ComGC could be found in the sheared fraction in a comGA knockout mutant (Figure 2A). These data strongly suggest that ComGC is part of an extra-cellular appendage and that ComGA is necessary to its assembly.
We inserted a FLAG tag at the C-terminus of ComGC to directly visualize the competence-induced appendages by immuno-fluorescence. It was not possible to insert the sequence encoding the tag at the comGC locus on the chromosome because comGC and comGD genes overlap in the comG operon. Therefore, a copy of comGC encoding a C-terminally FLAG-tagged ComGC (ComGC-FLAG) was integrated ectopically into the chromosome of S. pneumoniae under the control of a competence-induced promoter [17]. The transformation efficiency was not affected in this strain (Figure 2B). Using anti-FLAG antibodies, we could show by immuno-fluorescence that almost all the cells appeared to harbour one or a few ComGC foci or distinct fluorescent appendages (Figure 3A and B; Figure S1A). Due to sample preparation, many broken appendages were also found in the background. No preferential location of the foci/appendages at the cell surface was observed. They are absent in comGA knockout cells (Figure 3A). Note that anti-ComGC antibodies were not able to label the competent cells. They probably recognize epitopes that are masked when ComGC is included in the appendages.
Using electron microscopy, we observed filaments attached to the cell surface of negatively stained competent pneumococci (Figure 4A and B). These flexible filaments are 5–6 nm in diameter. Their length could reach up to 2–3 micrometers (Figure 4A). A maximum of 2–3 filaments per cell could be observed. Their average length was difficult to assess because they break easily into smaller fragments during sample preparation. Using the ComGC-FLAG expressing strain, we confirmed by immunogold-labelling that they contain ComGC (Figure 4C).
Appendages were then purified using anti-FLAG affinity chromatography after mechanical shearing. Appendage fragments of between 50 and 500 nm in length were observed by electron microscopy (Figure 5A), showing that these filamentous structures do not disassemble during purification. SDS-PAGE analysis of the purified fraction showed that ComGC is the major component of the appendages (Figure 5B). Using whole protein mass profiling by high-resolution mass spectrometry [18], we could only detect ComGC and ComGC-FLAG in the purified material (Figure 5C), confirming that ComGC is the major constituent of these appendages. Monoisotopic mass measurements of intact proteins and top-down fragmentation using a variety of activation techniques confirmed that the ComGC prepilin is cleaved after the alanine residue in position 15 and that the first amino acid of the mature protein is methylated, presumably by PilD (Figure S2). Indeed, PilD homologs in Gram-negative bacteria catalyze this post-translational modification of the Type IV pilins [19]. No other post-translational modification was detected in ComGC. Other proteins, including other ComG proteins, were not detected in the purified material by the methods used in this study. This suggests that these proteins are either absent, present in very low amount within the appendage or weakly bound to it and lost during sample preparation. These morphological and biochemical features are typical of Gram-negative Type IV pili. Therefore, we propose that the competence-induced appendage observed in S. pneumoniae belongs to the Type IV pilus family.
It was important to determine whether these competence-induced pili were involved in transformation. Indeed, it was previously shown that S. pneumoniae and B. subtilis comGA knockout could not be transformed (Figure 2B) [9] [13]. In this study, we were able to show in S. pneumoniae that comGA mutant cells lack pili (Figure 2A and 3A). It was enticing to conclude that competence-induced pili assembly is essential for transformation. However, it was recently shown that a comGA mutation could have a pleiotropic effect on transformation in B. subtilis [14]. Therefore, we generated a comGC mutant in S. pneumoniae in which the conserved glutamic acid in position 5 was substituted by an alanine (Figure 1B). Such a substitution was shown to impair Type IV pilus assembly in Gram-negative bacteria [20]. ComGC cellular level was not affected by this point mutation (Figure 2A). Our results show that this mutant strain could not assemble any pilus and that it was defective for transformation (Figure 2A and B). Therefore we conclude from the analysis of both comGA and comGC(E5A) mutants that the assembly of the competence-induced pilus is required for transformation.
The nature of the primary DNA receptor at the surface of transformable Gram-positive bacteria is not known. It is generally proposed that the transformation pseudopilus would bind extracellular DNA at the surface of competent Gram-positive bacteria [8], [21]. However, this hypothesis has never been confirmed experimentally. Using affinity purification, we show that DNA naturally released in the culture medium co-fractionates with the purified pili. No DNA could be found in the purified fraction in absence of the pilus (Figure 6A). These data were a first hint suggesting that DNA present in the environment could bind to the transformation pilus. However, it was not clear if this binding was related to the transformation process or fortuitous. By using specific electron microscopy methods [22], we visualized DNA directly bound to the transformation appendage after adding linear double stranded DNA (dsDNA) to competent bacteria. Long stretches of dsDNA interacting with the transformation pilus were observed with clearly visible multiple contact points (Figure 6 B–E). Interestingly, it was extremely difficult to see DNA bound on the pilus in the reference bacteria (R1501 strain), which are known to internalize exogenous DNA quickly [23]. On the other hand, in ComEC and comFA mutants, we could easily observe bound DNA on transformation pili. These strains are defective for DNA uptake and accumulate bound DNA at their surface [13]. Given that the dsDNA was added in large excess, no difference between the reference and mutant strains should be observed if DNA binding on the pili was a coincidental event. The fact that the uncoupling of DNA binding and uptake processes facilitates the observation of the DNA/pilus interaction is a strong indication that DNA binding on the transformation pilus is related to the transformation process.
The comG operon is conserved in all transformable Gram-positive bacteria. This operon encodes proteins that are homologous to proteins involved in Type IV pilus assembly in Gram-negative bacteria. Therefore it has been proposed that a pilus (or pseudopilus) could be assembled at the surface of competent Gram-positive bacteria. Since all comG genes are essential for transformation, this pilus could be directly involved in transformation. The first biochemical clues for the existence of a transformation pilus were found in B. subtilis although decisive observational support was lacking. In addition, it was not clear if the ComGC-containing macromolecular complex found in B. subtilis was a common feature of competent Gram-positive bacteria or specific to this species. Finally, the function of this putative transformation pilus, and in general of the ComG proteins, was unclear.
The pneumococcal transformation pilus represents a newly discovered pneumococcal surface structure. For a long time, no external appendage could be found at the surface of S. pneumoniae cells while many electron microscopy images were published in the literature. Recently, sortase-mediated pili have been discovered in some pathogenic S. pneumoniae strains [24]. To our knowledge, no specific ultrastructural study of competent S. pneumoniae has ever been described. Here, we analysed a laboratory strain that is commonly used to study the transformation process in S. pneumoniae [10] [13]. In this strain, competence can be induced in a rapid and synchronous manner upon addition of synthetic CSP in the medium of an exponentially growing culture [5], [25]. To make sure that the appearance of the transformation pilus is a common feature of competent pneumococci and not a mere one-off property of our reference strain, we observed negatively stained G54 and CP strains by electron microscopy. The G54 strain is a wild-type clinical strain. The CP strain is a laboratory strain that has a different genetic background than our reference strain [26]. In both cases, transformation pili were observed at the surface of competent cells (Figure S3). Therefore, we think that transformation pili are found at the surface of most, if not all, pneumoccocal strains, including clinical strains.
The pneumococcal transformation pilus is morphologically very similar to Type IV pili found in many Gram-negative bacteria. Its major component, the ComGC pilin, is cleaved and probably methylated by a PilD homolog. We therefore propose that the transformation pilus is a bona fide Type IV pilus. Since its length can reach up to 2–3 µm, we think that the “pseudo-pilus” appellation does not apply to the pneumococcal transformation appendage. By comparison, the type II secretion pseudo-pilus is just 50–100 nm long [27]. The transformation pilus is the first Type IV pilus clearly observed in a Gram-positive bacterium. So far, Type IV pilus-dependent gliding motility had been described in Clostridium species [28]. However, no clear picture of this pilus was provided. A recent genomic study show the existence of numerous and diverse Type IV pilus-like operons in a wide range of Gram-positive bacteria [29]. This suggests that many other Type IV-like pili remain to be discovered in these bacteria. The conservation of comG operons argues in favor of the presence of a transformation pilus in all naturally transformable Gram-positive bacteria. However, species-specific variations in pilus length can be anticipated because of variations in thickness of the capsule and/or the cell wall.
Cells were grown at 37°C under anaerobic condition, without agitation, in a Casamino Acid Tryptone medium (CAT) up to OD600 = 0.3 for stock cultures [40]. After addition of 15% glycerol, stocks were kept frozen at −80°C. For competence induction, cells were grown in CAT supplemented with BSA (2 g/L), calcium chloride (1 mM) and adjusted to pH = 7.8. Competence was triggered by incubating cells with the Competence Stimulating Peptide (CSP) at OD600 = 0.1 for 12 min as described previously [40]. For transformation, DNA was then added and transformants were selected on CAT agar plates [17]. Competence was induced following the same protocol in G54 and TCP1251 strains.
For transformation efficiency assays, 100 µL of competent bacteria were transformed by the addition of 100 ng of S. pneumoniae R304 genomic DNA (contains the streptomycin resistance gene str41). Bacteria were plated in presence and absence of streptomycin (100 µg/mL final concentration) and incubated at 37°C overnight before colony counting.
The annotated names of the comG genes in different strains of S. pneumoniae are listed in Table S1. The S. pneumoniae strains used derived from the non-capsulated R6 strain and are listed in Table S2. The comGC-FLAG gene was cloned by PCR using genomic DNA of pneumococcal R6 strain (ATCC BAA-255) as template. The resulting fragment was digested with NcoI and BamHI and inserted into the same sites of the pCEPx vector [17]. RL001 strain was constructed by transformation of R1501 cells with the pCEPx plasmid containing comGC-FLAG, followed by selection with kanamycin (Kan). RL002 was obtained by transformation of RL001 with R1062 chromosomal DNA and selection with spectinomycin (Spc). For RL003, a 2 kb genomic fragment of R6 genome containing comGC in the middle was amplified, and the codon 20 was changed from GAG to GTG by cross-over PCR. R1501 was transformed with this modified genomic fragment, and clones were screened by sequencing the comGC gene.
Chemically competent Escherichia coli BL21 Star (Life Technologies) were used for heterologous production of ComGC soluble domain. The corresponding DNA sequence was amplified from genomic DNA of strain R800 and cloned into pET15b expression vector (Novagen), using NdeI/XhoI. The protein was purified from the soluble fraction using IMAC affinity and gel filtration in 50 mM Tris/HCl pH = 8, 200 mM NaCl. The anti-ComGC were raised against the purified protein (Eurogentec).
Shearing experiments were adapted from Sauvonnet et al. [41]. Competence was induced exactly as described above in a 50 mL culture. Cells were harvested by centrifugation 15 min at 4,500 g, 4°C. The pellet was suspended in 1 mL LB and immediately vortexed for 1 min to apply mechanical pressure. The suspension was then centrifuged twice at 13,000 g for 5 min to separate the bacteria (pellet fraction) from the pilus-enriched supernatant (sheared fraction). The supernatant was then precipitated with 10% trichloroacetic acid for 30 min on ice. Both fractions were loaded on SDS 15% polyacrylamide gels and subjected to electrophoresis and immunoblot with rabbit polyclonal antibodies raised against ComGC soluble domain (38–108) or anti-FLAG M2 antibody (Sigma-Aldrich F1804).
The pili containing ComGC-FLAG were purified from the sheared fraction of a 1 L culture. Shearing was performed in 2 mL Tris Buffered Saline (TBS, Tris pH 7.6 0.05 M, NaCl 0.15 M, protease inhibitor cocktail Roche 11873580001) and incubated overnight on a rotating wheel at 4°C with ANTI-FLAG M2 affinity resin (Sigma-Aldrich A2220). After washing with TBS, the pili were eluted by adding 3×FLAG-peptide at 100 µg/mL (Sigma Aldrich F4799) 30 min at room temperature under agitation.
To prevent DNA aspecific binding on the ANTI-FLAG M2 affinity resin, the resin was saturated 2 h at 4°C with a 1.5 kb PCR fragment (20 ng/µL). For DNA detection, 20 µL of the eluted pili were run on a 1% agarose gel and stained with SYBR safe (Life technologies S33102).
Competence was induced exactly as described above in a 10 mL culture. Cells were harvested by centrifugation 15 min at 4,500 g, 4°C. The pellet was suspended in 60 µL phosphate-buffered saline (PBS) (Sigma-Aldrich P4417). A drop of this suspension was placed on a glow discharged carbon coated grid (EMS, USA) for 1 min. The grid was then placed on a drop of PBS-3% formaldehyde, 0.2% glutaraldehyde for 10 min, and washed on drops of distilled water. The grids were then treated with 2% uranyl acetate in water. Specimens were examined using a Philips CM12 transmission electron microscope operated at 120 kV. Pictures were recorded using a camera KeenView (SIS, Germany) and ITEM software. For immunogold labelling, additional steps were applied after fixation: 3 washes with PBS, PBS–50 mM NH4Cl (10 min), 3 washes with PBS, PBS with 1% BSA (5 min), 1 hour incubation with ANTI-FLAG M2 antibody (Sigma-Aldrich F1804) diluted 1/100 in PBS with 1% BSA, 3 washes with PBS-BSA 1% (5 min), 1 hour incubation with goat anti-mouse antibody (5 nm gold particles, BritishBioCell, UK) diluted 1/25 in PBS containing 1% BSA.
S. pneumoniae cells were grown in the same conditions as above for visualization by electron microscopy. Cells were harvested by centrifugation for 15 min at 4,500 g, 4°C. The pellet was suspended in 500 µL PBS and directly immobilized on poly-L-lysine-coated coverslips. Samples were fixed for 30 min with 3.7% formaldehyde, washed 3 times with PBS containing 1% BSA and incubated on a 100 µL drop of anti-FLAG antibodies (1∶300) and secondary Alexa Fluor 488- coupled anti-mouse IgG (Invitrogen). Samples were examined with an Axio Imager.A2 microscope (Zeiss). Images were taken with AxioVision (Zeiss) and processed in ImageJ [42].
Protein samples were desalted and eluted directly into a 10 µL spray solution of methanol∶water∶formic acid (75∶25∶3). Approximately 4 µL was loaded into a coated, medium sized, nano-ESI capillary (Proxeon) and introduced into an Orbitrap Velos mass spectrometer, equipped with ETD module (Thermo Fisher Scientific, Bremen, Germany) using the off-line nanospray source in positive ion mode. A full set of automated positive ion calibrations was performed immediately prior to mass measurement. The transfer capillary temperature was lowered to 100°C, sheath and axillary gasses switched off and source transfer parameters optimised using the auto tune feature. Helium was used as the collision gas in the linear ion trap. For MSn experiments, ions were selected with a 3 Da window and both CID and HCD were performed at normalised collision energies of 15–25%, with the appropriate HCD charge state set and other activation parameters left as default. For ETD the reagent gas was fluoranthene and the interaction time 10 ms. Supplemental activation was used as noted. The FT automatic gain control (AGC) was set at 1×106 for MS and 2×105 for MSn experiments. Spectra were acquired in the FTMS over several minutes with one microscan and a resolution of 60,000 @ m/z 400 before being summed using Qualbrowser in Thermo Xcalibur 2.1. Summed spectra were then deconvoluted using Xtract and a, b, c−1, y, z, z+1 ions assigned using in house software at a tolerance of 5 ppm. N-terminal ions were verified manually.
Five microliters of bacterial culture (wild-type, ΔcomFA or ΔcomEC) were diluted in 45 µL of Tris 10 mM, pH 8, NaCl 150 mM. Bacteriophage lambda DNA (0,1 mg/ml final) was then added to bacteria. Five µL of mix were immediately adsorbed onto a 600 mesh copper grid coated with a thin carbon film, activated by glow-discharge. After 1 min, grids were washed with 0,02% (w/vol) uranyl acetate solution (Merck, France) and then dried with filter paper. TEM observations were carried out with a Zeiss 912AB transmission electron microscope in filtered crystallographic dark field mode. Electron micrographs were obtained using a ProScan 1024 HSC digital camera and Soft Imaging Software system.
|
10.1371/journal.pcbi.1000266 | Partial Order Optimum Likelihood (POOL): Maximum Likelihood Prediction of Protein Active Site Residues Using 3D Structure and Sequence Properties | A new monotonicity-constrained maximum likelihood approach, called Partial Order Optimum Likelihood (POOL), is presented and applied to the problem of functional site prediction in protein 3D structures, an important current challenge in genomics. The input consists of electrostatic and geometric properties derived from the 3D structure of the query protein alone. Sequence-based conservation information, where available, may also be incorporated. Electrostatics features from THEMATICS are combined with multidimensional isotonic regression to form maximum likelihood estimates of probabilities that specific residues belong to an active site. This allows likelihood ranking of all ionizable residues in a given protein based on THEMATICS features. The corresponding ROC curves and statistical significance tests demonstrate that this method outperforms prior THEMATICS-based methods, which in turn have been shown previously to outperform other 3D-structure-based methods for identifying active site residues. Then it is shown that the addition of one simple geometric property, the size rank of the cleft in which a given residue is contained, yields improved performance. Extension of the method to include predictions of non-ionizable residues is achieved through the introduction of environment variables. This extension results in even better performance than THEMATICS alone and constitutes to date the best functional site predictor based on 3D structure only, achieving nearly the same level of performance as methods that use both 3D structure and sequence alignment data. Finally, the method also easily incorporates such sequence alignment data, and when this information is included, the resulting method is shown to outperform the best current methods using any combination of sequence alignments and 3D structures. Included is an analysis demonstrating that when THEMATICS features, cleft size rank, and alignment-based conservation scores are used individually or in combination THEMATICS features represent the single most important component of such classifiers.
| Genome sequencing has revealed the codes for thousands of previously unknown proteins for humans and for hundreds of other species. Many of these proteins are of unknown or unclear function. The information contained in the genome sequences holds tremendous potential benefit to humankind, including new approaches to the diagnosis and treatment of disease. In order to realize these benefits, a key step is to understand the functions of the proteins for which these genes hold the code. A first step in understanding the function of a protein is to identify the functional site, the local area on the surface of a protein where it affects its functional activity. This paper reports on a new computational methodology to predict protein functional sites from protein 3D structures. A new machine learning approach called Partial Order Optimum Likelihood (POOL) is introduced here. It is shown that POOL outperforms previous methods for the prediction of protein functional sites from 3D structures.
| Development of function prediction capabilities is a major challenge in genomics. Structural genomics projects are determining the 3D structures of expressed proteins on a high throughput basis. However, the determination of function from 3D structure has proved to be a challenging task; the functions of most of these structural genomics proteins remain unknown. Computationally based predictive methods can help to guide and accelerate functional annotation. The first step toward the prediction of the function of a protein from its 3D structure is to determine its local site of interaction where catalysis and/or ligand recognition occurs. Such capabilities have many important practical implications for biology and medicine.
We have reported on THEMATICS [1]–[4], for Theoretical Microscopic Titration Curves, a technique for the prediction of local interaction sites in a protein from its three-dimensional structure alone. In the application of THEMATICS, one begins with the 3D structure of the query protein, solves the Poisson-Boltzmann (P-B) equations using well-established methods, then performs a hybrid procedure to compute the proton occupations of the ionizable sites as functions of the pH. Residues involved in catalysis and/or recognition have different chemical properties from ordinary residues. In particular, these functionally important residues have anomalous theoretical proton occupation curves. THEMATICS exploits this difference and utilizes information from the shapes of the theoretical titration curves of the ionizable residues, as calculated approximately from the computed electrical potential function.
THEMATICS utilizes only the 3D structure of the query protein as input; neither sequence alignments nor structural comparisons are used. Recently it was shown [4] that, among the methods based on the 3D structure of the query protein only, THEMATICS achieves by far the best performance, as measured by sensitivity and precision for annotated catalytic residues.
The purpose of the present paper is five-fold: (1) We present a monotonicity-constrained maximum likelihood approach, called Partial Order Optimum Likelihood (POOL), to improve performance and expand the capabilities of active site prediction. (2) Then it is shown that POOL, with THEMATICS input data alone, outperforms previous statistical [4] and Support Vector Machine (SVM) [5] implementations of THEMATICS when applied to a test set of annotated protein structures. (3) It is then demonstrated that the inclusion of one additional 3D-structure-based feature, representing the ordinal size of the surface cleft to which each residue belongs, can result in some improved performance, as demonstrated by ROC curves and validated by Wilcoxon signed-rank tests. (4) With the introduction of environment features, POOL then can use the THEMATICS data to predict both ionizable and non-ionizable residues. This all-residues extension of THEMATICS, together with a cleft size rank feature, results in a simple 3D-structure-based functional site predictor that performs better than other 3D structure based methods and nearly as well as the very best current methods that utilize both the 3D structure and sequence homology. (5) Finally, the POOL approach is able to take advantage of sequence alignment-based conservation scores, when available, in addition to these structure-based features. When this additional information is included, the resulting classifier is shown to outperform all other currently available methods using any combination of structure and sequence information.
In prior implementations of THEMATICS for the identification of active-site residues from the 3D structure of the query protein [3]–[5], titration curve shapes were described by the moments of their first derivative functions. These first derivative functions are essentially probability density functions and give unity when integrated over all space. In Ko et al. [3], the third and fourth central moments μ3 and μ4 of these probability functions were used. These moments measure asymmetry (skewness) and, roughly, the area under the tails relative to the area near the mean (kurtosis), respectively. In Tong et al. [5], the first moment and second central moment were also used. In each of these approaches, the moments measure deviations from normal curve shape and the analyses identify the outliers, the residues that deviate most from the normal proton occupation behavior. These prior approaches all use spatial clustering, so that outlier residues are reported as positive by the method if and only if they are in sufficiently close spatial proximity to at least one other outlier. Thus the previous THEMATICS identifications involve two stages, where the first stage makes a binary (outlier / not an outlier) decision on each residue and the second stage finds spatial clusters of the outliers. In the new approach reported here, every residue is assigned a probability that it is an active-site residue. Here, as an alternative to the clustering approach, we introduce features that describe a residue's neighbors; we call these environment features. For a given scalar feature x, we define the value of the environment feature xenv(r) for a given residue r to be:(1)where r′ is an ionizable residue whose distance d(r′,r) to residue r is less than 9Å, and the weight w(r′) is given by 1/d(r′,r)2.
In this study, we use the same features μ3 and μ4 used in the Ko [3] approach, along with the additional features μ3env and μ4env. Thus every ionizable residue in any protein structure is assigned the 4-dimensional feature vector (μ3, μ4, μ3env, μ4env). The present approach has a number of advantages. Specifically, active residues may be selected in one step and they can be rank-ordered according to the probability of involvement in an active site. Furthermore, while THEMATICS previously has been applied to ionizable residues only, the present approach opens the door to direct prediction of non-ionizable active site residues, because the environment features μ3env and μ4env are well defined for all residues, including the non-ionizable ones. Finally, additional geometric features that are obtainable from the 3D structure only may be readily combined with the four THEMATICS features in order to enhance performance.
Geometric features, such as the relative sizes of the clefts on the surface of the protein structure, have been shown to correlate with active site location [6],[7]. For instance, for the majority of single-chain proteins, the catalytic residues are in the largest cleft. However geometric features alone do not perform comparatively well for active residue prediction, particularly because they are not very selective. It is shown here that cleft size information combined with THEMATICS electrostatic features yields high performance in purely 3D structure based functional site predictions.
The monotonicity-constrained maximum-likelihood approach underlying the POOL method described below is built on certain assumptions relating features used for classification to the probability that an instance having those features belongs to the positive class. Here we describe in detail the form these assumptions take when relating the THEMATICS features listed above to the probability that a residue is an active-site residue. Later we will also note that similar assumptions are reasonable when considering cleft rank and sequence conservation scores and apply them to those features as well. These THEMATICS feature-based monotonicity assumptions are as follows:
More precisely, for the first assumption, we treat μ3 and μ4 as measures of degree of perturbation, and for the second we treat μ3env and μ4env as measures of overall perturbation within the spatial vicinity. These assumptions then become: Given two residues in the same protein, let their corresponding 4-dimensional feature vectors be x = (x1, x2, x3, x4) and y = (y1,y2, y3, y4). If xi≤yi for each i, the probability that the first residue is an active-site residue is less than or equal to the probability that the second residue is an active-site residue. A more elegant formulation arises from the definition of a coordinate-wise partial order on the 4-dimensional feature space by x≤y iff xi≤yi for all i, and the above monotonicity assumptions then take the simple form x≤y implies P(active|x)≤P(active|y).
Finally, there is one additional subtlety that all implementations of THEMATICS have had to address, and the current approach is no exception: the need for some kind of normalization across proteins. In Ko's approach [3], the raw features were individually transformed into Z-scores, the deviations from the mean in units of the standard deviation, as calculated for the set of all ionizable residues within a given protein. Similarly, in Tong's SVM approach [5], the raw features were likewise transformed into robust Z-scores, defined as the deviations from the median in units of the interquartile distance for the set of all ionizable residues within a given protein. Here a very different type of transformation is applied to each feature across the population of residues within a given protein. We call this transformation rank normalization. Within each protein, each feature value is ranked from lowest to highest in that protein, and each data point is then assigned a number uniformly across the interval [0,1] based on the rank of that feature in that protein. The highest value for that feature is thus transformed to 1, and the lowest value is transformed to 0. Note that unlike the use of Z-scores or the robust Z-scores of Tong [5], this is a nonlinear transformation of the raw feature values. For each scalar feature x, denote its within-protein rank-normalized value as , which by definition lies in [0,1]. The use of this notation is extended to feature vectors in the obvious way, i.e. .
Note that this rank normalization transformation does not affect the within-protein partial order used in the assumptions. That is, x≤y is true for raw feature vectors x and y in the same protein if and only if . However, when data from multiple proteins is combined for training and the results are used to make predictions for new proteins, as is described in detail below, this actually implies an even stronger monotonicity assumption across proteins in which the within-protein rankings replace the raw feature values. This assumption is harder to justify intuitively, but such an approach is required to be able to train on multiple proteins and make predictions for novel proteins, and, as shown below, it appears to give good results.
After the normalization is performed, the labeled dataset may be regarded as a collection of pairs, one for each ionizable residue in the protein, where is the 4-dimensional rank-normalized feature vector for the ith residue and the label ci is either 1 (identified as an active-site residue) or 0 (not so identified). Given such a set of training data, the mathematical problem we wish to solve is to find a maximum likelihood estimator for as a function of in [0,1]4 based on this training dataset and satisfying the constraint that whenever in the coordinate-wise partial order described above. Letting n represent the number of training examples and pi the estimate of for each i from 1 to n, we seek to maximize:(2)subject to the constraints: for each (i,j) such that
This is a convex optimization problem with linear constraints. We have shown [8] that the solution to this convex optimization problem is the same as the solution to the quadratic programming problem of the minimization of:(3)subject to these same constraints. Note that while the equivalence of these solutions is well-known if there are no constraints, not every constrained maximum-likelihood problem is equivalent to the corresponding minimum squared-error problem with the same constraints. However, with these particular constraints, the two solutions are indeed identical.
This latter optimization problem is a special case of the general isotonic regression problem [9],[10] and this special form lends itself to a particularly straightforward solution technique. First, at an arbitrary point in the feasible region, the set of active constraints is determined by solving the corresponding dual quadratic programming problem of finding {λi} minimizingsubject to the constraints λi≥0 for all i, where is the negative gradient and are the normal vectors to the m constraint surfaces.
The kth constraint in the primal problem is active iff λk>0. Furthermore, by rescaling coordinates in the primal problem, its contours become circular and the negative gradient at any point points toward a single point, the unconstrained optimum. (Thus another formulation of the primal problem is to find the point in the feasible region closest to the unconstrained optimum.) As a consequence, the active set so determined at any feasible point is exactly the same as the active set at the solution point. But the active set simply represents equivalence classes of data points for which equality of the estimates must hold. Since equality-constrained maximum-likelihood estimates have the form (number of positives)/(total number of points), identifying which constraints are active at the solution leads immediately to the solution itself. Full details of this algorithm as well as the proof that the minimum sum-of-squared-errors solution is also the maximum-likelihood solution can be found in Tong's dissertation [8].
We call our algorithm for solving this maximum-likelihood problem the POOL algorithm. POOL is both an acronym for Partial Order Optimal Likelihood as well as an accurate characterization of the way the method first identifies the active constraints and then simply combines the corresponding data values into “pools” to be assigned probability estimates according to the proportion of positives in that pool.
The use of the POOL method with this 4-dimensional THEMATICS feature vector is denoted POOL(T4) in the Results section, where its performance is compared with other methods.
Previous studies have shown that active site residues tend to be located in one of the largest clefts in a protein structure [6],[7],[11]. Indeed it has been reported that in 83% of single-chain enzymes, the active site is located in the largest cleft [11]. Nearly all active sites are principally located in one of the five largest clefts of a protein structure, with the largest cleft containing the active site for the highest fraction of enzymes and with the fractions decreasing as the size rank progresses to smaller clefts [12]. If such purely geometric analyses were to be used for active site prediction, the result would be low precision and a high false positive rate, since the active site typically constitutes just a fraction of the area of the cleft. Since such geometric analyses are purely 3D-structure based, may be performed rapidly, and constitute a very different type of information from that of THEMATICS, it makes sense to combine these data in order to enhance overall performance. One straightforward way to do this is to combine the features from both THEMATICS and cleft size rank into a single vector of input to any appropriate classifier or probability estimator. Cleft size rankings may be readily incorporated into POOL, since there is an implicit monotonicity assumption that applies to this feature as well: The probability that a cleft contains an interaction site is highest for the largest cleft in a protein structure and decreases for clefts of smaller size rank. In this study we used CASTp [13], which uses computational geometry to define and measure pockets on the protein surface, to calculate cleft information for each residue in the protein. For present purposes, every residue in a given protein is assigned an integer number corresponding to the rank of the size of the cleft to which it belongs, where 1 is the largest, 2 is the second-largest, and so on. If the atoms of a residue belong to more than one cleft, the residue is assigned the rank of the largest of these clefts. When combined with THEMATICS features, the result is a 5-dimensional input vector to which coordinate-wise monotonicity constraints are applied on all five coordinates. POOL(T4,G) denotes the estimator resulting from applying the POOL method to this five-dimensional concatenation of the four THEMATICS features and one cleft size rank.
An interesting alternative to simply concatenating all features into a single vector and applying a single classifier or probability estimator to such vectors is to compute two separate probability estimates and then combine them. Consider the general problem of estimating the class probability P(c|x) for a feature vector x = (x1, x2, …, xk) formed as the concatenation of feature vectors x1, x2, …, xk. It is straightforward to show that if the naïve Bayes conditional independence assumption(4)holds for each class c, then(5)where α is a normalizing constant. This gives a computationally attractive way to consider combining probability estimates for combinations of feature sets when separate estimates are available for the individual feature sets. As with other applications of naïve Bayes, it is not necessary that the conditional independence assumption be strictly true for the results of this computation to give useful results, especially when it comes to relative rankings [14].
This then gives another approach, which we have dubbed chaining, to obtain active-site probability estimates using both THEMATICS features and cleft size rank. In this case, we use equation (5) to combine the POOL estimates based on THEMATICS with the POOL estimates for the one-dimensional cleft size rank feature. POOL estimates based on the four-dimensional THEMATICS input and those based on the one-dimensional cleft size rank are labeled POOL(T4) and POOL(G), respectively, where G stands for geometry. POOL(G) gives a simple set of active-site probabilities for each ranking. The probability estimator computed using equation (5) with POOL(T4) and POOL(G) we then call POOL(T4)xPOOL(G). Later we also incorporate a conservation score feature, based on sequence alignment, using this same technique.
Non-ionizable residues do not have titration curves and thus THEMATICS does not predict them directly. Nevertheless, the non-ionizable residues in interaction sites tend to have ionizable residues in their immediate vicinity and these ionizable residues generally have perturbed titration curves [1],[5]. This was the basis for the attempt by Tong et al. [5] to identify non-ionizable active site candidate residues by their proximity to the ionizable residues selected by THEMATICS. That approach, based on SVM results and called SVM-region, yields an unacceptably high false positive rate. Here we adopt a related strategy based on POOL and demonstrate substantially improved results.
Note that every non-ionizable residue has the environment features μ3env and μ4env; these serve as measures of the overall amount of titration curve perturbation in their spatial neighborhood. Thus we posit an extension to the THEMATICS monotonicity assumptions, namely: All other things being equal, a non-ionizable residue having more titration curve perturbation in its neighborhood is more likely to be an active-site residue. Thus we can apply the POOL method to non-ionizable residues separately by applying coordinate-wise monotonicity constraints to the probability estimates for the 2-dimensional feature vectors , once again using the transformations, rank-normalized within each protein, of these features. In this case, the rank normalization is performed separately on just the set of non-ionizable residues in a given protein.
Furthermore, we have the same options described above for incorporating cleft or other information for these non-ionizable residues. Finally, for any given protein, we can start with separate ordered lists of probability estimates for the ionizables and the non-ionizables, however computed, and then merge these into a single ordered list. This list then gives an estimated probability, and hence a ranking, for all residues.
Yet another feature that is generally taken to be predictive of functional activity in a monotonic fashion is the extent to which a given residue is found to be conserved across sequence homologues: The more conserved the residue, the more likely that residue is to be functionally important in the protein. Here we also examine combining such conservation information with THEMATICS and cleft size information. In particular, we use ConSurf [15], a sequence comparison based method that identifies functionally important regions on the surface of a protein of known three-dimensional structure, based on the phylogenetic relations between its close sequence homologues. If there are a sufficient number of sufficiently diverse homologues to the query protein, Consurf assigns a score between 1 and 9 to each residue in the query sequence based on how conserved this residue is among those homologues. The more conserved a residue is, the higher its score. We call this the conservation feature and in the results below we use C to denote its inclusion. Note that taken by itself, it is also a simple one-dimensional feature, like the cleft size rank. In this study, if Consurf returns more than 10 homologues, we use the score ConSurf assigns to each residue as its conservation feature value. For any proteins with 10 or fewer homologues, all residues in that protein are assigned a single common value for that feature. The effect of this is that all residues in such proteins have a tie for that feature, so it contributes nothing to the individual probability estimates or the residue rankings within that protein.
In the Results section it is shown that using information derived solely from the 3D structure, the present method outperforms all other 3D structure based methods. When sequence conservation information is included, the resulting classifier outperforms all other methods. Especially noteworthy is that in the absence of sequence conservation information, performance is nearly as good as that with such conservation information. This is particularly significant for structural genomics proteins, for which the present method is expected to perform well, even for novel folds and orphan sequences.
As described in more detail in the Materials and Methods section, the results presented in this paper are based on two sets of proteins, a set of 64 test proteins selected randomly from the CSA database [16],[17] and a 160-protein set covering most of the original CSA database. A detailed list of the names of the proteins, the PDB IDs of the structures, the E.C. classification, and the CSA-labeled positive residues within each protein in both test sets can be found in the Supporting Information; Dataset S1 contains the 64-protein test set and Dataset S2 contains the 160-protein test set. For each set of performance data reported here, the results are based on eight-fold cross-validation for the 64-protein set and ten-fold cross-validation for the-160 protein set.
The results presented here are based on several standard measures of performance. For a standard classification problem, performance is typically measured by recall (or true positive rate) and false-positive rate. Within a specific system with tunable parameters, recall and false positive rate typically involve a tradeoff: adjusting the parameters to lower the false-positive rate also lowers recall, while raising the latter also raises the former. So to judge the performance of such a system, it is important to know the tradeoff between these two, and thus ROC curves, which plot recall against false positive rate, are presented here. In the latter subsections, it is sometimes necessary to use other performance measures in order to compare our results against those reported by others.
Since our method outputs a ranked list (actually a list of probabilities) for all residues within a given protein and not a binary classification, considering the ROC curves is an especially useful way to characterize the behavior of any binary classification scheme derived from it. Among the many possibilities for creating a binary classification from such a list would be to select the top n or the top p percent or use a probability threshold. One advantage of ROC curves is that they are independent of the selection scheme.
One disadvantage of using ROC curves alone, however, is that unless the curve for one method dominates (i.e., lies completely above and to the left of) that of another, there may be no simple metric to compare these two methods. For this reason, a single number that is sometimes used as a reliable measure for comparing systems in the machine learning literature is the area under the ROC curve (AUC) [18]. We make use of this as a single numerical measure to which we can apply statistical significance tests to corroborate the apparent superiority of one method over another.
In order to generate ROC curves, we need to be able to calculate recall and false-positive rate values, which come from classification problems. In the POOL method, the result for each protein is a ranked list based on the probability of a residue being in the active site. A natural way to draw a ROC curve for every protein is to move the cutoff one residue at a time from the top to the bottom of the list. The resulting ROC curve has a staircase shape: only recall increases when an active site residue is encountered and only false positive rate increases when a non-active-site residue is encountered.
We define average specificity (AveS) for each protein in the set:(6)where r is the rank, N is the number of residues in a protein, pos(r) is a binary function that indicates whether the residue of a given rank r is annotated in the reference database in the active site (pos(r) = 1) or not (pos(r) = 0), and S(r) is the specificity at a given cut-off rank r. (Specificity is defined to be 1 – false positive rate.) It is not hard to see that AveS represents the area under the ROC curve (AUC) for that protein. We also compute the across-protein mean of AveS over a given set of proteins, which we call the mean average specificity (MAS) for that set.
To visually compare the performance from different methods, we also generate the averaged ROC curve for each method by computing the recall and false-positive rate after truncating the list after each of the positive residues in turn, followed by linearly interpolating the value at each recall value and computing the mean of the interpolated false-positive rate values across all proteins in the dataset.
From these average ROC curves we can get a strong sense of the apparent relative performance of different systems, but it is also important to be able to verify that such apparent differences are in fact statistically significant. To test the significance of the observed differences, we also perform the Wilcoxon signed-rank test [19] on AveS from these methods to estimate the probability of observing such a difference under the null hypothesis that the seemingly better-performing method is actually not better than the other. This test essentially determines which method is consistently better on a protein-by-protein basis (as measured by AUC or AveS), while the curves we display essentially demonstrate which methods perform better on average.
In this paper, we presented the application of the POOL method using THEMATICS plus some other features for protein active site prediction.
We started with the application of the POOL method just on THEMATICS features, with features similar to those used before in the SVM method [5], as well as those used in Ko and Wei's statistical analysis [3],[4]. These results show that the POOL method outperforms all of the earlier THEMATICS methods with no cleaning of the training data and no clustering after the classification. This suggests that by relying solely on the underlying THEMATICS monotonicity assumptions, the POOL method makes better use of the training data.
We also tested different ways of incorporating additional features into the learning system. Not surprisingly, the results show that in order to improve performance, we have to incorporate the right features in the right way. Even with features that were found to be helpful in improving the performance, how they are incorporated matters. The data show that chaining the results from separate POOL estimates is better than simply combining all the available features into a big POOL estimator over a higher-dimensional feature space. As mentioned earlier, the reason behind this might be overfitting, since combining features into a POOL table with high dimension causes the number of probabilities needed for estimation to grow exponentially, while the training data can only increase linearly in most cases. In other words, the high dimensionality makes the table too sparse and less accurate for probability estimates.
We also extended the application of THEMATICS to all residues, not just ionizable residues, in a natural way and showed that it is effective. Although the performance for non-ionizable residues is not as good as the performance for ionizable ones, this extension does provide a way to combine features from THEMATICS, which by itself can only be applied to ionizable residues directly, with some other features. The inclusion of the non-ionizable residues results in better overall performance and also makes performance comparison with other methods more accurate and fair.
The incorporation of sequence conservation information does improve the predictions when there are enough homologues with appropriate diversity. The POOL method gives us a means for easily utilizing this information when it is available, while not affecting the training and classification when it is not.
When comparing with other methods, especially if the other methods use binary classification instead of a ranked list, we have to commit to a specific cutoff value and turn our system into a binary classification system. The results in this paper clearly show that the POOL method using THEMATICS and geometric features achieves equivalent or better performance than the other methods in comparison, even in cases where their methods are tested on very special groups of proteins. This makes this method more widely applicable to proteins with few or no sequence homologues, such as some Structural Genomics proteins, than other methods that use sequence alignments from homologues. Performances of the previous best methods, those of Youn and of Petrova, will degrade significantly when sequence conservation information is not available. However with THEMATICS data the approach developed here is still robust in the absence of sequence conservation information. In effect, for those proteins having an insufficient number of sequence homologues, the POOL(T)xPOOL(G)xPOOL(C) method reduces to the still highly effective structure-only POOL(T)xPOOL(G) method.
Interestingly enough, when comparing the performance of POOL(T)xPOOL(G) and POOL(T)xPOOL(G)xPOOL(C) in Figure 4, it is apparent that the addition of the conservation information does improve the performance a little, but not to the extent observed previously for sequence-structure methods. Typically the conservation information is the most important input feature, and without it performance is substantially worse [24]. This suggests that the 3D structure based THEMATICS features are quite powerful compared with other 3D structure based features and can take the place of conservation information for purposes of active site prediction. This is also borne out by the analysis at the end of the Results section.
When looking at the recall and false positive rates of the results from all the protein active site prediction methods, one must keep in mind that the annotation of the catalytic residues in the protein dataset is never perfect. Since most of the labeling comes from experimental evidence, some active site residues are not labeled as positive simply because no experiment was ever carried out to verify the role of that specific residue. Since we have used the CatRes/CSA annotations as the sole criteria to evaluate the performance in order to keep the comparisons consistent, the reported false positive rate is probably higher than in reality. There is evidence available to support the functional importance of some residues that are not labeled as active in the CatRes/CSA database [3],[4] and these residues have high ranks in the list from the POOL method and are classified as positive by THEMATICS-SVM and THEMATICS-statistical analysis as well.
Although we evaluated the POOL method performance using filtration ratio values as a cutoff, it is just for the purpose of comparing with other protein active site prediction methods that use a binary classification scheme. The ranked list of residues based on their probability of being in the active site contains much more information than traditional binary classification. The rank of the first annotated positive residue analysis in this paper shows just one application of the extra information contained in a ranked list rather than a traditional binary label. There are many possible measurements of performance depending on the actual application by users, and in turn many possible applications that can benefit from using a ranked list form. It is noteworthy that P-Cats [25] also estimates the probability that a residue belongs to a protein active site, using a k-nearest neighbor method. The P-Cats server uses the probability estimates as the basis to assign binary labels; residues with probability larger than 0.50 are labeled as positive and the others as negative. The method of Cheng [26] also generates a rank-ordered list based on a scoring system; these scores could in principle be translated into probability estimates.
The POOL approach is amenable to the addition of other properties for the prediction of active sites [27]–[32]. We also note that the POOL methodology is applicable to other types of problems in a variety of different areas where probability depends monotonically on the input feature variables.
In conclusion, we have established that applying the POOL method, with THEMATICS and other features, appears to yield the best protein active site prediction system yet found and it provides more information than other active site prediction methods.
The three-dimensional coordinate files for the protein structures used for training and testing were downloaded from the Protein Data Bank (http://www.rcsb.org/pdb/). In order to predict the theoretical titration curve of each ionizable residue in the structure, finite-difference Poisson-Boltzmann calculations were performed using UHBD [33] on each protein followed by the program HYBRID [34], which calculates a corresponding titration curve of the form average net charge as a function of pH. These titration curves were obtained for each ionizable residue: Arg, Asp, Cys, Glu, His, Lys, Tyr, and the N- and C- termini. The pH range we simulated for all curves is from −15.0 to 30.0, in increments of 0.2 pH units. This wide theoretical pH range is necessary for proper numerical integration of the first derivative functions. The structures were processed and analyzed to obtain the central moments μ3 and μ4, as described earlier. These individual features were then rank-normalized within each protein, and thus assigned values in the interval [0,1], also as described above. This four-dimensional representation constitutes what we designate the THEMATICS features for each residue. The monotonicity assumptions for this multidimensional feature set are as described earlier.
For the geometric feature, we used CASTp [13], which uses a pocket algorithm for shape measurements to calculate the cleft information for each residue in the protein. The clefts were ranked based on their sizes in decreasing order and each residue having atoms located in any cleft is assigned the rank number of the largest of the clefts where its atoms are located. One special value is assigned to every residue not on the protein surface, and another is assigned to every residue on the surface but not within any cleft. Ignoring these special values, the monotonicity assumption is that the larger the cleft to which a residue belongs, the more likely that residue is to belong to the active site.
For the conservation feature we used ConSurf [15] to calculate a sequence conservation score for the residues in each protein. ConSurf takes a protein sequence and finds its closest sequence homologues using MUSCLE [35], a multiple-sequence alignment algorithm. Two sequences with similarity higher than a preset threshold are treated as homologues. ConSurf analyzes the homologues of the query sequence and determines how conserved each residue is in the query protein among these homologues. In order to normalize the result and make it comparable between different proteins with different numbers of homologues and with different degrees of overall conservation, the program labels each residue with a conservation score between 1 and 9, with 9 being the most conserved and 1 being the most variable. If there exist more than 50 homologues for the query sequence, the 50 homologues closest to the query sequence are analyzed. In this study, we only used the conservation score reported by ConSurf when there are at least 11 homologues for a protein. The monotonicity assumption applied to this feature is that the larger the conservation score for a residue, the more likely that residue is to belong to the active site.
The results reported here are based on eight-fold cross-validation on a set of 64 proteins or 10-fold cross-validation on a set of 160 proteins, both taken from the Catalytic Site Atlas (CSA) database [16],[17]. The labels were taken directly from the CSA database; if a residue is identified there as active in catalysis, it was labeled as positive in our dataset. If not so identified in the CSA, we labeled it as negative. The CSA annotations, although incomplete, constitute the best source of active residue labels for enzymes. In anticipation that the POOL method would not be overly sensitive to mislabeled data, no hand tuning of the labels was performed and no residues were omitted during training, in contrast to the SVM study reported by Tong [5].
For the eight-fold cross-validation procedure, the 64-protein set was randomly divided into eight folds of eight proteins each, with seven of the eight folds (56 proteins) used for training and the remaining fold (8 proteins) used for testing. This was repeated eight times, once for each of the eight folds. Likewise, for the ten-fold cross-validation procedure, the 160-protein set was randomly divided into ten folds of sixteen proteins each, with nine of these (144 proteins) used for training the remaining fold (16 proteins) used for testing, and this was repeated a total of ten times, once for each fold.
Training was performed by applying the POOL method to obtain a function for each rank-normalized feature vector in the appropriate feature space [0,1]k. Note that: k = 4 for the POOL method applied on the four THEMATICS features of ionizable residues as stated earlier, denoted by POOL(T4); k = 5 for the POOL method applied on the four THEMATICS features of ionizable residues plus the geometric feature of the cleft size, denoted as POOL(T4,G); k = 1 for the POOL method applied to the geometric feature of cleft size, denoted by POOL(G), as well as the POOL method applied to the conservation feature, denoted by POOL(C); and k = 2 for POOL applied to the environmental features for non-ionizable residues, denoted as POOL(T2).
An additional detail is that for training we quantize the multi-dimensional data points. For example, for POOL(T4), each rank-normalized feature falls into one of 20 bins whose sizes vary depending on their distance from 0.0. In particular, the lowest ranked bins cover the half-open intervals, [0.0, 0. 2), [0. 2, 0.4), [0.4, 0.6), [0.6, 0.7), and there are 16 more bins of width 0.02 above that, with one special bin for 1.0. Thus the lowest-ranking data are quantized more coarsely than the remaining data. This is appropriate since these data tend to have very low average probability of being in the active site anyway, because the vast majority of residues are negatives. Thus the inability to make fine distinctions among these low-probability candidates does not degrade the overall quality of the results. It does, however, improve the efficiency of the training procedure significantly, so this is an important component of the analysis. This is especially helpful in the 10-fold cross-validation on the 160-protein set. The typical training set of 144 proteins from this set contains about 14500 ionizable residues, which fall into more than 6000 quantized bins in the 4-dimensional space used for POOL(T4). The number of corresponding inequality constraints is about 35,000–40,000.
One final detail is that the probability estimates generated by the POOL method as described here tend to have numerous ties as well as some places where there is no well-defined value. The latter places occur because the method only assigns values to existing data points (or bins containing data in the case of our use of quantization). The locally constant regions occur both because of the quantization applied to the training data at the outset and because the data pools created by the algorithm acquire a single value. In cells where no value is defined, the interpolation scheme used is to simply assign a value linearly interpolated based on the Manhattan distance between the least upper bound and the greatest lower bound for that cell based on the monotonicity constraint. Finally, since both the data pooling performed by the algorithm and this interpolation scheme tend to lead to ties, the Manhattan distance from the origin of the four THEMATICS features is used as a tie-breaker for any residues whose probability estimates are identical. This simply imposes a slight bias toward strict monotonicity even though the mathematical formulation used to determine these probabilities is based on a non-strict monotonicity assumption, making it possible to obtain well-defined rankings for all the residues in a protein.
|
10.1371/journal.pgen.1005263 | Feeding and Fasting Signals Converge on the LKB1-SIK3 Pathway to Regulate Lipid Metabolism in Drosophila | LKB1 plays important roles in governing energy homeostasis by regulating AMP-activated protein kinase (AMPK) and other AMPK-related kinases, including the salt-inducible kinases (SIKs). However, the roles and regulation of LKB1 in lipid metabolism are poorly understood. Here we show that Drosophila LKB1 mutants display decreased lipid storage and increased gene expression of brummer, the Drosophila homolog of adipose triglyceride lipase (ATGL). These phenotypes are consistent with those of SIK3 mutants and are rescued by expression of constitutively active SIK3 in the fat body, suggesting that SIK3 is a key downstream kinase of LKB1. Using genetic and biochemical analyses, we identify HDAC4, a class IIa histone deacetylase, as a lipolytic target of the LKB1-SIK3 pathway. Interestingly, we found that the LKB1-SIK3-HDAC4 signaling axis is modulated by dietary conditions. In short-term fasting, the adipokinetic hormone (AKH) pathway, related to the mammalian glucagon pathway, inhibits the kinase activity of LKB1 as shown by decreased SIK3 Thr196 phosphorylation, and consequently induces HDAC4 nuclear localization and brummer gene expression. However, under prolonged fasting conditions, AKH-independent signaling decreases the activity of the LKB1-SIK3 pathway to induce lipolytic responses. We also identify that the Drosophila insulin-like peptides (DILPs) pathway, related to mammalian insulin pathway, regulates SIK3 activity in feeding conditions independently of increasing LKB1 kinase activity. Overall, these data suggest that fasting stimuli specifically control the kinase activity of LKB1 and establish the LKB1-SIK3 pathway as a converging point between feeding and fasting signals to control lipid homeostasis in Drosophila.
| Liver kinase B1 (LKB1), a serine/threonine kinase, controls 14 different AMP-activated protein kinase (AMPK) family kinases, including salt-inducible kinase 3 (SIK3), suggesting that it plays a variety of roles. Using the fruit fly as an in vivo model system, we reveal that LKB1 kinase activity is critical for lipid storage and controls the lipolysis pathway in the fat body, which is equivalent to mammalian adipose and liver tissue. We find that the lipolytic defects of LKB1 mutants are rescued by the expression of constitutively active SIK3 in the fat body. We show that LKB1 and SIK3 regulate lipid storage by altering the gene expression of brummer, the Drosophila homolog of human adipose triglyceride lipase (ATGL), a critical lipolytic gene. We also identify that LKB1-SIK3 signaling controls the nuclear and cytosolic localization of the class IIa deacetylase HDAC4 via SIK3-dependent phosphorylation in feeding and fasting conditions, respectively. Collectively, these data suggest that the LKB1-SIK3-HDAC4 pathway plays a critical role in maintaining fly lipid homeostasis in response to dietary conditions.
| Perturbation of energy homeostasis either directly or indirectly causes human health problems such as obesity and type II diabetes [1]. Lipid stores are the major energy reserves in animals and are dynamically regulated by alternating between the lipogenesis and lipolysis cycles in response to food availability. Dissecting the regulatory mechanisms of lipid homeostasis is therefore essential to our understanding of how energy metabolism is maintained.
Drosophila has emerged as a useful genetic model organism for studying lipid homeostasis and energy metabolism [2]. Drosophila lipid reserves are mainly stored as triacylglycerol (TAG) in the fat body, the insect equivalent of mammalian adipose tissue. In addition, lipolytic factors are evolutionarily conserved between insects and mammals. Brummer (Bmm) is the Drosophila homolog of ATGL, a key regulator of lipolysis. bmm mutant flies are obese and display partial defects in lipid mobilization [3]. Furthermore, hormonal regulation of lipid metabolism is also highly conserved in Drosophila. Under starvation conditions, the primary role of AKH, the functional analogue of glucagon and β-adrenergic signaling in mammals [4,5], is to stimulate lipid mobilization by activating AKH receptor (AKHR) [6] and consequently inducing cAMP/PKA signaling in the fat body [7]. A report demonstrated that AKH acts in parallel with Bmm to regulate lipolysis and that AKHR mutation leads to obesity phenotypes and defects in fat mobilization [7]. However, bmm expression is hyperstimulated in starved AKHR mutants [7], implying the existence of an unknown regulatory mechanism between Bmm and AKHR in Drosophila.
LKB1 (liver kinase B1, also known as STK11) is a serine/threonine kinase that was first identified as a tumor suppressor gene associated with Peutz-Jeghers syndrome [8,9]. LKB1 phosphorylates and activates AMP-activated protein kinase (AMPK) in response to cellular energy status, thus controlling cell metabolism, cell structures, apoptosis, etc. [10–13]. Moreover, LKB1 is the master upstream protein kinase for 12 AMPK-related kinases, including salt-inducible kinases (SIKs) [14], suggesting that it plays diverse roles. Although the metabolic functions of AMPK have been highly studied, the in vivo functions of LKB1 and AMPK-related kinases in metabolism, including lipid homeostasis, are still largely unknown [15]. Recent reports showed that LKB1 is required for the growth and differentiation of white adipose tissue [16] and that SIK3 maintains lipid storage size in adipose tissues [17]. In addition, we and others determined that Drosophila SIK family kinases regulate lipid levels and starvation responses [18,19]. However, to further understand the roles and mechanisms of LKB1 signaling in lipid metabolism, proper genetic animal models are urgently required.
Here we demonstrate the role of LKB1 and its downstream SIK3 in the regulation of lipid homeostasis using Drosophila as an in vivo model system. We demonstrated that LKB1-activated SIK3 regulates the nucleocytoplasmic localization of HDAC4 to control lipolytic gene expression. We also identified that DILPs modulate SIK3 activity via Akt-dependent phosphorylation and the AKH pathway regulates LKB1 activity in phosphorylating SIK3 to control its lipolytic responses upon short-term fasting. Furthermore, we identified that AKH-independent signaling modulates the LKB1-SIK3-HDAC4 pathway upon prolonged fasting. Altogether, these studies showed that the LKB1-SIK3 signaling pathway plays a crucial regulatory role in maintaining lipid homeostasis in Drosophila.
LKB1 functions in a complex with two scaffolding proteins, STE20-related adaptor (STRAD) and mouse protein 25 (MO25) [20,21]. As the first step toward elucidation of the role of LKB1 in lipid metabolism, we demonstrated the gene expression of each component of the LKB1 complex in the fat body (Fig 1A), suggesting that Drosophila LKB1 forms the heterotrimeric complex when activated in tissues. Additionally, we characterized an LKB1-null mutant line, LKB1X5 [22], and found that these flies showed markedly decreased lipid storage compared to wild-type flies, despite having similar food intake and retaining expression of the lipogenic genes (SREBP, FAS, and ACC) (Figs 1B, 1C and S1A). However, expression of bmm and lipolysis activity were elevated in LKB1X5 mutants (Figs 1C and S1B, respectively). Moreover, transgenic expression of wild-type LKB1 with two different fat body drivers (FB-Gal4 and cg-Gal4) rescued the decreased lipid levels and increased bmm expression phenotypes of LKB1X5 mutants, whereas expression of the kinase-dead form of LKB1 (LKB1 K201I) did not (Figs 1E, 1F, S2A and S2B). Additionally, overexpression of LKB1 induced significant increases in the lipid levels and decreases in bmm expression in a dose-dependent manner (S3A and S3B Fig). The implication behind these observations is that LKB1 plays a critical role in lipid storage in Drosophila by regulating the lipolysis pathway in a kinase activity-dependent manner.
To identify the lipolytic target of LKB1 among AMPK-related kinases in Drosophila, we determined mRNA levels of SIKs and AMPK, which are heavily involved in various metabolic pathways. As shown in Fig 1D, SIK3 and AMPKalpha were more highly expressed in the fat body. Furthermore, transgenic expression of constitutively active SIK3 (SIK3 T196E) in the fat body rescued the lipid accumulation and bmm expression defects of LKB1-null mutants, whereas expression of constitutively active AMPK (AMPK T184D) or inactive SIK3 with a mutation in the LKB1 phosphorylation site (SIK3 T196A) failed to rescue the lipid levels of the null mutants (Figs 1E, 1F, S4A and S4B). These results clearly suggest a specific role for SIK3 in the LKB1-mediated regulation of lipid storage in the fat body of Drosophila. Supporting this conclusion, overexpression of LKB1 highly augmented the phosphorylation of conserved Thr196 in SIK3 (Fig 1G), but this phosphorylation was completely lost in LKB1X5 mutants (Fig 1H).
Drosophila SIK3, one of the AMPK-related kinases, shares considerable sequence homology with the kinase domain of mammalian SIK3 (Fig 2A). To assess the in vivo role of SIK3, SIK3 loss-of-function mutants were generated by mobilizing the EP-element from SIK3G7844 (Fig 2B). From 600 EP excision alleles, we generated SIK3Δ5–31 mutant, which lacks 2,476 bp (2R14578001~14580477) that encodes for the translation start site and the ATP-binding site of SIK3 (Fig 2B and 2C). Confirming that SIK3Δ5–31 is a null mutant, SIK3 mRNA was not detected in the mutant (Fig 2D). However, the internal gene (CG15071) in the coding region of SIK3 was not affected (Fig 2D).
SIK3Δ5–31 mutant flies died before the mid-pupal stage and showed a decreased survival rate (Fig 2E and 2F). The SIK3 null mutant also exhibited a lipodystrophic phenotype (Fig 2G), and FB-Gal4-driven EGFP expression further confirmed the lean fat body phenotype of SIK3Δ5–31 mutants compared to control flies (Fig 2H). Consistently, SIK3Δ5–31 mutant had decreased lipid stores despite having a similar food intake in the larval stage (Figs 2I and S1A, respectively). Surprisingly, the fat body-specific expression of exogenous wild-type SIK3 rescued the lethality of SIK3Δ5–31 mutant, while the expression of a kinase-dead SIK3 (SIK3 K70M) failed to rescue the mutant (Fig 2E and 2F). These results demonstrated that the phosphotransferase activity of SIK3 in the fat body is crucial for its function.
To further investigate the role of SIK3 in lipid metabolism, we analyzed bmm gene expression in SIK3Δ5–31 mutant. Expectedly, the mutant showed markedly increased expression of bmm and increased lipase activity (Figs 2J and S1B, respectively), a phenotype similar to the LKB1 null mutant. Transgenic expression of either wild-type (SIK3 WT) or constitutively active SIK3 (SIK3 T196E) in the fat body of SIK3Δ5–31 mutant resulted in full recovery of lipid levels and bmm expression compared to wild-type controls (Figs 2K, 2L, S2C and S2D). In contrast, expression of either inactive SIK3 harboring a mutation in the LKB1 phosphorylation site (SIK3 T196A) or a kinase-dead mutant (SIK3 K70M) failed to rescue the SIK3Δ5–31 mutant phenotypes (Fig 2K and2L, respectively). Therefore, the kinase activity of SIK3 controlled by LKB1 is critical for the lipid storage in Drosophila fat body.
LKB1 and AMPK-related kinases play a major role in the inhibition of hepatic gluconeogenesis in response to high glucose levels via phosphorylation of the class IIa HDACs and the CREB co-activator CRTC [23–25]. To test whether HDAC4 or CRTC is involved in the LKB1 and SIK3 pathway, we analyzed the genetic interactions of LKB1 and SIK3 with HDAC4 and CRTC in Drosophila. We found that ablation of CRTC exacerbated the lethality of LKB1 and SIK3 null mutants (S5A and S5B Fig). However, strikingly, the loss of HDAC4 rescued the lethality of SIK3 null mutants, but did not affect the lethality of LKB1 null mutants (S6A–S6D Fig), suggesting that HDAC4 participates in LKB1-SIK3 signaling of Drosophila.
To evaluate whether HDAC4 is crucial for the regulation of lipid storage by LKB1 and SIK3, we expressed HDAC4 RNAi in the fat body of LKB1 and SIK3 mutants. Surprisingly, knockdown of HDAC4 in the fat body fully rescued the TAG levels and bmm gene expression of LKB1 and SIK3 null mutants (Fig 3A and 3B, respectively), indicating that HDAC4 is indeed a critical downstream target of LKB1 and SIK3 in lipid metabolism of Drosophila.
SIKs can regulate target gene expression by directly phosphorylating the class IIa HDACs and consequently inhibiting their translocation to the nucleus [26,27]. Expression of wild-type SIK3 (SIK3 WT) or constitutively active SIK3 (SIK3 T196E) augmented the phosphorylation of HDAC4 but not of the phosphorylation-defective HDAC4 (HDAC4 3A), demonstrating that SIK3 induces HDAC4 phosphorylation in Drosophila (Fig 3C). HDAC4 localized to both the cytoplasm and nuclei of larval fat body cells under feeding conditions, but localized mostly to the nucleus under fasting conditions (Fig 3D). However, HDAC4 accumulated in the nuclei of the fat body cells of LKB1 and SIK3 null mutants even under feeding conditions (Fig 3D). In addition, HDAC4 3A was retained in the nuclei of the fat body cells under both feeding and fasting conditions (Fig 3D). These results indicated that Drosophila SIK3, under the control of LKB1, phosphorylates HDAC4 in the fat body and regulates its nucleocytoplasmic localization in different dietary conditions.
The class IIa HDACs deacetylate and activate FOXO transcription factors [19,24], and the activated FOXO then induces ATGL/Bmm expression [19,28]. Overexpression of wild-type HDAC4 increased the mRNA levels of bmm (Fig 3E), indicating that HDAC4 regulates bmm gene expression in the fat body. Furthermore, overexpression of constitutively active SIK3 completely blocked the increased bmm expression induced by HDAC4 overexpression (Fig 3E), and bmm knockdown in the fat body blocked the decreases in TAG levels induced by LKB1 or SIK3 null mutation (Fig 3F and 3G). Altogether, these results suggested that the LKB1-SIK3 signaling pathway controls HDAC4-dependent Bmm activity in Drosophila fat body.
Under fasting conditions, AKH activates the mobilization of fat body triglyceride by triggering AKHR and consequent activation of cAMP signaling in the fat body [7]. Consistently, we showed that AKHR mutation highly increased TAG levels (Fig 4A) and decreased bmm gene expression (Fig 4B). To determine the functional interaction between AKHR signaling and the LKB1-SIK3 signaling pathway, we crossed LKB1 or SIK3 null mutant flies with AKHR mutant flies. Interestingly, deletion of LKB1 or SIK3 reversed both the lipid accumulation and the reduced bmm expression phenotypes of AKHR mutant flies (Fig 4A and 4B, respectively), suggesting that the LKB1-SIK3 pathway likely acts downstream of AKHR. Furthermore, SIK3 Thr196 phosphorylation was reduced in both fasting and AKH overexpression conditions compared to that in feeding conditions (Fig 4C), supporting that the AKH pathway inhibits the kinase activity of LKB1 as shown by decreased SIK3 Thr196 phosphorylation.
On the basis of the observation that fasting induces the nuclear translocation of HDAC4, we also examined subcellular localization of HDAC4 in AKHR mutant flies. Intriguingly, HDAC4 in AKHR mutants localized to both the cytoplasm and the nucleus of larval fat body cells in 4 hr fasting condition compared to control (Figs 3D and 4D), suggesting that AKHR-dependent regulation is critical for HDAC4 localization. In addition, overexpression of the phosphorylation-defective HDAC4 by SIK3 (HDAC4 3A) suppressed the TAG levels and enhanced bmm expression in AKHR mutants (Fig 4E and 4F, respectively). Collectively, these results suggested that the LKB1-SIK3-HDAC4 pathway acts downstream of the AKH pathway to control lipolysis activity in Drosophila.
We showed that AKHR-dependent regulation of LKB1-SIK3 activity is critical for HDAC4 nuclear localization in ~4 hr fasting condition (Fig 4D). However, Gronke et al. showed that bmm gene expression is stimulated in AKHR mutant flies in 6 hr fasting condition [7]. Notably, in contrast to 4 hr fasting condition (Fig 4D), HDAC4 accumulated in the nuclei of the fat body cells in AKHR mutant flies after prolonged fasting (~10 hr) (Fig 5A), indicating that there should be AKHR-independent HDAC4 regulation during prolonged fasting. Furthermore, knockdown of HDAC4 in the fat body blocked the increased bmm gene expression in AKHR mutant flies after 10 hr fasting (Fig 5B), indicating that AKHR-independent signaling promotes HDAC4 nuclear localization to induce bmm gene expression under prolonged fasting conditions. Interestingly, expression of constitutively active SIK3 blocked the prolonged fasting-induced nuclear localization of HDAC4 (Fig 5C), suggesting that LKB1-SIK3 activity is critical for bmm expression under prolonged fasting. Taken together, these results demonstrated that the LKB1-SIK3-HDAC4 pathway acts as the primary lipolytic signaling upon both short-term and prolonged fasting while AKH plays a major role only in short-term fasting. Thus, it is of particular interest to investigate novel signaling mechanisms regulating the LKB1-SIK3-HDAC4 pathway under prolonged fasting conditions.
Our study provides evidence that LKB1 is necessary for maintaining Drosophila lipid storage via the regulation of lipolysis through the activation of SIK3. Consistent with our results in Drosophila, adipose tissue-specific LKB1 knockout mice showed decreased serum triglycerides [16], and the basal lipogenesis activity of adipocytes was significantly lower in LKB1 hypomorphic mice [29]. Recently, SIK3 null mice were also found to display a malnourished phenotype with lipodystrophy and were resistant to high-fat diets [17]. Thus, the LKB1-SIK3 pathway is indeed an evolutionally conserved regulatory mechanism for lipid homeostasis.
LKB1 is ubiquitously expressed and constitutively active in mammalian cells [15], which raises the question of how dietary conditions change the activity of LKB1 and SIK3 to control lipid homeostasis. Our findings suggested that fasting and the AKH pathway inhibit LKB1 activity to regulate SIK3 Thr196 phosphorylation (Figs 4C and 6C). It is possible that fasting- and AKH-induced inhibition of LKB1 activity can be achieved by altered subcellular localization, protein conformation, stability, and/or protein-protein interactions of LKB1 and its associated proteins. Interestingly, in HEK-293 cells, fasting triggers autophosphorylation of human LKB1 at Thr336 [30] that corresponds to Thr460 in Drosophila LKB1 [31]. This phosphorylation promotes the protein-protein interaction between LKB1 and 14-3-3 proteins [30] and inhibits the ability of LKB1 for suppressing cell growth [31].
In addition, the AKH pathway activates cAMP/PKA signaling in Drosophila [7]. Mammalian PKA inhibits SIK activity by phosphorylating a conserved serine residue [32,33] that corresponds to Ser563 in Drosophila SIK3 [19], suggesting that the AKH pathway also controls SIK3 activity via PKA-dependent phosphorylation (Fig 6C). On the other hand, the Drosophila insulin-like peptides (DILPs) did not increase SIK3 Thr196 phosphorylation (Fig 6A), but induced Akt-mediated SIK3 phosphorylation (Fig 6B), suggesting that DILPs directly regulate SIK3 activity independently of affecting LKB1 activity [19,34] (Fig 6C). Interestingly, these Drosophila signaling circuits are highly similar to mammalian insulin and glucagon pathways in controlling lipid metabolism and storage, raising questions regarding whether the LKB1-SIK3-HDAC4 signaling pathway is also conserved in mammalian systems as a converging point between feeding and fasting signals to control lipid homeostasis.
Is SIK3 also involved in the modulation of other LKB1 functions, such as the regulation of cell polarity and mitosis? SIK3 null mutants showed normal epithelial polarity and mitosis (S7A and S7B Fig). Additionally, transgenic expression of constitutively active SIK3 (SIK3 T196E) failed to suppress the cell polarity and mitosis defects of LKB1 mutants (S8A and S8B Fig), suggesting that SIK3 does not participate in the regulation of cell polarity and mitosis by LKB1. In addition, both fat body-specific expression of LKB1 and ablation of HDAC4 failed to rescue the lethality of LKB1 null mutants (S6C Fig), indicating that LKB1 has SIK3/HDAC4-independent roles and additional targets in other tissues and developmental processes.
In summary, we have demonstrated that the LKB1-SIK3 pathway is important for maintaining lipid homeostasis in Drosophila. As alterations in lipolysis are closely associated with human obesity [35], future studies will be required to unravel the relationship between LKB1-SIK3-HDAC4 signaling and obesity-related metabolic diseases.
The following fly stocks were used in this study: LKB1X5, UAS-LKB1WT, and UAS-LKB1KI [22], HDAC4KG09091 (Bloomington #15159), UAS-HDAC4 RNAi (VDRC #20522), UAS-bmm RNAi (Bloomington #25926), UAS-InRCA (Bloomington #15159), cg-Gal4 (Bloomington #7011), hs-Gal4 (Bloomington #1799), UAS-2xEGFP (Bloomington #6874), UAS-HA-AMPKTD [12], CRTC25-3 [36], UAS-FLAG-HDAC4WT and UAS-FLAG-HDAC43A [19], AKHR1 [7], and FB-Gal4 [37]. SIK3Δ5–31 was generated by imprecise excision of SIK3G7844 line (KAIST Drosophila Library Facility, Daejeon, Korea). To generate UAS-SIK3 flies, SIK3 EST cDNA (Berkeley Drosophila Genome Project accession no. LD07105) was cloned into the Myc-tagged pUAST vector and microinjected into w1118 embryos. All flies were grown on food containing approximately 35 g cornmeal, 70 g dextrose, 5 g agar, 50 g dry active yeast (Ottogi, Inc., Korea), 4.6 ml propionic acid, and 7.3 ml Tegosept (100 g/l in ethanol) per liter at 25°C. All flies were backcrossed for a minimum of 6 generations into w1118 background.
The QuickChange kit (Stratagene) was used for site-directed mutagenesis. For generation of a kinase-dead mutant SIK3 (Lys70Met, SIK3K70M), 5’-CAAGACAAAGGTGGCCATCATGATCATAGACAAAACATGTC-3’ and 5’-GACATGTTTTGTCTATGATCATGATGGCCACCTTTGTCTTG-3’ primers were used. For generation of a SIK3 mutant non-phosphorylatable by LKB1 (Thr196Ala, SIK3 T196A), 5’-GGGTGCCACCTTAAAAGCTTGGTGTGGATCAC-3’ and 5’-GTGATCCACACCAAGCTTTTAAGGTGGCACCC-3’ primers were used. For generation of a SIK3 mutant mimicking LKB1-dependent phosphorylation (Thr196Glu, SIK3 T196E), 5’-GAGGGTGCCACCTTAAAAGAATGGTGTGGATCACCGCCC-3’ and 5’-GGGCGGTGATCCACACCATTCTTTTAAGGTGGCACCCTC-3’ primers were used.
Larvae were collected, and RNA was extracted using the RNeasy Mini Kit (QIAGEN). Total RNA (1 μg) was reverse-transcribed by M-MLV Reverse Transcriptase (Promega) to generate cDNA for quantitative real-time RT-PCR (Bio-Rad CFX96 Real-Time PCR detection system, SYBR Green) using a 500 nM primer concentration and 2 ng of cDNA template. The primers were used in S1 Table. The relative values were calculated using the ΔΔCt method via normalization to rp49 mRNA levels. Results were expressed in arbitrary units, with each control value as 1 unit.
Third instar larvae were dissected in Drosophila Ringer’s solution and fixed with 4% formaldehyde in phosphate buffered saline (PBS) for 10 min at room temperature. After being washed with 0.1% Triton X-100 in PBS (PBST), the samples were blocked for 1 hr incubation at room temperature with 5% bovine serum albumin (BSA) in PBST. The samples were further incubated at 4°C for 16 hr with the indicated antibodies: anti-FLAG-M2 (Sigma, F1804), anti-Myc (Cell Signaling Technology, #2272), anti-aPKC (Santa Cruz, sc-216), and anti-PH3 (Millipore, 06–570). Following three washes with PBST, the samples were incubated with appropriate secondary antibodies (and with Hoechst 33258 used for staining DNA, if required) for 3 hr at room temperature. The samples were washed with PBST and mounted with 80% glycerol in PBS, then observed by a confocal microscope LSM710 (Zeiss).
Feeding assay was performed according to previously described with minor modifications [38]. Blue food dye (Erioglaucine Disodium Salt, Sigma, #861146) was added at 1% (w/v) to fly food. Larvae were switched from normal food to blue-color food for 2 hr. After feeding, larvae were frozen immediately. Samples were homogenized in PBS buffer and centrifuged for 25 min at 13,200 rpm. The absorbance of the supernatant was measured at 625 nm using Infinite M200 spectrophotometer (Tecan).
Larvae were lysed in a lysis buffer (20 mM Tris-HCl (pH 7.5), 1 mM EDTA, 5 mM EGTA, 150 mM NaCl, 20 mM NaF, 1% Triton X-100, 1 μg/ml leupeptin, and 1mM PMSF) for 30–60 min on ice. After centrifugation for 15 min at 13,200 rpm, supernatants were reserved for SDS-PAGE analysis, and proteins were then transferred to nitrocellulose membranes (GE Healthcare, #BA85). Membranes were incubated in a blocking solution (Tris-buffered saline (TBS) containing 0.1% Tween-20, 5% BSA) for 1hr. The primary antibodies used were anti-LKB1 [22], anti-phospho-Thr196 SIK3 [39], anti-phospho-Ser239 HDAC4 (Cell Signaling Technology, #3443), anti-phospho-Akt substrate (Cell Signaling Technology, #9614), anti-FLAG-M2 (Sigma, #F1804), anti-Myc (Cell Signaling Technology, #2272), anti-AKH (a gift from Dr. Veenstra), and anti-β-tubulin antibody (Developmental Studies Hybridoma Bank, E7). Protein detection was done using the LAS-4000 imaging system (Fujifilm), and densitometric analysis was performed using Multi Gauge 3.0 software.
TAG measurement was performed according to previously described methods using Free Glycerol Reagent (Sigma, #F6428) and Triglyceride Reagent (Sigma, #T2449) [40]. A standard curve was generated with a glycerol standard solution (Sigma, #G7793). Samples were assayed at 540 nm using Infinite M200 spectrophotometer (Tecan). In each homogenate, amounts of TAG (in mg) were normalized to those of protein (in mg) using Bradford protein assay (Bio-Rad). Lipase activity was determined according to the manufacturer’s instructions with QuantiChrom kit (BioAssay Systems, DLPS-100).
The analysis for survival rate was performed as previously described [12]. The eggs from flies with the appropriate genotypes were laid on 60 mm dishes containing standard apple juice-agar with yeast paste for 4 hr. The hatched larvae were collected using selection markers and transferred to plates containing normal food media. The green balancer chromosome (CyO, Actin-GFP) was used to select the homozygous SIK3Δ5–31 mutant. The number of larvae was scored for viability at each developmental stage, and dead larvae were removed. At least 100 larvae were studied per genotype.
All quantitative data are analyzed using Student’s t tests or ANOVA with a post Tukey’s multiple comparison test, and P < 0.05 was considered statistically significant. Each experiment was repeated at least three times, data are presented as the average ± standard error of the mean (SEM). The P values given in the survival data are the result of a log rank test using GraphPad Prism 5 software.
|
10.1371/journal.pgen.1000693 | FON2 SPARE1 Redundantly Regulates Floral Meristem Maintenance with FLORAL ORGAN NUMBER2 in Rice | CLAVATA signaling restricts stem cell identity in the shoot apical meristem (SAM) in Arabidopsis thaliana. In rice (Oryza sativa), FLORAL ORGAN NUMBER2 (FON2), closely related to CLV3, is involved as a signaling molecule in a similar pathway to negatively regulate stem cell proliferation in the floral meristem (FM). Here we show that the FON2 SPARE1 (FOS1) gene encoding a CLE protein functions along with FON2 in maintenance of the FM. In addition, FOS1 appears to be involved in maintenance of the SAM in the vegetative phase, because constitutive expression of FOS1 caused termination of the vegetative SAM. Genetic analysis revealed that FOS1 does not need FON1, the putative receptor of FON2, for its action, suggesting that FOS1 and FON2 may function in meristem maintenance as signaling molecules in independent pathways. Initially, we identified FOS1 as a suppressor that originates from O. sativa indica and suppresses the fon2 mutation in O. sativa japonica. FOS1 function in japonica appears to be compromised by a functional nucleotide polymorphism (FNP) at the putative processing site of the signal peptide. Sequence comparison of FOS1 in about 150 domesticated rice and wild rice species indicates that this FNP is present only in japonica, suggesting that redundant regulation by FOS1 and FON2 is commonplace in species in the Oryza genus. Distribution of the FNP also suggests that this mutation may have occurred during the divergence of japonica from its wild ancestor. Stem cell maintenance may be regulated by at least three negative pathways in rice, and each pathway may contribute differently to this regulation depending on the type of the meristem. This situation contrasts with that in Arabidopsis, where CLV signaling is the major single pathway in all meristems.
| The body plan of plants is regulated by the function of apical meristems that are generated in the embryo. Leaves and floral organs are derived from cells supplied by stem cells in the vegetative shoot apical meristem (SAM) and the floral meristem (FM), respectively. Thus, genetic regulation of stem cell maintenance is a central issue in plant development. In the model plant Arabidopsis thaliana, CLAVATA3 (CLV3) functions as a key signaling molecule to restrict the size of the stem cell population in both the SAM and the FM. In rice, however, we show here that two CLV3-like genes, FLORAL ORGAN NUMBER2 (FON2) and FON2 SPARE1 (FOS1), redundantly regulate maintenance of the FM. We also show that FOS1 is likely to be involved in maintenance of the vegetative SAM, whereas FON2 plays no role in regulation in this meristem. FOS1 appears to act via a putative receptor that differs from the FON2 receptor, suggesting that these two signaling molecules function in independent pathways to restrict stem cells in different ways depending on the type of meristem. In addition, we show that the FOS1 gene was compromised in the standard rice, Oryza sativa spp. japonica, during the evolution of rice.
| Intercellular communication plays a crucial role in the determination of cell fate in plant development. Cell fate is determined by positional information emanating from neighboring or distant cells. Recent molecular genetic studies have revealed that peptide signaling molecules are involved in intercellular communication to regulate various aspects of plant development, such as stem cell maintenance, vascular differentiation, stomata patterning, and leaf size control [1]–[5].
The CLE genes encode small secreted proteins with a plant-specific domain called the CLE domain [6]. The CLAVATA3 (CLV3) gene of Arabidopsis thaliana, and the FLORAL ORGAN NUMBER2 (FON2) and FON2-LIKE CLE PROTEIN1 (FCP1) genes of rice (Oryza sativa) are involved in stem cell maintenance in the shoot apical meristem (SAM) and the floral meristem (FM) [1],[7],[8]. Tracheary element differentiation inhibitory factor (TDIF) has a role in suppressing the differentiation of tracheary elements in Zinnia elegans [3]. Recent biochemical studies have revealed that functional peptides of CLV3 and TDIF in vivo are dodeca peptides derived from the conserved CLE domains [3],[4].
In Arabidopsis, stem cell identity in the SAM is maintained by a regulatory feedback loop comprising the CLV and WUSCHEL (WUS) genes [9],[10]. CLV3 acts as a negative regulator of stem cell maintenance by repressing WUS, which encodes a novel homeodomain transcription factor that is expressed in the organizing center and promotes the identity of the stem cells overlying its expression domain [9]–[11]. Conversely, WUS positively regulates CLV3 expression in the stem cell region. CLV3 peptide secreted from the stem cell appears to act through putative receptor complexes, consisting of CLV1, CLV2 or CORYNE/SOL2 [1], [12]–[15]. A recent biochemical study has revealed that CLV3 peptide binds directly to the extracellular domain of CLV1 [16]. When negative regulation of CLV signaling is removed by severe mutations of the CLV1 and CLV3 genes, enlargement of the SAM and the FM occurs, resulting in a fasciated stem and an increase in the number of flowers and floral organs [17],[18].
A similar genetic mechanism to regulate stem cell maintenance seems to be conserved in monocots. Mutations in the FON1 and FON2 genes in rice cause enlargement of the FM, resulting in an increase in the number of floral organs such as stamens and carpels [7],[8],[19],[20]. A double mutant of fon1 and fon2 shows no additive phenotype, suggesting that the two genes act in the same genetic pathway [8]. FON1 encodes a receptor-like kinase with a leucine-rich repeat (LRR) structure in the extracellular domain that is closely related to Arabidopsis CLV1 [7]. FON2 is a member of the CLE gene family, and the CLE domain of FON2 is similar to that of CLV3 [8],[21]. Likewise, in maize (Zea mays), the thick tassel dwarf1 (td1) gene encodes a CLV1-like receptor kinase, and the fasciated ear2 (fea2) gene, like Arabidopsis CLV2, encodes an LRR protein that lacks a cytoplasmic domain [22],[23]. Loss of function of these genes results in enlargement of the inflorescence meristem (IM) and the FM, causing fasciation of the inflorescences and an increase in floral organ number in maize. Constitutive expression of the FON2 gene results in a severe decrease in the number of flowers and floral organs, probably because of a reduction in the size of the IM and the FM in rice, resulting in a phenotype similar to the wus flower [8],[11]. The effect of FON2 overexpression is not observed in the fon1 mutant, suggesting that FON1 is a putative receptor of FON2. Thus, CLV-related genes negatively regulate stem cell proliferation in the reproductive meristems in both rice and maize, as they do in Arabidopsis.
Despite this conservation, meristems in the vegetative phase are not affected by mutations in these CLV-related genes in the grasses, unlike in Arabidopsis [7],[8],[22],[23]. In rice, constitutive expression of FON2 does not affect meristem maintenance in the vegetative phase [8],[24]. We previously showed that FCP1 is probably involved in stem cell maintenance in the vegetative SAM because constitutive expression of FCP1 causes consumption of the SAM, similar to overexpression of CLV3 in Arabidopsis [24]. This action of FCP1 is also observed in fon1 mutants, suggesting that FCP1 requires a receptor other than FON1. Thus, it is likely that, depending on the type of meristem, two independent pathways negatively regulate stem cell maintenance in rice. In maize, expression of td1 is excluded from the vegetative SAM [23]. Thus, meristem maintenance in the vegetative phase is regulated differently from that in the reproductive phase in the grasses.
During the positional cloning of FON2, we found that expressivity of the fon2 mutation is markedly reduced in F2 plants from a cross between the fon2 mutant (O. sativa japonica) and Kasalath (O. sativa indica) [8]. To explain this difference, we hypothesized that the indica genome might contain genes that suppress the fon2 mutation. In this paper, we describe the isolation and characterization of a gene, named FON2 SPARE1 (FOS1), that suppresses the fon2 mutation. FOS1 encodes a secreted protein with a CLE domain, and is expressed in the SAM, IM and FM. Genetic and molecular analyses indicate that FOS1, together with FON2, is likely to be involved in stem cell maintenance in the FM, in rice species in the Oryza genus including O. sativa indica; by contrast, FOS1 function seems to be severely compromised in O. sativa japonica. In addition, FOS1 is likely to be involved in maintenance of the SAM in the vegetative phase, because overexpression of FOS1 caused the formation of abnormal shoots with a terminated meristem. Analysis of the FOS1 sequence from a large number of domesticated and wild rice species reveals that a nucleotide substitution related to the function of FOS1 may have occurred during divergence of the domesticated rice O. sativa japonica from its wild ancestor.
In wild-type rice, a single pistil derived from congenitally fused carpels develops into a floret, and a single ovule is formed in the pistil. After fertilization, a single seed is formed within the husks, which are derived from the palea and lemma in a floret (Figure 1A). In fon1 and fon2 mutants, by contrast, the number of floral organs such as pistils increases due to an enlargement of the FM (Figure 2B) [8],[19]. Therefore, “twin seeds” are formed within the husks in these fon mutants (Figure 1B) when two or more pistils are produced in a floret (the third and fourth seeds cannot develop to maturity). Here we used the twin seed phenotype as an indication of fon2 mutation.
In a previous screen of the fon2 phenotype for positional cloning [8], we found that expressivity of the fon2 mutation was reduced in F2 plants from a cross between the fon2-1 mutant (japonica) and Kasalath (indica). To estimate quantitatively the frequency of the appearance of the fon2 phenotype, we assessed the numbers of twin-seed phenotypes in this study. First, we counted the number of plants producing the twin-seed phenotype among F2 plants from the fon2-1 and Kasalath cross. As a result, we found that the number of F2 plants showing a fon2 phenotype was reduced markedly in the F2 plants: only 4.5% of F2 plants showed a fon2 phenotype (Table 1). Second, we found that the number of the twin-seed phenotypes per panicle was also reduced markedly (Figure 1C). The median frequency of the appearance of the twin-seed phenotype per panicle was 71% in the fon2-1 mutant (japonica); by contrast, it was reduced to 17% in the F2 plants showing a fon2 phenotype. These results suggest that the indica (Kasalath) genome causes a reduction in the expressivity of the fon2 mutation in the F2 progenies. In other words, it is likely that the indica genome has one or more genes that suppress the phenotype caused by the fon2 mutation.
To address this possibility, we performed quantitative trait locus (QTL) analysis (see Materials and Methods). For QTL analysis, we checked the genotypes of the F2 plants showing the fon2 phenotype to confirm that the fon2 locus was homozygous for the mutation. As a result, a major QTL that suppressed the fon2 mutation was detected in the region between 40 and 80 cM on chromosome 2. In a group of F2 plants showing high suppressor activity, the frequency of genotypes homozygous for the indica genome was very high, whereas that of genotypes homozygous for the japonica genome was very low (Figure 1D). In a group of F2 plants showing low suppressor activity, by contrast, the opposite result was obtained (Figure 1E). Thus, in this region of chromosome 2, high or low suppressor activity was closely associated with a genotype homozygous for indica or japonica, respectively.
The above results indicated that a putative gene that suppresses the fon2 mutation is located in the 40–80 cM region of chromosome 2 in the indica genome. We assumed that if indica has a gene that is functionally redundant to FON2 in this region, this gene should behave as a suppressor-like gene function in genetic analyses. A strong candidate for such a gene would be a CLE gene, like FON2. A survey of rice genomic sequences identified a candidate CLE gene located at about 54 cM on chromosome 2.
Next, we examined whether the CLE gene on chromosome 2 in indica (tentatively named CLE-C2) could suppress the fon2 mutation by conducting the following two experiments. First, we introduced a 3.3-kb genomic fragment from Kasalath containing the CLE gene into the fon2-1 mutant by Agrobacterium-mediated transformation. We found that the defect in the fon2 flowers was completely rescued in the transgenic plants, suggesting that the CLE-C2 gene suppressed the fon2 mutation (Figure 2C, 2G, and 2H). Second, we applied a genetic approach using a Nipponbare/Kasalath chromosomal segment substitution line (N/K CSSL#9), in which the chromosomal segment of Nipponbare (japonica) encompassing the CLE-C2 gene was replaced by that of the Kasalath (indica) genome. We crossed fon2-1 (japonica) with N/K CSSL#9, and then screened for F2 plants that were homozygous for both fon2-1 and the indica CLE-C2 allele by determining the genotype with molecular markers. The results indicated that the flower phenotype of the plants screened was identical to that of wild type, suggesting that the fon2 mutation was also suppressed (Figure 2D, 2G, and 2H). Taken together, these results clearly indicate that the CLE-C2 gene located at about 54 cM on chromosome 2 functions as a fon2 suppressor. Thus, we designated this CLE gene as FON2 SPARE1 (FOS1) because this gene can substitute for FON2.
Sequence analysis revealed that FOS1 consists of one exon with a single open reading frame and encodes a putative small protein of 131 amino acids (Figure S1). FOS1 has a signal peptide that is rich in hydrophobic amino acids at its N-terminus and a CLE domain at its C-terminus (Figure 3A). Among the CLE proteins in rice and Arabidopsis, the CLE domain of FOS1 is more similar to those of CLE8 and CLE13 than to those of FON2, FCP1 and CLV3 (Figure 3B).
Notably, the nucleotide sequence of FOS1 in the fon2-1 mutant (background, Fukei71) was identical to that of standard japonica wild-type strains such as Nipponbare and Taichung65 (T65). By contrast, we found an amino acid difference at the putative cleavage site of the signal peptide between japonica (all three stains; AB455109) and indica (Kasalath; AB455108) (Figure 3A; Figure S1). It is possible that this amino acid substitution in japonica FOS1 causes a defect in the processing of FOS1 and a reduction in the amount of active CLE peptide in japonica. Except for this mutation, no nucleotide change causing indels of amino acids was detected in the FOS1 gene between indica and japonica.
We analyzed the spatial and temporal expression patterns of FOS1 by in situ hybridization. In indica, FOS1 was expressed in all aerial apical meristems, not only in the FM and the IM in the reproductive phase, but also in the SAM in the vegetative phase (Figure 4A–4C). In japonica, FOS1 transcripts were also detected in all apical meristems in a spatial distribution pattern similar to that observed in indica (Figure 4E–4G). These spatial expression patterns suggest that FOS1 may be involved in meristem maintenance in rice. In addition to the meristems, FOS1 transcripts were also detected in the primordia of lateral organs such as the leaf and the floral organs. No significant differences were observed at the transcriptional level in the expression patterns of FOS1 between indica and japonica, suggesting that the functional difference between indica FOS1 and japonica FOS1 is not due to differences at the transcriptional level. No signals were detected in the SAM and FM when sense probes were used (Figure 4D and 4H).
Next, we expressed constitutively indica FOS1 by using the rice Actin1 promoter. Unlike shoots transformed with a control vector, Actin1:indica-FOS1 shoots stopped growing at the seedling stage after a few abnormal and malformed leaves were produced (Figure 4I and 4P). A longitudinal section of the shoot apex revealed that a dome-shaped shoot apical meristem (SAM) was strongly compromised in Actin1:indica-FOS1 plants, as compared with transgenic seedlings carrying a control vector (Figure 4K and 4L). Next, we examined the expression pattern of rice OSH1, an ortholog of Arabidopsis SHOOT MERISTEMLESS (STM) and maize knotted1 (kn1), which marks undifferentiated cells in the meristem [25]. OSH1 was expressed uniformly in the meristem except for the site of leaf initiation (P0) (Figure 4M). By contrast, OSH1 expression was not observed in the meristem of Actin1:indica-FOS1 shoots (Figure 4N). These results indicate that constitutive expression of indica-FOS1 terminated meristem function, suggesting that FOS1 is involved in the maintenance of stem cells in the vegetative SAM as well as in the FM.
To address whether, FON1, a putative receptor of FON2, is required for FOS1 function, we introduced an indica genomic fragment containing FOS1 into the fon1-5 mutant, which has a severe mutation and is thought to be a null allele of fon1. The numbers of floral organs such as stamens and pistils in fon1-5 transformed with the indica FOS1 gene were identical to those in wild type, suggesting that FOS1 functions normally in this fon1 mutant (Figure 2F–2H). Next, we expressed indica-FOS1 constitutively in fon1-5. The resulting shoot showed a phenotype identical to that of wild type constitutively expressing indica FOS1 (Figure 4J). These results suggest that the FON1 receptor is not required for the function of FOS1 and that FOS1 is likely to function in an independent signaling pathway.
To elucidate when the mutation observed in japonica FOS1 occurred during rice evolution, we compared the FOS1 sequence of a number of varieties/species of domesticated rice and wild rice species such as O. rufipogon. To encompass genetic diversity, we examined a core collection of domesticated rice from around the world (WRC, 67 accessions) and from Japan (JRC, 50 accessions) (Table S1; Table S2) [26],[27].
As described above, the one-base change that causes an amino acid substitution at the putative cleavage site of the signal peptide in FOS1 and is associated with its function was found in three japonica strains (fon2-1, Nipponbare and T65). Hereafter, we call this mutation a functional nucleotide polymorphism (FNP) without reference to the japonica or indica type. Sequence analysis showed that 66 out of 68 accessions of japonica had the FNP (Haplotype B, see below), whereas 59 out of 60 accessions of indica did not (Haplotype A) (Table S1; Table S2; Figure S2). Thus, the FNP was closely associated with japonica except for three accessions (Calotoc, Pinulupot 1, Padi Perak). Although the accessions in the WRC have been designated indica or japonica by phenotypic analysis, it seemed likely that the genome of two subspecies might have been introgressed into each other during recent breeding programs. Thus, we examined the type of genome around the FOS1 locus in the three exceptional accessions by using molecular markers. The results clearly indicated that the FNP is consistent with the japonica genome, but not with the indica genome (Table 2).
Next, we compared the FOS1 sequence from five wild rice species (22 accessions) and the African domesticated rice O. glaberrima (2 accessions), all of which have an AA genome (Table S3). Nucleotide polymorphisms were found in FOS1 among the wild and domesticated rice accessions (Figure S2). We classified the FOS1 sequences into 13 haplotypes, and generated a network of these haplotypes (Figure 5). The network indicated that the prototype of FOS1 is haplotype C, which is shared by two wild rice species, O. rufipogon and O. glumaepatula. The FOS1 sequence in wild rice species and domesticated rice may have been derived from this haplotype. None of the accessions of wild rice species showed the FNP at the processing site. Therefore, the FNP in FOS1 is specific to the genome of O. sativa japonica. Because Asian domesticated rice species, namely japonica and indica, are thought to have derived independently from O. rufipogon [28]–[30], this FNP may have occurred during the diversification of japonica from O. rufipogon.
In this paper, we identified a new CLE gene, FOS1, in rice by screening for a suppressor of the fon2 mutation. Like FON2, FOS1 is likely to regulate stem cell maintenance negatively in the FM, but the action of FOS1 is independent of FON1, the putative receptor of FON2. In addition, FOS1 appears to be associated with maintenance of the SAM in the vegetative phase. Genetic analysis suggests that the function of FOS1 in japonica appears to be reduced by an FNP occurring at the putative cleavage site of the signal peptide. Distribution of the FNP suggests that this mutation might have occurred during the divergence of japonica from its wild ancestor.
The presence of a factor that suppresses fon mutations in indica was initially assumed from the low expressivity of the fon phenotype in F2 plants from a cross between japonica and indica. Although there are two possible explanations for this low expressivity – namely, differences in the genetic background of japonica and indica, or the presence of a major gene in the genome of indica – QTL analysis provided evidence in support of the latter possibility.
Several lines of evidence suggest that the function of FOS1 is likely to be compromised in japonica. As a result, mutations at the FON2 locus result in enlargement of the FM and an increase in the floral organ number in japonica [8]. In indica, by contrast, functional FOS1 probably masks fon2 mutations by substituting for FON2 function in regulating maintenance of the FM (Figure S3). Likewise, in F2 plants from a cross between japonica and indica, FOS1 derived from indica is likely to mask the fon2 mutation. The frequency (4.5%) of the appearance of the fon2 phenotype, which is also confirmed by the genotype, in those F2 plants is roughly consistent with that expected for the appearance of double mutants. Overexpression of japonica FOS1 produced an abnormal shoot, as did overexpression of indica FOS1, suggesting that japonica FOS1 is not a complete loss-of-function mutant. In wild-type japonica, however, FOS1 CLE peptide, even if produced in part, would be insufficient to restrict stem cells in the FM.
Because our rice research is principally based on japonica, indica FOS1 appears to behave as though it is a suppressor of the fon2 mutation. A more likely interpretation is, however, that FOS1 regulates maintenance of the FM redundantly with FON2 in a wide range of species in the genus Oryza (see below) and that japonica is a mutant for the FOS1 locus.
In plant development, it is well known that genes that encode closely related proteins have redundant functions. APETALA1 (AP1) and CAULIFLOWER (CAL), which that encode MADS-box transcription factors, regulate floral meristem identity together with LEAFY [31],[32]. The ap1 cal double mutant has a striking phenotype, showing excessive proliferation of the inflorescence meristem, which resembles a cauliflower. This phenotype differs from that of the ap1 single mutant. Because CAL has less effect on floral meristem identity, its single mutation shows no phenotype. CAL was identified as an enhancer of the ap1 phenotype in F2 plants from a cross between the ap1 mutant on a Landsberg electa background and wild-type Wassilewskija [31]. Thus, the identification of FOS1 in this study resembles the discovery of CAL in Arabidopsis, although FOS1 has the opposite effect; that is, it appears to be a suppressor of fon2. In the case of AP1 and CAL, functional redundancy is due to the factors themselves; by contrast, signaling pathways comprising a different signaling molecule and its receptor might be redundant in the case of meristem maintenance in rice, as discussed below.
Our previous study demonstrated that FON2 is a negative regulator of stem cell maintenance in the FM [8]. In this study, introduction of indica FOS1 into fon2-1 by genetic methods using a chromosomal segment substitution line or by Agrobacterium-mediated transformation completely suppressed the fon2 mutation. This finding suggests that FOS1 can substitute for the function of FON2. Thus, FOS1 is likely to play an important role in maintenance of the FM in indica and either one of FOS1 and FON2 appears to be sufficient to restrict stem cell proliferation in the FM.
There are two possible explanations for the redundancy of FOS1 and FON2. Both CLE peptides may be involved in the same pathway and may share their receptors. Alternatively, there may exist two independent pathways: one involving FOS1 and one involving FON2 as signaling molecules. Two experiments in this study supported the latter possibility. First, a genomic fragment containing indica FOS1 was able to rescue a severe mutant of fon1, in which the putative receptor of FON2 is defective [7]. Second, constitutive expression of indica FOS1 in fon1 mutant showed abnormal shoots, a phenotype that is similar to that of wild type overexpressing indica FOS1. These results suggest that FON1 is not required for FOS1 function and that the signaling pathways involving FON2 and FOS1 are independent of each other. Because wild species in the Oryza genus have no mutation in the functional region of FOS1, these two pathways may function in the FM in all Oryza species except for japonica (Figure 6).
We found that FOS1 is expressed in the vegetative phase, and constitutive expression of FOS1 generates abnormal shoots with malformed leaves. It is, therefore, likely that FOS1 may be involved in maintenance of the vegetative SAM. Constitutive expression of FON2, by contrast, does not cause abnormalities in the shoot [24]. In this respect, FOS1 and FON2 may have diversified functionally (Figure 6).
In contrast to the FM, the vegetative SAM seems to be unaffected by loss of both FOS1 and FON2 because shoot morphology is normal in fon2 mutants in japonica [8]. Therefore, stem cell maintenance in the vegetative SAM may be regulated by an as yet unidentified negative pathway. FCP1 is likely to be involved in this pathway because its constitutive expression consumes stem cells in the vegetative SAM [24]. Thus, stem cell maintenance may be regulated by at least three negative pathways in rice, and each pathway may contribute differently to this regulation depending on the type of the meristem (Figure 6). This situation contrasts with that in Arabidopsis, where CLV signaling is the major single pathway in all meristems.
Recent genetic and phylogenetic analyses have revealed that indica and japonica arose independently from a genetically distinct population in a wild ancestor, O. rufipogon [28]–[30],[33]. Our haplotype network of FOS1 is also consistent with an independent origin of the two subspecies. In our network, haplotype C would have been the prototype of FOS1 for all domesticated and wild rice species. Haplotype A, associated with indica, and haplotype B, associated with japonica, would have been produced by the occurrence of a single nucleotide change in haplotype C during rice evolution. Indica may have been derived from an O. rufipogon species with haplotype A. The FNP at the cleavage site of the signal peptide is responsible for the generation of haplotype B. Although there is no O. rufipogon accession with haplotype B, it is possible that japonica might have been domesticated from an unidentified ancestor with this FNP.
Many types of mutation are found in FOS1 of wild rice species, including not only amino acid substitutions but also insertions or deletions (Figure S2). There are, however, no mutations that affect the function of FOS1 in the coding region, such as an amino acid change in the CLE domain or a frameshift mutation. This observation suggests that defects in FOS1 may not be neutral and that FOS1 may be essential for the growth and survival of wild rice species under natural conditions. In line with this hypothesis, it is unlikely that an O. rufipogon species that has the FNP in FOS1 will be found in the natural population at present.
Taichung 65 (T65) and Kasalath were used as representative strains of wild-type japonica and indica, respectively, in molecular genetic and histochemical analyses. Nipponbare/Kasalath chromosomal segment substitution line #9 (N/K CSSL#9) was obtained from the Rice Genome Resource Center, Japan. Core collections of O. sativa (World Rice Collection (WRC) and Japanese Rice Collection (JRC)) were obtained from the Genebank of National Institute of Agrobiological Sciences, Japan (Table S1; Table S2) [26],[27]. Wild rice species were obtained from the National Institute of Genetics, Japan (Table S3).
F2 plants from a cross between fon2-1 and Kasalath were used to search for a gene that suppresses the fon2 mutation. We obtained 154 F2 plants showing a fon2 phenotype from about 2,000 F2 plants and checked their genotypes to confirm that the fon2 locus has the mutant allele. For QTL analysis, the suppressor activity in each F2 plant that had the fon2 mutation was estimated by calculating the frequency of the twin seeds, an indication of the fon2 mutation. Next, the genotypes of about 90 loci in the 89 F2 plants were determined by using molecular markers [34]. As a result, a major QTL was found at around 40–80 cM on chromosome 2 (LOD score: 6.3). A gene (FOS1) encoding a protein with a CLE domain was then identified at around 40 and 80 cM on chromosome 2 by searching the rice genomic sequence database using the amino acid sequence of the FON2 CLE domain as a query.
FOS1 cDNA was amplified with the appropriate primers (Table S4) from total RNA isolated from young panicles of T65 (japonica) and Kasalath (indica). After sequencing of the RT-PCR product, the open reading frame was predicted.
To introduce indica FOS1 into the fon2 mutant, a 3.3-kb FOS1 genomic fragment, including 2.6 kb of sequence directly upstream of the initiation codon of FOS1, from the Kasalath genome was used. For constitutive expression of FOS1, a FOS1 cDNA derived from T65 or Kasalath was placed under the rice Actin1 promoter [35]. The resulting plasmids, designated Actin1:japonica-FOS1 (T65) and Actin1:indica-FOS1 (Kasalath), were introduced into Agrobacterium tumefaciens strain EHA101 and transformed into rice as described previously [36].
For the in situ hybridization probe for FOS1, a 646-bp fragment consisting of the entire coding region, the 5′ UTR (137 bp) and the 3′ UTR (113 bp) was amplified with the appropriate primers (Table S4). The fragment was cloned into a T-vector by TA-cloning (Novagen, Madison). The OSH1 probe was prepared as described in the original paper [25]. Probe synthesis, preparation of sections, in situ hybridization, and microscopic observation were performed as described previously [7],[24].
The genomic region of FOS1 was amplified with the appropriate primers (Table S4). The amplified fragments were purified with Montage PCR Filter Units (Millipore, Billerica) and sequenced with the same primers used for amplification. The haplotype network was constructed by using the program TCS1 [37].
|
10.1371/journal.pgen.1002130 | Graded Nodal/Activin Signaling Titrates Conversion of Quantitative Phospho-Smad2 Levels into Qualitative Embryonic Stem Cell Fate Decisions | Nodal and Activin are morphogens of the TGFbeta superfamily of signaling molecules that direct differential cell fate decisions in a dose- and distance-dependent manner. During early embryonic development the Nodal/Activin pathway is responsible for the specification of mesoderm, endoderm, node, and mesendoderm. In contradiction to this drive towards cellular differentiation, the pathway also plays important roles in the maintenance of self-renewal and pluripotency in embryonic and epiblast stem cells. The molecular basis behind stem cell interpretation of Nodal/Activin signaling gradients and the undertaking of disparate cell fate decisions remains poorly understood. Here, we show that any perturbation of endogenous signaling levels in mouse embryonic stem cells leads to their exit from self-renewal towards divergent differentiation programs. Increasing Nodal signals above basal levels by direct stimulation with Activin promotes differentiation towards the mesendodermal lineages while repression of signaling with the specific Nodal/Activin receptor inhibitor SB431542 induces trophectodermal differentiation. To address how quantitative Nodal/Activin signals are translated qualitatively into distinct cell fates decisions, we performed chromatin immunoprecipitation of phospho-Smad2, the primary downstream transcriptional factor of the Nodal/Activin pathway, followed by massively parallel sequencing, and show that phospho-Smad2 binds to and regulates distinct subsets of target genes in a dose-dependent manner. Crucially, Nodal/Activin signaling directly controls the Oct4 master regulator of pluripotency by graded phospho-Smad2 binding in the promoter region. Hence stem cells interpret and carry out differential Nodal/Activin signaling instructions via a corresponding gradient of Smad2 phosphorylation that selectively titrates self-renewal against alternative differentiation programs by direct regulation of distinct target gene subsets and Oct4 expression.
| Nodal and Activin are extracellular signaling molecules that diffuse from the source of secretion and induce recipient stem cells to become new cell types according to a concentration gradient. In the early embryo, they are important for the specification of tissues at the correct place and time, but paradoxically they drive the opposite function in embryonic and epiblast stem cells where they maintain the stem cell state instead of promoting differentiation. The molecular basis of how the level of signaling determines stem cell fate decisions remains poorly understood. We found that Smad2, the main transcription factor of the Nodal/Activin pathway was phosphorylated according to the level of signaling. By mapping where phospho-Smad2 binds in the embryonic stem cell genome and how this affects transcription of the associated target genes, we show that phospho-Smad2 can recruit and regulate different sets of target gene depending on the signaling level. Moreover, phospho-Smad2 also directly regulates Oct4, a master gene controlling the stem cell state thereby reconciling the opposing functions of the Nodal/Activin pathway in differentiation versus self-renewal programs. The pathway can mediate the exit from self-renewal via Oct4 and simultaneously drives differentiation towards particular lineages by recruiting the relevant gene subsets for this purpose.
| Morphogens are secreted signaling molecules that orchestrate the spatial distribution and sequence of cellular differentiation events throughout embryonic development. The specific cell types, their localization and order of induction from recipient stem cell populations are determined by the concentration gradient of morphogens diffusing from the source of secretion. Previous studies have proposed some of the models by which morphogen gradients are initiated, established and stabilized including the level of receptor occupancy, positive/negative feedback and feed forward mechanisms [1]–[3]. However, little is understood about the transcriptional mechanisms responding to variable receptor activation and how they permit pluripotent stem cells to interpret signaling levels and direct the appropriate differentiation programs during mammalian development.
Nodal and Activin are morphogens of the TGFβ superfamily of signaling molecules. In Xenopus embryos, Activin is a potent concentration-dependent inducer of mesoderm, mesendoderm and endoderm in animal cap cells [2], [4], [5]. Nodal has also been shown to be a classical morphogen in zebrafish where it functions in a concentration gradient independently of any relaying mechanisms [6]. In the early mouse embryo, mutations that perturb the level of Nodal/Activin signaling show that the pathway plays crucial roles in the induction of the primitive streak/mesoderm, mammalian organizer (node), mesendoderm and endoderm during the establishment of the anterior-posterior axis [7]–[11]. In contrast to in vivo evidence that Nodal/Activin signaling predominantly promotes differentiation events, the pathway also paradoxically has important roles in the maintenance of self-renewal and pluripotency. Indeed Activin A is frequently used directly in culture for the continued propagation and expansion of human embryonic and mouse epiblast stem cells [12]–[15].
The signaling level of the Nodal/Activin pathway is determined by the overall activity of its components many of which have been identified. Both the Nodal and Activin ligands bind to the same type I/II serine-threonine receptor kinase complexes consisting of ActRIIA/B and Alk4/5/7 respectively in the mouse [16]. Nodal requires the cofactors Cripto/Criptic for receptor activation as opposed to Activin that can bind directly to the receptors and is inhibited by Cripto [17]–. Upon ligand docking, the Type I receptors phosphorylate the downstream signal transducers Smad2 and Smad3 (Smad2/3) which form hetero- or homodimers and trimers [20]. Both Smad2/3 are also phosphorylated by crosstalk with EGF/ERK/MAPK signaling [21]–[23] but only the serine residues of the SSXS motif on the extreme carboxy terminus are specifically phosphorylated by Nodal/Activin/TGFbeta signaling. This phosphorylation is important for the translocation of Smad2/3 to the nucleus in association with Smad4 [24], [25] where the complex recruits a number of transcription factors including FoxHI, p53, β-catenin and Jun/Fos for the direct regulation of target genes [20]. Specificity of the Smads for their direct target genes is partly conferred by a DNA domain in the MH1 region to the Smad-binding DNA element (SBE) consisting of a basic CAGA sequence or its complement [26]. The other partner transcription factors within the complex are required for additional target gene affinity and specificity.
While Smad2/3 share more than 90% protein homology, they are not functionally equivalent. Full-length Smad2 differs from Smad3 as the presence of an inhibitory domain in the MH1 region prevents direct DNA binding while Smad3 can bind directly to SBE boxes [27]. However, an alternatively spliced variant of Smad2 that lacks the inhibitory domain can bind DNA directly and has been shown to be the isoform that accounts for all developmental Smad2 functions in vivo [28]. The developmental roles of Smad2/3 are also disparate. Smad2 knockout mouse embryos fail to form mesoderm and endoderm due to defects in primitive streak specification after implantation at 6.5 dpc [29] closely phenocopying Nodal mutants [10]. In contrast, Smad3 mutant mice are born alive and are fertile but develop chronic intestinal inflammation leading to colorectal cancer [30]. This suggests that Smad2 is the primary transcriptional mediator of early developmental events while Smad3 is involved in immune function and possibly acts as a tumor suppressor postnatally.
Our focus here is to clarify how mechanistically different levels of Nodal/Activin signaling lead to different embryonic stem (ES) cell fate decisions. ES cells were differentiated using three different quanta levels of Nodal/Activin signaling. We showed that ES cells are able to arbitrate between three distinct cell fate decisions. Maintenance of endogenous Nodal/Activin signaling is required for self-renewal of ES cells where any perturbation leads to an exit from self-renewal and pluripotency programs towards mesendoderm induction at high signaling and trophectoderm differentiation at low signaling.
One obvious question to resolve is whether different levels of Nodal/Activin signaling recruit different sets of genes. While genome wide transcriptome studies have suggested possible Nodal/Activin targets, the identity of many transcriptional targets directly regulated by Smad2/3 remains unknown. One ChIP-chip study to date has been performed to address endogenous Smad2/3 binding in transformed human keratinocytes [31] while none have been carried out in the context of stem cell fate decisions, graded Nodal/Activin signaling or examining beyond promoter regions. Here we performed quantitative chromatin immunoprecipitation (ChIP) of phospho-Smad2 (pSmad2) during graded Nodal/Activin signaling followed by massively parallel sequencing (ChIP-Seq) covering the full extent of pSmad2 binding to the ES cell genome including 5′/3′ UTRs, exons/introns and gene deserts. PSmad2 binding and regulation of direct target gene expression does not vary uniformly across the genome but changes in both a qualitative and quantitative manner with different signaling levels. Some targets are up- or downregulated proportionate to the activity of the Nodal/Activin pathway. However, separate subsets of target genes are regulated only during high or low signaling conditions. The downstream consequences of this is differential expression of the target genes that combine dose-dependent genes with different subsets of genes activated or repressed specifically for each signaling level. Thus ES cells carry out alternative cell fate decisions via the recruitment of target gene subsets in a pSmad2 dose-dependent manner.
To reconcile some of the conflicting functions of Nodal/Activin signaling in self-renewal and pluripotency versus differentiation cell fate decisions, we examined the regulation of the Oct4 pluripotency and self-renewal master gene. Oct4 was directly regulated by pSmad2 binding in the promoter region independent of all other cis regulatory elements. Consistent with the modulation of pSmad2 binding, both endogenous mRNA and protein levels of Oct4 were also repressed by inhibition of Nodal/Activin signaling. Hence pSmad2 is a direct upstream regulator of Oct4 transcription where it permits an exit from maintenance of the stem cell state towards mesendoderm or trophectoderm differentiation programs as specified by the signaling level.
In conclusion, the molecular switching of binding locations and target genes by pSmad2 across the ES cell genome in a dose-dependent manner provides a mechanism for the shift in the balance between maintenance of the stem cell state and the opposing induction of differentiation. Key signaling pathways have been predominantly studied in a binary context where they are either present or absent in a biological system. This view has only been able to account for some of their many and often conflicting roles. Our findings challenge this view and support multi-level signaling in stem cells where different signaling strengths can engender different cell fate decisions reflective of the in vivo development of embryos directed not just by Nodal/Activin signaling but possibly Hedgehog, FGF, Wnt and other morphogen pathways.
The direct cellular function of the Nodal/Activin pathway notably of the downstream components Smad2/3/4 is for the regulation of transcription. To address the relation between graded signaling and how they affect transcription, we quantified the changes in expression of known target genes under different signaling levels in chemically defined KSR media conditions. Pluripotent mouse embryonic stem (ES) cells were used to assess the mechanism of morphogen activity as they can differentiate into all tissue types of the adult and express all components of the pathway permitting response to manipulated Nodal/Activin signaling.
Some of the known target genes include Pitx2 and Lefty2 which are responsible for the establishment of left-right asymmetry during early embryogenesis, a key developmental role of Nodal/Activin signaling [32]. In addition, both Lefty2 and Smad7 function as inhibitors of the pathway in a negative feedback mechanism for the attenuation of Nodal/Activin signaling strength [33], [34]. Although direct Smad2/3 binding and regulation of the Pitx2 and Lefty2 genes have not yet been demonstrated, in vivo reporter assays suggest that specific enhancers are responsive to Nodal/Activin signaling and are active only on the left side of the embryo [35], [36]. Moreover, these enhancers have been shown to contain FoxH1 binding sites, a known key transcriptional copartner of Smad2/3. Smad7 has been shown to be a direct target of Smad2/3/4 binding in the promoter region by gel shift assays [37], [38] and it antagonizes the interaction of Smad2/3 with the Type I kinase receptors [39] during negative feedback.
Using real-time PCR quantitation, the expression of the 3 target genes was examined in the ES cells following the induction of high signaling by direct treatment with Activin in a time-course. In the reciprocal experiment, the small chemical inihibitor SB-431542 that specifically prevents the kinase domains of the Type I kinase receptors from phosphorylating Smad2/3 [40] was used to generate low Nodal/Activin signaling conditions. Pitx2, Lefty2 and Smad7 were up- and downregulated in direct correlation with the level of signaling under chemically defined conditions compared to the DMSO carrier control representing endogenous or medium signaling (Figure 1). Over the course of 24 hours, the maximum expression of Pitx2 and Lefty2 occurred at 18 hours (Figure 1A and 1B) while that of Smad7 (Figure 1C) occurred earlier at 12 hours. We therefore conclude based on these known target genes that Nodal/Activin signal transduction and its effects on transcription require up to 18 hours to fully develop and any earlier time points result in weaker inductions.
We further confirmed that Pitx2, Lefty2 and Smad7 are direct targets of the Nodal/Activin pathway by conducting chromatin immunoprecipitation of phosphorylated Smad2 in the ES cells under the same chemically defined conditions at 18 hours followed by quantification of the enriched genomic DNA fragments by real-time PCR using tiling primers (Figure S1 and Table S3). The antibody used for the pulldown was raised against the phosphorylated serines 465 and 467 on the carboxy-terminus of Smad2 that are specifically targeted by TGFbeta signaling and not by EGF/ERK/MAPK signaling. At 18 hours where there is maximum expression of the 3 target genes, there was also a robust divergence in the level of pSmad2 binding according to the signaling level for the enhancers of Pitx2 and Lefty2 (Figure S1A and S1B). Interestingly pSmad2 binding was invariant on the known TGFbeta response element of the Smad7 promoter (Figure S1C). This suggested that Nodal/Activin target genes had different binding efficiencies for pSmad2 at each Nodal/Activin signaling level and this was not uniformly changed for all target genes.
In conclusion, we confirm that Pitx2, Lefty2 and Smad7 were direct targets of Nodal/Activin signaling and graded pSmad2 binding. Differential signaling sustained for 18 hours also leads to the maximum level of differential gene expression with clear changes in pSmad2 binding on the Pitx2 and Lefty2 genes.
Given the downstream changes in pSmad2 binding and transcription of the known direct target genes, we next addressed how extracellular signal levels are translated into intracellular levels of signal transduction. We hypothesized that this could be directly related to changes in pSmad2 levels in ES cells as a consequence of Type I receptor kinase activity. Hence ES cells subjected to differential morphogen signaling conditions may be able to produce different amounts of pSmad2 in cells generating a corresponding differential level of intracellular signaling that leads to differential transcription.
It has recently been shown that overexpression of the constitutively active Alk4 type I kinase receptor is sufficient to drive phosphorylation of Smad2 independent of all other Nodal/Activin receptor complex components [41]. Here we show that direct treatments of the ES cells with Activin and the specific Type I receptor kinase inhibitor SB-431542 in chemically defined conditions also tightly regulates receptor complex activity and produces the phosphorylation of Smad2 in a signaling dependent manner (Figure 2A). During Activin stimulation (high signaling) for 18 hours, there is a defined 2-fold increase in pSmad2 levels while repression with 10 µM SB (low signaling) leads to a 2-fold decrease that is within the limits of physiological change compared to the DMSO vehicle control (equivalent of medium signaling). Differential signaling had no effect on the equilibrium of total Smad2 suggesting that only phosphorylation and not regulation of the total Smad2 population is mediated by Nodal/Activin signaling. Hence extracellular signaling levels are translated into an equivalent gradient of intracellular Smad2 phosphorylation in ES cells.
Subsequently we addressed the long-term consequences of increased and decreased signaling on ES cell fate decisions by examining how manipulation of the pathway recapitulates in vivo cell fate decisions by direct treatment with Activin or SB for 6 days. Analysis of a broad range of early cell fate markers (Figure 2B) shows that enhanced Nodal/Activin signaling promotes mesendoderm differentiation in ES cells with strong upregulation of mesendodermal lineage genes including Gsc, Mixl, Eomes and Fgf8. The marker for mesoderm, Brachyury (T), was also strongly induced although this was not reflected by the other mesodermal markers such as Flk1 and Tbx6. This was consistent with the finding that T is also co-expressed in mesendoderm in vivo at the anterior primitive streak [42]. Taken together, this suggests that high signaling induced by Activin predominantly drives mesendoderm differentiation.
Conversely, inhibition of the pathway with SB (Figure 2B) led to the upregulation of trophectoderm specific markers including Dlx3, Esx and Hand1 and a less significant induction of extraembryonic primitive endoderm markers such as Gata4/6 and Pdgfra. Similar results were obtained when the ES cells were treated with recombinant Lefty1 protein for the same period of time (data not shown) suggesting that the trophectoderm induction was specific to low Nodal/Activin signaling. Interestingly there was no induction of mesendodermal markers as in the Activin treatment and instead some of these such as Gsc, Mixl and Fgf8 were strongly downregulated. Together, these results suggest that perturbation of the level of Nodal/Activin signaling and consequently endogenous Smad2 phosphorylation led to an exit from self-renewal in ES cells towards highly divergent cell fate decisions of either mesendoderm or trophectoderm differentiation.
To confirm these results, fluorescent immunostaining was carried out to assess the protein markers of trophectodermal and mesendodermal lineages (Figure S2) after differentiation in serum containing media. The cell fates obtained under these conditions are similar to the results from the marker analysis performed in chemically defined conditions. Differentiated cells staining positive for Mixl and Lim1 in the nucleus could be detected in Activin cultures. Similarly, Hand1 and placental Cadherin (P-cad) positive giant cells could also be derived from SB treated ES cells. Control treatments with a low dose of DMSO carrier (1/5000 dilution) contained large populations of ES cells that stained strongly for Oct4 and SSEA-1. These results confirmed that the level of Nodal/Activin signaling is responsible for at least 3 cell fate decisions. The endogenous level of signaling is permissive for self-renewal and maintenance of pluripotency, an increase in signaling leads to the induction of mesendoderm like cells while reduction of signaling results in trophectoderm differentiation.
We hypothesized that for divergent differentiation programs to be initiated in ES cells, differential gene expression mediated by pSmad2 transcription would be a pre-requisite, which is in turn dependent on the level of Nodal/Activin signaling. Each discrete signaling threshold should induce an independent and unique transcriptional signature distinct from other thresholds. To determine the genetic targets regulated downstream of Nodal/Activin signaling and their pattern of expression, microarray analysis was carried out to examine genome-wide gene expression following Activin, DMSO or SB treatments in chemically defined KSR media for 18 hours.
No significant changes in gene expression out of 26,000 probes could be detected between the DMSO and KSR media control suggesting that the effect of the low concentration of DMSO was negligible on ES cells (Figure 3A). In contrast, Activin and SB treatments induced specific changes in gene expression compared to the DMSO and KSR media controls. Most significantly, we were able to identify subsets of target genes that were regulated by one signaling level and not the other consistent with our hypothesis of threshold specific target gene regulation. For example, 19 genes including Gdf15, Msmb and Orai3 were consistently upregulated in Activin treated cells while showing no significant changes in SB. In contrast, a larger subset of 131 target genes were specifically up- and down-regulated only in the SB treatment and not in Activin. A core subset of 12 targets was co-regulated by both high and low signaling changing their expression in correlation with the treatment including the known Nodal/Activin target genes Lefty1/2 and Pitx2 that were upregulated by Activin and downregulated in SB.
Interestingly, the number of SB regulated targets significantly exceeds that of Activin targets, suggesting that endogenous Nodal/Activin signaling in ES cells is high or near saturation levels such that a 2-fold increase in pSmad2 could only induce a smaller subset of genes compared to a 2-fold downregulation. Higher doses of Activin treatments and greater than 2-fold increases in pSmad2 may be required to mirror the strength of SB inhibition providing an explanation for asymmetric up- or downregulation of gene expression during different levels of signaling. Some of the target genes driven by Nodal/Activin signaling were indeed implicated in the mesoderm, endoderm and trophectoderm lineages. Fgf15 plays an important role in the development of cardiac mesoderm [43] and Chst15 is specifically expressed in definitive endoderm in vivo [44] with both targets being upregulated by Activin. For SB treatments, Gata3, Tcfap2c and Igf2 were specifically upregulated. Gata3 is a driver of trophectoderm development [45], [46] while Tcfap2c is expressed specifically in the placenta where it regulates essential ADA expression [47], [48] and Igf2 is an imprinted gene that modulates nutrient supply between the placenta and fetus [49], [50]. Together these target genes support some of the mesendodermal and trophectodermal differentiation programs that may be initiated at 18 hours after the induction of differential Nodal/Activin signaling. With longer-term graded Nodal/Activin signaling over 6 days differentiation, it is likely that additional target genes reinforcing the specification of both lineages may be brought into play over time.
Lefty1, Pitx2, Fgf15 and Spsb1 were validated by RT-PCR (Figure 3B) to be co-regulated target genes of high, medium and low signaling displaying a gradient of expression following the signaling level. Cripto, Bcar3, Nphs1 and Cdh3 were targets that were predominantly downregulated by SB inhibition of signaling showing no significant change during Activin stimulation. Conversely, the ID1/2/3 family of transcriptional repressors and Serping1 are specifically upregulated only by the SB treatment showing no difference in response to either Activin or the DMSO control. Hence we conclude that different thresholds of Nodal/Activin signaling are indeed able to regulate the expression of specific subsets of target genes providing an important explanation for the establishment of divergent differentiation programs.
While whole genome microarrays are able to identify the putative subsets of target genes differentially expressed during specific Nodal/Activin signaling levels, this does not provide a molecular mechanism for how different target genes can be directly regulated by the same pathway at different signaling strengths. To address this question, we examined the recruitment of the pSmad2 transcription factor to target genes after subjecting ES cells to Activin, SB or DMSO control treatments in chemically defined KSR media that produce 2-fold up- and downregulation of Smad2 phosphorylation by 18 hours. ChIP-Seq of pSmad2 was employed to identify where pSmad2 was binding on a whole genome scale in parallel cultures of ES cells under the 3 signaling conditions. ChIP samples from each condition were sequenced to a similar depth of 10-13 million tags. Interestingly the number and magnitude of pSmad2 binding events did not correspond to the 2-fold up- or downregulation of pSmad2 in ES cells under Activin and SB treatments. In fact the greatest number of binding peaks (7423) occurred in the control DMSO condition that maintains self-renewal and pluripotency of ES cells (Figure 4A). When homeostatic Nodal/Activin signaling was perturbed by Activin and SB treatments, the number of binding events decreased to 5094 and 4859 respectively suggesting that any change in the levels of endogenous pSmad2 from the ES cell undifferentiated condition also caused a dynamic change in pSmad2 binding across the ES cell genome. The lower numbers may also be reflective of the transition where pSmad2 is dissociating from former target genes and establishing the recruitment of new genes. This was further supported by the percentage of overlapping peaks that were common to the 3 treatments being relatively small at 10.3% with a significantly larger number of unique peaks appearing in specific treatments (37.25% in DMSO, 20.44% in SB and 19.5% in Activin out of 12979 total peaks in the union).
A previous study has profiled Smad2/3 binding sites using promoter arrays in human keratinocytes [31]. However, reporter assays on Nodal/Activin responsive target genes such as Lefty1/2, Nodal [36], [51] and Pitx2 [35] suggest that Smad2/3 may also regulate DNA elements in the introns rather than at the promoter region. Consistent with the reporter assay studies, our ChIP-Seq data showed that the majority of pSmad2 binding (Figure 4B) occurs in introns (∼30%) with only a minority of sites in the proximal promoters (∼10%). Furthermore, there was a significant shift in pSmad2 binding from the distal 5′ and 3′ regions towards the promoters of genes in the SB treatment compared to DMSO and Activin (Figure 4B). Examination of binding specifically in the promoter region showed a clear preference for pSmad2 to associate in the +/−600 bp proximal region of transcriptional start sites (TSS) with a steady decrease in binding further away from the TSS (Figure 4C). In addition, the increase in number of pSmad2 binding peaks during low signaling with the SB treatment can be confirmed in the promoter region both up- and downstream of the TSS. In conclusion, pSmad2 binding, similar to the changes in gene expression identified by microarrays, also demonstrates binding to distinct subsets of genomic locations at different signaling levels.
We next examined the relationship governing the degree of pSmad2 binding and the level of transcription across the genome (Figure 5A). In all 3 conditions, a clear trend emerges suggesting that more pSmad2 binding drives higher levels of gene expression. However, the possibility that pSmad2 is not driving expression but preferentially associates with more transcriptionally active genes and open chromatin cannot be excluded. To distinguish between the 2 possibilities, we examined the trend between pSmad2 binding events and differential gene expression from the microarray analysis in the 3 signaling conditions. Indeed, a significant majority (64.2%) of microarray target genes had pSmad2 binding within +/−50 kb and all displayed >1.5 fold change in binding in each signaling condition or had different number of binding events or changed the location of pSmad2 binding (Table S1) suggesting that the pattern of gene expression was indeed dynamically driven by pSmad2-DNA interactions.
To account for how pSmad2 is able to switch binding locations during differential Nodal/Activin signaling, we examined its preference for specific DNA motifs under each condition. It is known that Smad2/3 are able to bind directly the basic CAGA motif and at the same time they possess a number of partner transcription factors that modulate the specificity and strength of binding. Here we see that there is strong pSmad2 association with the basic CAGA SBE specifically in the Activin treatment (Figure 5B). This was also confirmed when we examined the strong CAGAC canonical SBE as defined by the TRANSFAC PWM database which also appears with high frequency at the center of pSmad2 ChIP-seq peaks in the Activin treatment and not in DMSO, SB or the random mouse genome sequence control (Figure 5C). This suggests that both CAGA and CAGAC displayed graded pSmad2 binding that varied with the signaling level and were preferentially bound in the Activin condition. To compare the contribution of CAGA against non-CAGA sequences towards pSmad2 binding, the top 10 de novo motifs in each condition were identified using the Weeder program (Figure 5D). Motifs that occurred with significant frequency but were not enriched in the center of pSmad2 ChIP-Seq peaks were excluded to remove the influence of comotifs around the peaks. A number of non-CAGA motifs that occurred with similar or greater frequency than CAGA were isolated. Interestingly, these de novo motifs also showed a graded effect on pSmad2 association similar to the CAGA SBE. Other non-CAGA motifs were preferentially bound by pSmad2 only in the DMSO and SB condition and depleted during the high signaling Activin condition. This suggested that while CAGA binding was significant, binding to non-CAGA sequences accounted for the majority of pSmad2 association within the ES cell genome suggesting that this was primarily mediated by transcriptional co-partners. Indeed, when the top consensus motifs in the center of all ChIP-seq peaks in each signaling condition and in the combined dataset were studied (Figure 5E), there was a strong enrichment for motifs belonging to transcription partners such as E2f and Ap1 instead of Smad binding CAGA boxes. To confirm the association of the putative transcriptional cofactors and establish their identity, we expanded the analysis to TRANSFAC co-motifs occurring within +/− 1 kb range of pSmad2 binding sites (Table S2). A large number of known pSmad2 transcription partners such as Ap1, Sp1 and E2f are indeed associated within the vicinity of pSmad2 peaks regardless of the level of Nodal/Activin signaling. However, there were additional co-motifs bound by transcription factors such as Oct4, Stat3 and p53 that only appear prominently in Activin treatments and Hes1, Lrf and Plzf appearing in SB. This is supportive of an exchange of transcription partners in association with pSmad2 that was governed by the level of Nodal/Activin signaling which was likely to be responsible for the change in specificity of pSmad2 transcriptional complexes for target gene subsets and their level of expression. Furthermore, while pSmad2 does bind to its own CAGA sequence, transcriptional copartners played a greater role both in binding affinity and specificity of pSmad2 protein complexes for the ES cell genome.
To investigate the different models of pSmad2 binding during differential Nodal/Activin signaling, we examined the ChIP-Seq profiles including those of the transcriptionally regulated microarray targets and identified at least 4 types of pSmad2 binding. The first model is that of “graded” target genes that follow closely the changes in Nodal/Activin signaling with increased binding and transcription during high signaling, have moderate response in endogenous baseline signaling and showed a loss of binding with decreased mRNA levels during signaling repression. This category of pSmad2 binding comprises 23.87% of high confidence ChIP-Seq peaks corresponding to 16.28% of target genes associated within +/−50 kb of these peaks (Figure S4). Radil a Rap GTPase effector that plays a role in the migration of neural crest progenitors [52] exemplified such pSmad2 binding and transcriptional regulation (Figure 3A, Figure 6A, and Table S1) in the first intron with normalized relative enrichments of 107 tags in Activin compared to 51 in the DMSO control and complete loss of binding indistinguishable from background sequencing levels in SB. The known target gene Pitx2 showed reproducible results with the ChIP data obtained by real-time PCR (Figure S1A) both in terms of the binding location in the intronic enhancer as well as the level of pSmad2 enrichments under graded Nodal/Activin signaling. There were normalized enrichments of up to 201 tags in Activin, 156 in DMSO control and again complete loss of binding in SB (Figure S3A). Interestingly, Pitx2 had 2 graded binding sites, one of which is in the known intronic region and a novel site in the 3′ region. The graded binding in the Pitx2 locus also correlates with transcriptional consequences showing strong induction/inhibition of Pitx2 mRNA levels from 0 to 24 hours (Figure 1A).
The two inducers of the mesendoderm cell fate Mixl [53] and Nodal [54] also show evidence of graded pSmad2 binding within 50 kb of the genes (Figure S5A and S5B) suggesting that they may be directly regulated by Nodal/Activin signaling for this purpose. The binding location in the first intron of Nodal also corresponds to the intronic enhancer previously described to be important for left side expression in the early embryo via the Nodal/Activin signaling autoregulatory loop [55], [56] confirming that Nodal is itself a direct target.
It was also unclear if pSmad2 binding and regulation of target genes only exists in a 1-to-1 relationship or if the same binding sites were capable of regulating multiple targets in the genomic vicinity. While Lefty2 was a known direct target with pSmad2 binding in its promoter region (Figure S1B), for the first time, to our knowledge, we characterized an important pSmad2 transcriptional hotspot in the entire 100 kb Lefty1/2 locus where all the genes within this region were co-regulated by pSmad2 binding suggesting a coordinated mode of transcriptional regulation (Figure S3B). This was further confirmed in the microarray analysis (Figure 3A) demonstrating that Lefty1/2, Pycr2 and Tmem63a display the same pattern of gene expression following a graded response to Nodal/Activin signaling. This was consistent with the real-time PCR quantification of the pSmad2 pulldown of the Lefty2 promoter (Figure S1B) that corresponds to the most upstream pSmad2 binding site in the Lefty1/2 hotspot as did a time course profiling of Lefty2 expression from 0 to 24 hours (Figure 1B).
In the second model of pSmad2 binding, we describe “low signaling dominant” conditions that permit pSmad2 binding but less so under other signaling levels. The Id1/2/3 (Figure 6B, Figure S6, and Table S1) family of transcriptional repressors shows pSmad2 binding to these genes only in the SB treatments and not in Activin or the DMSO control. Statistically, 32.73% of pSmad2 binding sites display this mode of behavior associated with 23.44% of target genes (Figure S4). In contrast, the third model showed the opposite “high signaling dominant” mode of binding such as in the case of 220011C2Rik (Figure 6C and Table S1) where pSmad2 only binds strongly in the Activin condition but to a lesser degree in DMSO or SB also resulting in transcriptional consequences (Figure 3A and Figure S4). Another known component of the mesendodermal cell fate Fgf8 [57] also shows strong pSmad2 binding in the promoter region specifically during high signaling (Figure S5C). Intriguingly, findings in the chick embryo show that Fgf8 also plays important roles in left-right asymmetry where it can be induced by Activin [58] in agreement with our results.
In the fourth model which accounts for the regulation of the largest proportion (33.69%) of target genes associated with pSmad2 ChIP-Seq peaks (), the same target gene may be regulated by “multimodal pSmad2 binding” events. Copz2 has two pSmad2 association sites in the intron and promoter region (Figure 6D). The promoter site only binds pSmad2 in the SB condition while the intronic enhancer shows a graded response to the signaling level. In the case of the known target gene Smad7, we have shown that the pSmad2 binding peak in the promoter region is invariant in all 3 signaling conditions (Figure S1C) which could not explain how Smad7 was differentially expressed during graded Nodal/Activin signaling (Figure 1C). In confirmation with these results, the ChIP-Seq data showed the same pSmad2 association on the Smad7 proximal region with no change in binding under all 3 signaling conditions. Surprisingly, we discovered a previously undescribed pSmad2 regulatory element in the distal Smad7 promoter region that binds pSmad2 in a graded manner (Figure S3C) and could account for why Smad7 was responsive to different Nodal/Activin signaling levels. Hence pSmad2 binding in the Smad7 proximal region may not be the dominant regulatory region for Nodal/Activin signaling but may depend instead on the dynamically changing pSmad2 distal promoter element for Smad7 regulation. Indeed the proximal promoter element may be more of a Smad3 regulated region instead of Smad2 as previously described [38].
In conclusion we demonstrate that pSmad2 dependent binding and transcription during graded Nodal/Activin signaling occurs in the ES cell genome in a graded, low or high signaling dominant, many-to-one or one-to-many multimodal manner in relation to the target genes that they regulate.
The mesendodermal and trophectodermal cell fate decisions brought about by graded Nodal/Activin signaling strikingly resemble the ES cell response to a less than 2-fold up- or downregulation of the Oct4 master regulator of stemness in driving differentiation towards similar cell fates [59]. Furthermore, an important mechanism for trophectoderm differentiation depends on the Oct4 repression of Cdx2 expression and the induction of this lineage is thought to be indicative of loss of stemness [60]. We therefore hypothesized that Oct4 may be a key downstream target under Nodal/Activin control during the specification of divergent cell fate decisions and investigated how the pathway may be governing Oct4.
We discovered that the Oct4 locus was rich in multiple pSmad2 binding events from ChIP-Seq profiling (Figure 7A). During graded Nodal/Activin signaling in chemically defined conditions however, only a pSmad2 peak in the promoter region of Oct4 showed a similarly graded response suggesting that this was the functional Nodal/Activin signaling response element. We examined the transcript levels of endogenous Oct4 expression (Figure 7D) upon inhibition of Nodal/Activin signaling with SB in serum containing media and found that it was also significantly downregulated within 24 hours. In agreement with the transcript data, Oct4 protein levels were similarly downregulated in SB treated ES cells (7E). Analysis of the 503 bp promoter region encompassing the beginning and end of the pSmad2 binding peak showed that it contained eight CAGA sites or their inversion (Figure 7B). To determine if this regulatory sequence was indeed a Nodal/Activin response element of the Oct4 promoter, we cloned this into luciferase reporter constructs and transfected ES cells subjected to the 3 signaling conditions with Activin, DMSO and SB (Figure 7C) in serum containing media. The reporter activity of the wild type Oct4 promoter construct was >100X higher than that of the empty reporter construct in the DMSO control signaling condition suggesting that the 503 bp sequence had strongly driven Oct4 promoter activity in ES cells. Crucially, the Oct4 promoter reporter displayed a specific graded response to Nodal/Activin signaling while the control empty reporter did not. To confirm that the Oct4 response to graded Nodal/Activin signaling was functionally driven by pSmad2 binding, we determined the exact SBE responsible for Oct4 inducibility (Figure 7C). Mutagenesis experiments on the Oct4 promoter region in luciferase assays revealed that the strong CAGAC consensus SBE site in the middle of the 503 bp fragment was indispensable for graded Oct4 promoter activity. Loss of this site completely abolished the promoter response to both high and low Nodal/Activin signaling. Further point mutations of two minimal CAGA SBEs flanking the CAGAC site led to no further significant effects on the Oct4 promoter.
We therefore conclude that Oct4 is a direct target of pSmad2 binding and Nodal/Activin signaling regulates both its promoter activity and endogenous expression. The 503 bp Oct4 promoter response element with the essential CAGAC SBE was sufficient and independent of all other pSmad2 binding events in the Oct4 locus or other DNA regulatory elements in cis that may be mediated by Nodal/Activin signaling. The regulation of Oct4 is well known for its importance in cell fate decisions and its downregulation during loss of Nodal/Activin signaling is significant not only as an impetus for trophectoderm differentiation but also reconciles the alternative role of Nodal/Activin signaling in maintaining self-renewal and pluripotency.
The molecular basis of extracellular signaling instructions governing differential cell fate decisions in the Nodal/Activin pathway has been postulated but not shown conclusively. Primarily, the transcriptional events occurring at the interface between pSmad2 signal transduction from the activated cell surface receptors to manipulation of the global stem cell transcriptome driving specific lineage programs have not been well characterised. This study provides an important insight into how quantitative signaling is translated into qualitative cell fate decisions by showing for the first time, to our knowledge, that the same transcription factor pSmad2 is able to bind and transcriptionally regulate different subsets of target genes in a dose-dependent manner.
The specification of cell fate decisions is governed by 2 distinct events. The first requires an exit from self-renewal and maintenance of stemness programs by direct control of pSmad2 over key pluripotency factors. Previous studies have revealed that Nanog is a direct target of Smad2/3 transcription in human ES cells [61]. Here we show an additional level of control over the stem cell program by direct transcriptional regulation of the Oct4 master pluripotency gene by pSmad2. The second event requires an entry into a specific differentiation program that is in turn brought about by direct and indirect pSmad2 regulation of differentiation genes such as Gata3, Tcfap2c and Igf2 that are known to be important factors for trophectoderm cell fates. This cell fate decision is further reinforced by loss of Oct4 with inhibition of Smad2 phosphorylation as the former is known to be a potent repressor of the trophectoderm gene Cdx2 in the blastocyst [60]. The pSmad2 binding target genes driving mesendodermal differentiation include Mixl, Fgf8 and Nodal itself, while other genes such as Chst15 expressed in definitive endoderm and Fgf15 for cardiac mesoderm have also been identified as strong Nodal/Activin transcriptional targets. It is likely that over the course of long-term differentiation for 6 days, additional target genes may be recruited for the specification of both lineage decisions that may not be apparent at the 18 hours time point in this study which may be too early for endpoint differentiation. Indeed, strong regulation of Mixl and Fgf8 and to a lesser extent Nodal could be detected at 3 and 6 days (Figure 2B and data not shown) of treatment in correlation with the level of Nodal/Activin signaling.
Consistent with the role of Nodal/Activin as morphogens, we found that many components of the pathway were themselves feedback targets that were directly regulated by pSmad2 binding in ChIP-Seq and/or differentially expressed in our microarray analysis. These include the negative feedback inhibitors such as Lefty1/2 and Smad7 which are already known targets of Nodal/Activin signaling. In this study, graded pSmad2 binding could be detected in the intronic region of Tmepai (Figure S6) which sequesters Smad2/3/4 from receptor kinase activity [62]. Similarly, SnoN [63] and Ski [64] also present graded intronic binding of pSmad2 (Figure S6) and both function as transcriptional repressors of Smad2/3/4. There were also positive feedback components such as Nodal, its cofactor Cripto and FoxH1 the transcriptional copartner of Smad2 (Figure 8 and Figure S6) that show graded binding in the intron and promoter regions. The preponderance of the negative components in the autoregulatory loop of Nodal/Activin signaling is significant, as it suggests that the pathway mainly dampens and attenuates its own signaling via negative feedback and less so by positive feedback loops mediated by Nodal, Cripto and FoxH1.
One of the intriguing findings is that extracellular signaling gradients were translated into a gradient of Smad2 phosphorylation that we have now shown to be able to recruit different target genes in a dose-dependent manner. This was possibly achieved by an exchange of transcriptional copartners that permits the shifting of the pSmad2 transcriptional complex to different target gene subsets as suggested by the differential recruitment of non-CAGA motifs and comotifs under each signaling condition. The fact that pSmad2 contains only CAGA sequence binding domains and not transcription activation domains suggest that it is further dependent upon copartners for transcription, binding affinity and specificity. In some cases graded pSmad2 transcription complex binding drives graded target gene response that follows signaling strength with high fidelity. In other cases, the target genes are only regulated and responsive at defined signaling thresholds (Figure 8). The consequence is that a relatively modest stimulation with Activin leading to a physiological 2-fold increase in Smad2 phosphorylation eventually drives mesendodermal differentiation while the reciprocal SB inhibition resulting in a 2-fold decrease of pSmad2 is able to promote trophectoderm cell fates. During this process, the master regulator of pluripotency Oct4 is itself titrated by the same Nodal/Activin signaling gradients in the ES cells undergoing differentiation. Hence the same pathway is able to tilt the balance in favor of maintenance of pluripotency or mediate an exit from self-renewal and entry into a specific lineage program. In conclusion, this study for the first time, to our knowledge, reconciles the multiple divergent roles of Nodal/Activin signaling in both pluripotency and differentiation with pSmad2 playing a central role in the cell fate decision making process.
E14 TG2A mouse embryonic stem cells (ATCC) were propagated in FBS media consisting of 20% ES cell-qualified FBS in DMEM supplemented with 100 µM non-essential amino acids, 100 U/ml penicillin, 100 µg/ml streptomycin, 2 mM GlutaMAX-I (Invitrogen), 55 µM β-mercaptoethanol (Sigma) and 1X homemade Leukemia inhibitory factor (LIF). For the establishment of Nodal/Activin signaling gradients in chemically defined conditions, KSR media containing 20% Knockout Serum Replacement (KSR, Invitrogen) in place of FBS with all other components of ES media excluding LIF were used. For acute (0 to 48 hours) signaling conditions, 25000 ES cells/cm2 were plated for 18 hours in FBS media followed by adaptation of the cells to chemically defined conditions with 10 µM SB-431542 (Tocris) in KSR media for 6 hours as previously described [41]. High signaling was induced by treatment with KSR media containing 25 ng/ml Activin (R&D Systems) or low signaling with 10 µM SB and maintenance of endogenous signaling with control KSR media or 1/5000 dilution of DMSO vehicle as indicated. For long-term differentiation, 2000 ES cells/cm2 were plated and 18 hours later directly treated with Activin, DMSO and SB in FBS media without LIF or KSR media for 6 days with media change everyday. The DMSO vehicle used to dissolve SB can induce differentiation and loss of pluripotency in ES cells [65], [66]. In the microarray analysis of the 3 signaling conditions, the effect of DMSO on differential gene expression was determined by comparing against the unsupplemented KSR media control (Figure 3A). The SB inhibitor was used at a high stock concentration of 50 mM permitting 5000X dilution of DMSO in ES cell cultures which was well below the limit required for differentiation. The cultures and treatments were carried out for the microarray study in 4 biological replicates consisting of ES cells at 4 different passages from P20 to P24 to identify and eliminate any cell culture variation effects from analysis.
ES cells were lysed in RIPA buffer (150 mM NaCl, 1% NP-40, 0.5% Sodium Deoxycholate, 0.1% SDS, 50 mM Tris pH 8.0) for protein extracts. SDS-PAGE was performed on 10% polyacrylamide gels and transferred on Immun-Blot PVDF membranes (Bio-rad Laboratories) followed by probing with 1∶1000 dilutions of rabbit anti-Smad2 (pSer465/467, Calbiochem), rabbit anti-Smad2 (Invitrogen), mouse anti-Pcna (Santa Cruz Biotechnology) and goat anti-Oct4 (Santa Cruz Biotechnology). Secondary antibodies used were 1∶1000 donkey anti-rabbit IgG-HRP (GE Healthcare), 1∶1000 goat anti-mouse IgG-HRP and 1∶2500 donkey anti-goat IgG-HRP (Santa Cruz Biotechnology). Densitometry measurements of protein bands on western blots were acquired using Photoshop CS3 (Adobe Systems Incorporated).
For gene expression, total RNA was extracted from cells using the RNeasy Mini kit (Qiagen) as per manufacturer instructions. This was reverse transcribed into cDNA using the High Capacity RNA-to-cDNA Master Mix (Applied Biosystems). Quantitative real-time PCR was performed on the 7900HT Fast Real-Time PCR System (Applied Biosystems) or the Biomark System (Fluidigm Corporation) on cDNA or ChIP DNA according to manufacturer instructions. For RT-PCR, products amplified for 25 to 33 cycles were resolved on a 2.5% agarose gel. Primer sequences for both ChIP-qPCR and gene/marker expression can be found in Table S3.
Total RNA was reverse transcribed into cDNA and in vitro transcribed into biotin-labeled cRNA using the Illumina TotalPrep RNA Amplification kit (Ambion). This was hybridized on MouseRef-8 v2.0 Expression BeadChips (Illumina). Raw intensity values were subjected to background subtraction on the BeadStudio Data Analysis Software (Illumina) and normalized using the cross-correlation method [67]. Differential gene expression was identified based on a fold change cutoff of >1.5 compared to the DMSO control. The microarray data was deposited in NCBI GEO with accession number GSE23239.
Chromatin Immunoprecipitation (ChIP) using a certified ChIP-grade rabbit polyclonal anti-Smad2 (phospho S465+S467) antibody (ab16509, Abcam) was carried out in ES cells under chemically defined high, medium or low Nodal/Activin signaling conditions according to the Agilent Mammalian ChIP-on-chip protocol v9.1 up to the ChIP DNA purification step. Adapter ligation, library amplification and size selection in the 200–300 bp range were performed according to the Illumina ChIP Sample Prep protocol (#11257047, Rev. A). Massively parallel sequencing was carried out for ChIP samples in all 3 signaling conditions with their respective input DNA controls on the Genome Analyzer (Illumina) up to a sequencing depth of at least 10×106 tags pass filter. The ChIP-Seq data was deposited in NCBI GEO with accession number GSE23581. Details of the ChIP-Seq, motif and statistical analysis can be found in the Text S1.
The 503 bp fragment of the mouse Oct4 promoter region corresponding to chr17:35640683–35641185 was cloned into the pGL4.23[luc2/minP] Firefly luciferase reporter construct (Promega) to generate pGL4.23 Oct4. This construct was point mutated by oligo cloning into unique StuI and NsiI sites to produce pGL4.23 m4 Oct4 (CAGAC mutated to CATGC) and pGL4.23 m345 Oct4 (TCTGGGCAGACGGCAGA mutated to TATGGGCATGCGGCATA). The constructs were transfected into mouse ES cells in an 80∶1 ratio with the pGL4.75[hRluc/CMV] Renilla luciferase co-transfection control (Promega) using Lipofectamine 2000 (Invitrogen). A control transfection was included with an 80∶1 ratio of the empty pGL4.23 vector to pGL4.75 Renilla control. Immediately after lipofections, the ES cells were pretreated with FBS media without LIF and with 10 µM SB for 6 hours. Subsequently the cells were split into replicates and plated in FBS media without LIF containing 25 ng/ml Activin, 1/5000 DMSO vehicle or 10 µM SB, which induces high, medium or low Nodal/Activin signaling respectively for 8 hours. The cells were washed once in PBS and lysed in 1XPassive Lysis Buffer and luciferase assays were performed using the Dual-Luciferase Reporter Assay System on the GloMax 96 Microplate Luminometer with Dual Injectors (Promega).
|
10.1371/journal.ppat.1002978 | Deciphering the Acylation Pattern of Yersinia enterocolitica Lipid A | Pathogenic bacteria may modify their surface to evade the host innate immune response. Yersinia enterocolitica modulates its lipopolysaccharide (LPS) lipid A structure, and the key regulatory signal is temperature. At 21°C, lipid A is hexa-acylated and may be modified with aminoarabinose or palmitate. At 37°C, Y. enterocolitica expresses a tetra-acylated lipid A consistent with the 3′-O-deacylation of the molecule. In this work, by combining genetic and mass spectrometric analysis, we establish that Y. enterocolitica encodes a lipid A deacylase, LpxR, responsible for the lipid A structure observed at 37°C. Western blot analyses indicate that LpxR exhibits latency at 21°C, deacylation of lipid A is not observed despite the expression of LpxR in the membrane. Aminoarabinose-modified lipid A is involved in the latency. 3-D modelling, docking and site-directed mutagenesis experiments showed that LpxR D31 reduces the active site cavity volume so that aminoarabinose containing Kdo2-lipid A cannot be accommodated and, therefore, not deacylated. Our data revealed that the expression of lpxR is negatively controlled by RovA and PhoPQ which are necessary for the lipid A modification with aminoarabinose. Next, we investigated the role of lipid A structural plasticity conferred by LpxR on the expression/function of Y. enterocolitica virulence factors. We present evidence that motility and invasion of eukaryotic cells were reduced in the lpxR mutant grown at 21°C. Mechanistically, our data revealed that the expressions of flhDC and rovA, regulators controlling the flagellar regulon and invasin respectively, were down-regulated in the mutant. In contrast, the levels of the virulence plasmid (pYV)-encoded virulence factors Yops and YadA were not affected in the lpxR mutant. Finally, we establish that the low inflammatory response associated to Y. enterocolitica infections is the sum of the anti-inflammatory action exerted by pYV-encoded YopP and the reduced activation of the LPS receptor by a LpxR-dependent deacylated LPS.
| Lipopolysaccharide (LPS) is one of the major surface components of Gram-negative bacteria. The LPS contains a molecular pattern recognized by the innate immune system. Not surprisingly, the modification of the LPS pattern is a virulence strategy of several pathogens to evade the innate immune system. Yersinia enterocolitica causes food-borne infections in animals and humans (yersiniosis). Temperature regulates most, if not all, virulence factors of yersiniae including the structure of the LPS lipid A. At 21°C, lipid A is mainly hexa-acylated and may be modified with aminoarabinose or palmitate. In contrast, at 37°C, Y. enterocolitica expresses a unique tetra-acylated lipid A. In this work, we establish that Y. enterocolitica encodes a lipid A deacylase, LpxR, responsible for the lipid A structure expressed by the pathogen at 37°C, the host temperature. Our findings also revealed that the low inflammatory response associated to Y. enterocolitica infections is the sum of the anti-inflammatory action exerted by a Yersinia protein translocated into the cytosol of macrophages and the reduced activation of the LPS receptor complex due to the expression of a LpxR-dependent deacylated LPS.
| Lipopolysaccharide (LPS) is one of the major surface components of Gram-negative bacteria. The molecular structure of LPS is rather unique: an amphiphile with a hydrophobic region, the so-called lipid A, adjacent to a densely negatively charged polysaccharide. In Escherichia coli K-12, the lipid A is a β(1′-6)-linked disaccharide of glucosamine phosphorylated at the 1 and 4′ positions with positions 2, 3, 2′, and 3′acylated with R-3-hydroxymyristoyl groups, the so-called lipid IVA. The 2′and 3′R-3-hydroxymyristoyl groups are further acylated with laureate (C12) and myristate (C14), respectively, by the action of the so-called late acyltransferases LpxL (HtrB) and LpxM (MsbB), respectively [1]. When E. coli is grown at 12°C, LpxP, the cold-temperature-specific late acyltransferase, acts instead of LpxL adding palmitoleate (C16∶1) [1]. Although the enzymes required to synthesize the lipid A are conserved throughout all Gram-negative bacteria there is heterogeneity on lipid A structure among Gram-negative bacteria compared to the E. coli K-12. This is due to differences in the type and length of fatty acids, in the presence of decorations such as aminoarabinose or phosphoethanolamine and even in the removal of groups such as phosphates or fatty acids from lipid A [2].
LPS plays a crucial role during recognition of microbial infection by the host immune system. In fact, the lipid A moiety is a ligand of the Toll-like receptor 4 (TLR4)/myeloid differentiation factor 2 complex [3]. The stimulation of this receptor complex triggers the activation of signalling cascades resulting in the induction of antimicrobial genes and release of cytokines, thereby initiating inflammatory and immune defence responses. Perusal of the literature demonstrates that changes in the number of acyl chains and in the phosphorylation status of the headgroup greatly affect the biological activity of lipid A. It is not surprising that some pathogens modulate their lipid A structure to alter their detection by the host; being these regulated changes important virulence traits (for a review see [4]). Furthermore, given the importance of the LPS structure to the homeostasis of the outer membrane, it is possible that the aforementioned changes may also affect the physiology of the outer membrane as was recently demonstrated for Salmonella [5].
The genus Yersinia includes three human pathogens: Y. pestis, Y. pseudotuberculosis and Y. enterocolitica. The latter can cause food-borne infections in animals and humans (yersiniosis), with symptoms such as enteritis and mesenteric lymphadenitis [6]. Y. enterocolitica is endowed with a repertoire of virulence factors that help bacteria to colonize the intestinal tract and to resist host defence mechanisms [7], [8]. Temperature regulates most, if not all, virulence factors of yersiniae [7], [8]. Recent studies have shown that temperature also regulates the structure of yersiniae lipid A [9]–[14]. Thus the number and type of the lipid A fatty acids and the substitutions of the 1- and 4′-positions in the glucosamine disaccharide can vary. Rebeil and co-workers [12] elegantly demonstrated that a shift in temperature induces a change in the number and type of acyl groups on the lipid A of the three Yersinia species. At 21°C, lipid As are mainly hexa-acylated whereas at 37°C they are tetra-acylated [12]. The temperature-dependent regulation of the lipid A acyltransferases underlines the shift in lipid A acylation both in Y. pestis and in Y. enterocolitica [12], [14]. Pathogenic yersiniae also express hepta-acylated lipid A due to the addition of C10, in Y. pestis and Y. pseudotuberculosis, or C16 (palmitate), in Y. enterocolitica [12], [14], [15]. PagP is the acyltransferase responsible for the addition of palmitate to the lipid A in Y. enterocolitica [15]. Other lipid A species are consistent with the substitution of the phosphate at the 4′ end of the glucosamine disaccharide with aminoarabinose [15]. The aminoarabinose content is temperature-regulated in Y. pestis and in Y. enterocolitica [12], [15], [16]; being higher in bacteria gown at 21°C than at 37°C. Similar to other Gram-negative bacteria, the products of ugd and pmrHFIJKLM (arnBCADTEF)(hereafter pmrF operon) are required for the synthesis and addition of aminoarabinose to lipid A in Y. enterocolitica [15]. Finally, we and others [9]–[14], [17] have reported a unique tetra-acyl lipid A species (m/z 1388) found only in Y. enterocolitica grown at 37°C. Evidence support the notion that this species lacks the ester-linked R-3-hydroxymyristoyl group further acylated with laureate (C12) [12], [14], [17]. Indeed, mass spectrometry analysis did confirm that the nonreducing glucosamine of the lipid A is substituted with only one (amide-linked) R-3-hydroxymyristoyl group further acylated with myristate (C14) [17]. Altogether, these findings strongly suggest that the tetra-acyl lipid A species (m/z 1388) may be caused by a deacylase removing the 3′-acyloxyacyl residue of the lipid A. The work described in this article gives experimental support to this hypothesis and explores the impact of the lipid A structure on Y. enterocolitica virulence traits.
Further confirming previous findings [14], [15], lipid A isolated from Y. enterocolitica 8081 serotype O:8 (hereafter YeO8; Table 1) grown at 37°C appeared to be identical to those reported by Rebeil et al. and Oertelt et al. [12], [17]. The main species were a 3′-O-deacylated form (m/z 1388) containing two glucosamines, two phosphates, three 3-OH-C14, and one C14; and a hexa-acylated form (m/z 1797) (Figure 1A). In bacteria grown at 21°C, a minor species (m/z 1414) was detected and may represent a 3′-O-deacylated form containing three 3-OH-C14 and one C16∶1 [12]. S. enterica serovar typhimurium and Helicobacter pylori also express 3′-O-deacylated lipid A species [18], [19]. A membrane located hydrolase, named LpxR, removes the 3′-acyloxyacyl residue of lipid A in both organisms [18], [19]. In silico analysis of the YeO8 genome (accession number AM286415; [20]) revealed that this pathogen may encode an LpxR orthologue (locus tag YE3039). The predicted YeO8 LpxR (YeLpxR) has 73% and 20% amino acid identities to S. enterica and H. pylori LpxR proteins, respectively. Furthermore, YeLpxR has 100% amino acid identity to Y. enterocolitca Y11 serotype O:3 (locus tag Y11_05741; accession number FR729477) and Y105 serotype O:9 (locus tag YE105_C2442; accession number CP002246) LpxR homologs. Analysis of the available Y. pestis and Y. pseudotuberculosis genomes revealed that they do not encode any gene similar to lpxR.
YeO8 lpxR was mutated to determine whether this gene is indeed responsible for removing the 3′-acyloxyacyl residue of lipid A. MALDI-TOF mass spectrometry studies showed that, at 37°C, the lpxR mutant (YeO8-ΔlpxRKm) produced a lipid A which lacked the unique tetra-acyl lipid A species (m/z 1388) found in YeO8 and only contained the hexa-acylated species (m/z 1797; four 3-OH-C14, one C12 and one C14) (Figure 1B). At 21°C, lipid A isolated from YeO8-ΔlpxRKm was similar to that of YeO8 although without the minor species m/z 1414 (Figure 1B). Complementation of the mutant with pTMLpxR restored the presence of the tetra-acyl species (Figure 1C).
In summary, our results confirmed the predicted function of Y. enterocolitica O:8 lpxR homolog as the lipid A 3′-O-deacylase.
The LpxR-dependent lipid A deacylation was more evident on bacteria grown at 37°C than at 21°C, hence suggesting that the expression and/or function of the deacylase might be temperature-regulated, being higher at 37°C than at 21°C. To monitor transcription of lpxR quantitatively, a transcriptional fusion was constructed in which a promoterless lucFF gene was under the control of the lpxR promoter region (see Material and Methods); thereafter lpxR::lucFF was introduced into YeO8 and the luciferase activity was determined. The expression of the fusion was higher at 21°C than at 37°C (Figure 2A). Real time (RT) quantitative PCR (RT-qPCR) experiments showed that lpxR mRNA levels were also higher at 21°C than at 37°C (Figure 2B).
To assess LpxR levels, the C-terminus of the protein was tagged with a FLAG epitope and the construct was cloned into the medium-copy plasmid pTM100 to obtain pTMLpxRFLAG (see Materials and Methods). This plasmid restored the presence of the tetra-acyl species (m/z 1414 and m/z 1388) in the lipid A of YeO8-ΔlpxRKm (data not shown). Western blot analysis of purified membranes from YeO8-ΔlpxRKm containing pTMLpxRFLAG showed that LpxR levels were higher in membranes from bacteria grown at 21°C than at 37°C (Figure 2C). Altogether, it can be concluded that the expression of lpxR is indeed temperature-regulated but, in contrast to our initial hypothesis, its expression is higher at 21°C than at 37°C.
The apparent contradiction between the mass spectrometry analysis, more deacylation at 37°C, and the Western blot data, higher levels of LpxR at 21°C than at 37°C, led us to explore whether low temperature may affect the function of the enzyme. Since E. coli has been used as surrogate host to characterize Salmonella LpxR (StLpxR) function [18], we mobilized pTMLpxR into E. coli MG1655 to analyze lipid A species by mass spectrometry in bacteria grown at 21°C and 37°C. Results shown in figure 3 demonstrate that LpxR did deacylate the E. coli lipid A from bacteria grown either at 21 or 37°C as detected by the presence of species m/z 1360 (Figure 3C–D). This species was found previously in E. coli expressing StLpxR [18]. Of note, the species m/z 1414, which is consistent with the deacylation of the species m/z 1850 containing palmitoleate (C16∶1) instead of laureate (C12), was observed only in E. coli grown at 21°C. LpxP is the cold-temperature-specific late acyltransferase responsible for the addition of palmitoleate [1]. Altogether, our results indicate that the reduced LpxR-dependent deacylation found in YeO8 grown at 21°C cannot be attributed to a general lack of function of the enzyme at this temperature.
We sought to determine why LpxR activity was not observed in YeO8 grown at 21°C despite the detection of the enzyme in the membrane. Among other possibilities, we speculated that specific features of YeO8 lipid A found only at 21°C might be responsible for the reduced LpxR activity. Furthermore, these features should be absent in E. coli grown at 21°C since LpxR-dependent activity was observed here. A conspicuous difference between YeO8 and E. coli lipid As is the presence of aminoarabinose and palmitate (m/z 1954 and 2063, respectively) only in the former [14], [15]. Therefore, we explored whether any of these modifications could account for the reduced LpxR activity. In YeO8, similarly to other Gram-negative pathogens, the products of the pmrF operon are required for the synthesis and addition of aminoarabinose to lipid A whereas the acyltransferase PagP is required for the addition of palmitate to lipid A [15]. The lipid A from the pagP mutant, YeO8-ΔpagPGB, grown at 21°C resembled that of the wild-type strain, except that the species containing palmitate (m/z 2063) was not detected (Figure 4A). In contrast, the tetra-acylated species (m/z 1414) was clearly observed in the lipid A from YeO8-ΔpmrF grown at 21°C (Figure 4C). This was dependent on LpxR activity since the peak was absent in the double mutant YeO8-ΔpmrF-ΔlpxRKm (Figure 4E). LpxR-dependent deacylation of lipid A (m/z 1388) observed in bacteria grown at 37°C was not affected in either pmrF or pagP single mutants (Figure 4B, D). Control experiments revealed that lpxR expression was not affected in YeO8-ΔpmrF since the expression of the lpxR::lucFF fusion was not significantly different between YeO8 and the pmrF mutant either grown at 21°C or at 37°C (Figure 4G).
On the whole, these results are consistent with the notion that the reduced LpxR activity observed in YeO8 at 21°C is associated with the lipid A modification with aminoarabinose.
Our findings might suggest that aminoarabinose-containing LPS may directly inactivate the lipid A deacylase activity of YeLpxR. Alternatively, modification of lipid A with aminoarabinose could inhibit the physical interaction of LPS with YeLpxR. To explore this, the 3-D structure of YeLpxR was modeled (Figure 5A). The amino acids 1–296 (following the putative signal sequence) could be modeled based on the crystal structure of StLpxR (PDB code 3FID; [21]) and the sequence alignment between StLpxR and YeLpxR (Figure S1). The fold of the resulting model is likely to be of good quality, since YeLpxR has such a high sequence identity to StLpxR (75%). Additionally, the important StLpxR amino acids identified by Rutten and co-workers [21] are conserved in YeLpxR. Six amino acids differ between the YeLpxR and the StLpxR active sites (Figure S1). Major differences are D31 and Q35 in YeLpxR, of which D31 is closer to the active site (Figure 5B). The corresponding amino acids are much smaller in StLpxR, glycine and an alanine, respectively, which cause StLpxR to have a bigger cavity. StLpxR has a protruding cavity close to K67, which cannot be found in YeLpxR (Figure 6A). The difference is induced by D31 in YeLpxR, which occupies more space than G31 in StLpxR. As a consequence, the conserved K67 adopts a different conformation in the YeLpxR model. Due to D31, the cavity in YeLpxR is divided into two parts with a narrow connection, and this amino acid also prevents YeLpxR from forming an inward protruding cavity similar to the one found near G31 in StLpxR (Figure 6A).
Docking of a modified Kdo2-lipid A molecule (see Materials and methods) to the model of YeLpxR showed that the phosphate group, which attaches aminoarabinose to Kdo2-lipid A, binds into the cavity in the vicinity of K67 and D31 (Figure 6B). Docking of the same molecule to the crystal structure of StLpxR yielded a result where the phosphate group was located in the protruding cavity close to K67 (Figure 6C). As expected, docking of the modified Kdo2-lipid A molecule with aminoarabinose to the YeLpxR model did not give any valuable result. On the other hand, when the same molecule was docked to the StLpxR crystal structure, aminoarabinose was bound close to G31. It occupies the space corresponding to the narrow connection of the two larger cavities in YeLpxR (Figure 6D)
As a result from the modeling and docking studies, we suggest that Kdo2-lipid A with aminoarabinose cannot fit into the active site of YeLpxR due to D31, hence leading to the inability of YeLpxR to deacylate Kdo2-lipid A with aminoarabinose.
To confirm our predictions, we constructed LpxR mutants by site-directed mutagenesis (see Material and Methods). In addition to the amino acids corresponding to the active site amino acids in StLpxR, we wanted to study the effect of the D31G mutation for YeLpxR as the modelling and docking studies suggested that D31 has an important role in the YeLpxR specificity for the Kdo2-lipid A species. The constructs were introduced into E. coli MG1655 and the lipid A from the transformants grown at 37°C was analyzed by MALDI-TOF mass spectrometry. Most of the constructs containing LpxR mutants did trigger the deacylation of E. coli lipid A, detected by the presence of species m/z 1360, (Table 2). In contrast, constructs containing LpxR mutants, LpxR(N9A), LpxR(D10A), LpxR(S34A), and LpxR(H122A) did not deacylate E. coli lipid A. These results were expected since Rutten and co-workers have reported that these residues are located in the StLpxR active site and all of them are conserved in LpxR homologues [21]. Next, only those constructs triggering deacylation of E. coli lipid A were introduced into YeO8. When the YeO8 strains were grown at 37°C, all LpxR mutants restored the presence of the tetra-acyl species (m/z 1388) in the lipid A of YeO8-ΔlpxRKm (Table 2). Additionally, the mass spectrometry analysis revealed that LpxR(D31G) mutant did trigger the deacylation of lipid A in bacteria grown at 21°C as it was detected the presence of lipid A species m/z 1414 and m/z 1545 (Figure 7B). The latter is consistent with the deacylation of the lipid A species modified with aminoarabinose (m/z 1954).
In summary, our results further confirmed the amino acids important for the catalytic activity of YeLpxR. Moreover, our results confirmed the molecular modelling predictions, thereby demonstrating that the presence of D31 in the active site pocket of YeLpxR causes steric hindrance for the binding and deacylation of lipid A species modified with aminoarabinose.
In YeO8 the expression of the loci responsible for the lipid A modification with aminoarabinose, ugd and pmrF operon, is temperature regulated, being higher at 21°C than at 37°C [15]. Mechanistically, this is so because the expression of the positive regulators phoPQ and pmrAB, which control the expression of ugd and the pmrF operon, is also higher at 21°C than at 37°C [15]. In turn, the temperature-dependent regulation of phoPQ and pmrAB is explained by H-NS-dependent negative regulation alleviated by RovA, another major regulator of Yersinia [22], [23], at 21°C [15]. Moreover, there is cross-talk between the regulators in such way that PhoPQ and PmrAB regulate positively the expression of rovA and the effect of PhoPQ is more important [15].
The inverse correlation between the substitution of the lipid A with aminoarabinose and lipid A deacylation, prompted us to evaluate whether phoPQ and pmrAB might negatively regulate lpxR. Results shown in figure 8 revealed that the expression of lpxR::lucFF was significantly up-regulated in the phoPQ and pmrAB mutants at 21°C and 37°C (Figure 8A). However, the expression of lpxR reached wild-type levels in the double phoPQ-pmrAB mutant regardless the bacteria growth temperature (Figure 8A). RT-qPCR experiments showed that the levels of lpxR mRNA were higher in the phoPQ and pmrAB mutants than in the wild type and double phoPQ-pmrAB mutants, which were not significantly different (Figure S2).
Recently, we have shown that rovA expression is downregulated in the phoPQ and pmrAB single mutants, being the lowest in the phoPQ mutant, whereas in the phoPQ-pmrAB double mutant rovA expression is not significantly different to that in the wild type [15]. Therefore, the fact that lpxR expression follows the opposite trend in these mutants led us to analyze whether rovA negatively regulates the expression of lpxR. Indeed, luciferase activity was higher in the rovA mutant than in the wild type and the levels were not significantly different that those observed in the phoPQ mutant when bacteria were grown either at 21°C or 37°C (Figure 8A). Similar results were obtained when the lpxR mRNA levels were analyzed by RT-qPCR (Figure S2). The increased lpxR expression observed in rovA and phoPQ single mutants at 21°C was no longer found in the double mutant rovA-phoPQ (Figure 8A and Figure S2). When bacteria were grown at 37°C, lpxR expression in the rovA-phoPQ mutant was significantly lower than those observed in the rovA and phoPQ single mutants (p<0.05 for each comparison versus rovA-phoPQ mutant) although still higher than that in the wild type (Figure 8A and Figure S2). Of note, the expression of lpxR was no longer temperature regulated in the rovA-phoPQ mutant (Figure 8B). The fact that the expression of lpxR::lucFF in the triple mutant rovA-phoPQ-pmrAB at 21°C was less than in the wild-type strain may support the notion that, in the absence of the negative regulator RovA, PmrAB and/or a PmrAB-modulated regulator positively regulates lpxR. At 37°C, lpxR expression in the triple mutant was not significantly different than those found in the double mutant phoPQ-pmrAB and the wild type (Figure 8A and Figure S2).
Collectively, our data revealed that the expression of lpxR is negatively controlled by the same regulators that activate the loci necessary for the substitution of the phosphate at the 4′ end of the glucosamine disaccharide with aminoarabinose.
In a previous study, we observed the down regulation of YeO8 virulence factors in mutants lacking the lipid A late acyltransferases LpxM, LpxL or LpxP [14]. These results raised the possibility that lipid A acylation may act as a regulatory signal by acting on a transduction pathway(s) [14]. In this context, we sought to determine the impact of LpxR to the expression/function of YeO8 virulence factors.
Virulence genes can be regulated as part of the flagellar regulon, indicating that this regulon contributes to Y. enterocolitica pathogenesis [24]. YeO8 is motile when grown at 21°C but not at 37°C [25] and previously we showed that LpxM and LpxP mutants are less motile than the wild type [14]. We examined the influence of LpxR on the flagellar regulon. We quantified the migration of the wild type and YeO8-ΔlpxRKm in motility medium (1% tryptone-0.3% agar plates). Figure 9 shows that YeO8-ΔlpxRKm was less motile than the wild type. Yersinia motility is related to the levels of flagellins which, in turn, are regulated by the expression of flhDC, the flagellum master regulatory operon [25], [26]. We hypothesized that the expression of flhDC could be lower in the lpxR mutant than in the wild type. To address this, the flhDC::lucFF transcriptional fusion [26] was introduced into the chromosome of the strains and the luciferase activity was determined. At 21°C, luminescence was lower in the lpxR mutant than in the wild type (Figure 9B). Complementation of the lpxR mutant with pTMYeLpxR restored flhDC::lucFF expression to wild-type levels (Figure 9B). Notably, the catalytic inactive LpxR mutants LpxR(N9A) and LpxR(S34A), encoded by pTMLpxR(N9A) and pTMLpxR(S34A) respectively, also complemented the lpxR mutant (Figure 9B). Western blot analysis of purified membranes from YeO8-ΔlpxRKm containing pTMLpxR(N9A)FLAG or pTMLpxR(S34A)FLAG showed that the mutant proteins were expressed (Figure S3).When the strains were grown at 37°C, YeO8 and YeO8-ΔlpxRKm produced the same luminescence (Figure 9B).
One virulence gene that is regulated as part of the flagellar regulon is yplA and hence its expression is regulated by flhDC [24], [27], [28]. Considering that flhDC expression was downregulated in the lpxR mutant, we speculated that yplA expression could be affected in this mutant. The transcriptional fusion yplA::lacZYA [29] was introduced into the chromosome of the wild type and the lpxR mutant and their β-galactosidase activities were measured. Indeed, the β-galactosidase activity was lower in YeO8-ΔlpxRKm than in the wild type (Figure 9C). Plasmids pTMYeLpxR, pTMLpxR(N9A) and pTMLpxR(S34A) complemented the phenotype (Figure 9C).
In summary, these results indicate that the flagellar regulon is downregulated in the lpxR mutant with a concomitant decrease in motility and downregulation of yplA expression.
Inv is an outer membrane protein of Y. enterocolitica responsible for invasion of the host [30], [31]. Since YeO8 lipid A mutations affect inv expression [14], we asked whether inv expression is altered in the lpxR mutant. An inv::phoA translational fusion [32] was introduced into the genome of YeO8 and YeO8-ΔlpxRKm and inv expression was monitored as alkaline phosphatase (AP) activity (Figure 10A). AP activity was significantly lower in the lpxR mutant than in the wild type. Plasmids pTMYeLpxR, pTMLpxR(N9A) and pTMLpxR(S34A) restored AP activity to wild-type levels. These differences in inv expression prompted us to study the ability of YeO8-ΔlpxRKm to invade HeLa cells by using a gentamicin protection assay. The amount of intracellular bacteria was 55% lower when cells were infected with the lpxR mutant than with the wild type (Figure 10B).
RovA is required for inv expression in Y. enterocolitica [33]. Therefore, among other possibilities, the low inv expression found in the lpxR mutant could be caused by downregulation of rovA expression. To address this, the rovA::lucFF transcriptional fusion [14] was introduced into the genome of the wild type and the lpxR mutant and the luminescence was determined. Results shown in figure 10C demonstrate that rovA expression was dowregulated in YeO8-ΔlpxRKm. This phenotype was complemented with plasmids pTMYeLpxR, pTMLpxR(N9A) and pTMLpxR(S34A).
Together, our data show that the down-regulation of inv expression found in the lpxR mutant is most likely caused by downregulation of rovA expression, the positive transcriptional regulator of inv.
Y. enterocolitica harbours a plasmid (pYV)-encoded type III secretion system which is required for virulence. A set of virulence factors, called Yops, are secreted by this system and enable Y. enterocolitica to multiply extracellularly in lymphoid tissues [34]–[36]. In several pathogens, LPS polysaccharide status affects the expression of the type III secretion systems [37]–[39]. Therefore, we asked whether the production of the Yersinia pYV-encoded type III secretion system is altered in the lpxR mutant. At 37°C and under low calcium concentrations, this system secretes the Yops to the culture supernatant [40]. Analysis of Yop secretion revealed that the wild type and the lpxR mutant secreted similar levels of Yops (Figure 11A). We sought to determine whether the translocation of Yops to the cytosol of eukaryotic cells is affected in the lpxR mutant. Detection of cytoskeleton disturbances upon infection of epithelial cells is one of the most sensitive assays to establish Yop translocation [41]. The injection of YopE into the cytosol of A549 cells by wild-type bacteria induced disruption and condensation of the actin microfilament structure of the cells whereas this was not the case when cells were infected with YeO8-ΔyopE mutant (Figure 11B). YopE translocation to A549 cells was not affected in the lpxR mutant background (Figure 11C). As expected, A549 cells infected with YeO8-ΔlpxRKm displayed similar cytoskeleton disturbances than those cells infected with the wild type (Figure 11B).
yadA is another pYV-encoded virulence gene whose expression is only induced at 37°C [42]. YadA is an outer membrane protein mediating bacterial adhesion, bacterial binding to proteins of the extracellular matrix and complement resistance (for a review see [43]). Analysis of YadA expression by SDS-PAGE demonstrated that YeO8-ΔlpxRKm and YeO8 produced the same amount of the protein (Figure 11D). To assess YadA functionality, we asked whether the YadA-dependent binding to collagen is altered in the lpxR mutant. To this end, we analyzed the binding of YadA-expressing whole bacteria to collagen type I by immunofluorescence (see Material and Methods). In contrast to the negative control, a pYV-cured strain (YeO8c), YeO8 and YeO8-ΔlpxRKm bound to collagen without differences between them (Figure 11E–F).
Taken together, these results suggest that the production and function of the pYV-encoded virulence factors Yops and YadA are not altered in the lpxR mutant.
Cationic antimicrobial peptides (CAMPs) belong to the arsenal of weapons of the innate immune system against infections. In the case of Gram-negative bacteria, CAMPs interact with the lipid A moiety of the LPS [44]–[47] and lipid A modification is one of the strategies employed by Gram-negative bacteria to counteract the action of CAMPs. We and others have used polymyxin B as a model CAMP since it also binds to lipid A. Furthermore, resistance to this peptide reflects well the resistance to other mammalian peptides and correlates with virulence [48]–[51]. Therefore we evaluated the resistance of the lpxR mutant to polymyxin B. Results shown in figure 12A demonstrate that the mutant was as resistant as the wild type to the peptide when grown either at 21°C or at 37°C. Of note both strains were more susceptible to polymyxin B when grown at 37°C than at 21°C (Figure 12A).
The mammalian immune system recognizes and responds to E. coli LPS via the TLR4 complex, resulting in the synthesis and secretion of pro-inflammatory cytokines that recruit immune cells to the site of infection. The ability of LPSs to evoke inflammatory responses and the potency of them are directly related to the structure of the molecule. It has been reported that underacylated LPSs are less inflammatory than hexa-acylated ones, being the E coli lipid A (m/z 1797) the prototype of hexa-acylated LPSs [52]. Therefore, the dramatic changes in lipid A acylation displayed by the lpxR mutant at 37°C led us to evaluate the immunostimulatory properties of YeO8 and YeO8-ΔlpxRKm. As cellular read-out, we determined TNFα levels secreted by macrophages infected either with the wild type or the lpxR mutant grown at 21°C and 37°C. YeO8 and YeO8-ΔlpxR induced similar levels of TNFα although the levels induced by bacteria grown at 37°C were significantly lower than those triggered by bacteria grown at 21°C (p<0.05 for comparison of TNFα levels between temperatures for a given strain) (Figure 12B). This was dependent on the well known anti-inflammatory action of the pYV-encoded YopP [53], [54], since a yopP mutant grown at 37°C induced similar levels of TNFα than those induced by wild-type bacteria grown at 21°C (Figure 12B). Therefore we sought to determine whether YopP could be counteracting the inflammatory response induced by YeO8-ΔlpxRKm. Indeed, YeO8-ΔyopP-ΔlpxRKm induced the highest levels of TNFα(Figure 12B). Further sustaining this notion, the TNFα levels induced by the lpxR mutant cured of the pYV virulence plasmid grown at 37°C were significantly higher than those induced by the virulence plasmid negative wild-type strain but not different than the YeO8-ΔyopP-ΔlpxRKm-triggered TNFα levels (Figure 12B). Of note, the TNFα levels induced by the virulence plasmid negative wild-type strain grown at 37°C were significantly lower than those triggered by bacteria grown at 21°C hence further highlighting the importance of lipid A acylation on the immunostimulatory properties of YeO8.
Pathogenic yersiniae show a temperature-dependent variation in lipid A acylation [9]–[14]. At 21°C, Y. enterocolitica synthesizes hexa-acylated lipid A containing four 3-OH-C14, one C12 and either one C16∶1 or one C14. At 37°C, Y. enterocolitica lipid A presents a tetra-acylated species (m/z 1388) and a hexa-acylated one containing four 3-OH-C14, one C12 and C14. In a previous work, we identified and characterized the acyltransfreases, lpxM, lpxL and lpxP, responsible for the addition of C12, C14 and C16∶1, respectively, to lipid A [14]. Moreover, we demonstrated that the expressions of these enzymes are temperature regulated [14]. However, the unique tetra-acyl lipid A found in the wild type grown at 37°C (m/z 1388) remained to be explained at the molecular level. We and others have established that this species is consistent with 3′-O-deacylation of lipid A [12], [14], [17]. In this work by combining biochemistry, genetics and molecular modelling we present evidence that LpxR is the lipid A 3′-O-deacylase of Y. enterocolitica.
YeLpxR is one of the closest homologues to StLpxR. Despite the presence of StLpxR in the Salmonella outer membrane, the bacterium does not produce 3′-O-deacylated lipid A species under any growth conditions tested to date [18]. This has been termed as enzyme latency and similar findings have been reported for the Salmonella lipid A 3-O-deacylase PagL and E. coli PagP [55], [56]. Our data revealed that YeLpxR is also latent in the membrane of YeO8 grown at 21°C. However, this is not a general feature of lipid A deacylases since H. pylori LpxR is constitutively active [19]. Several explanations could underlie YeLpxR latency at 21°C. Firstly, we explored whether low temperature may affect the function of the enzyme. The fact that YeLpxR did deacylate E. coli lipid A when grown at 21°C does not support that low temperatures grossly inhibit the enzyme activity. Nevertheless, we do not by any means completely rule out that temperature may affect YeLpxR activity, and thorough biochemical analyses are warranted to rigorously define the functional parameters of YeLpxR activity. This will be the subject of future studies. We next hypothesized that specific features of YeO8 lipid A, which do not exist in the E. coli lipid A, may be responsible for YeLpxR latency. The first conspicuous difference is the type of secondary fatty attached to the lipid IVA. In E. coli the late acyltransferases LpxL and LpxM add laureate (C12) and myristate (C14) respectively [1] whereas in YeO8 these enzymes transfer myristate (C14) and laureate (C12) respectively [14]. However, this cannot account for the reduced LpxR activity since the enzyme did deacylate E. coli lipid A. The presence of palmitoleate in YeO8 lipid A at 21°C but not at 37°C cannot be the reason since YeLpxR deacylated E. coli lipid A containing palmitoleate, found in E. coli grown at 21°C. Instead, our results revealed that the lipid A substitution with aminoarabinose is associated with YeLpxR latency since LpxR-dependent lipid A deacylation was clearly observed in the pmrF mutant grown at 21°C. Notably, the lack of aminoarabinose also releases Salmonella PagL from latency [56], hence suggesting a key role for the lipid A modification with aminoarabinose in LPS remodelling.
The molecular modelling and docking experiments further highlighted the importance of lipid A substitution with aminoarabinose for YeLpxR function. D31 in YeLpxR forces the conserved K67 to adopt a different conformation compared to StLpxR. According to the docking results, the resulting loss of cavity space in the vicinity of K67 in YeLpxR, causes the phosphate at the 4′ end of Kdo2-lipidA to bind somewhat differently to YeLpxR than to StLpxR. In the latter, the phosphate binds in the cavity near K67, while in YeLpxR it is forced to bind more outwards from the enzyme. The docking of Kdo2-lipidA with aminoarabinose to StLpxR showed that aminoarabinose occupies the cavity space, which corresponds to a narrow connection between two larger cavities in YeLpxR. The large reduction in cavity volume at this particular site causes this space to be too small for the accommodation of aminoarabinose. Hence, D31 seems to cause steric hindrance for the binding of aminoarabinose-containing Kdo2-lipidA to YeLpxR. Therefore, we predicted that D31 could have an important role for the YeLpxR substrate specificity. Indeed, the site-directed mutagenesis experiments validated that the presence of D31 in the active site pocket of YeLpxR causes a steric hindrance for the binding and deacylation of lipid A species modified with aminoarabinose. Nevertheless, at present we do not rule out that other residues of YeLpxR also contribute to its latency. In this regard, Salmonella PagL is released from latency when specific amino acid residues located at extracellular loops of the enzyme are mutated and it has been postulated that these residues are involved in the recognition of aminoarabinose-modified lipid A [56]–[58]. Studies are going to explore whether residues located at extracellular loops of LpxR also contribute to enzyme latency.
The inverse correlation between the aminoarabinose content in the LPS and the LpxR-dependent lipid A deacylation prompted us to evaluate whether the same regulatory network governing the expression of the pmrF operon and ugd could regulate lpxR. Recently, we have shown that the global regulators RovA, PhoPQ, and PmrAB positively control the expression of the loci necessary for aminoarabinose biosynthesis at 21°C [15]. Furthermore, there is a cross-talk between these regulators since the expressions of phoPQ and pmrAB are downregulated in the rovA mutant whereas rovA expression is downregulated in phoPQ and pmrAB single mutants [15]. Our findings support the notion that RovA and PhoPQ are negative regulators of lpxR since its expression was higher in phoPQ and rovA single mutant backgrounds than in the wild type. In turn, the two-component system PmrAB and/or a PmrAB-regulated system may act as a positive regulator because lpxR expression was similar in the wild-type and rovA-phoPQ backgrounds.
One striking finding of our study is that motility and invasion of eukaryotic cells were reduced in the lpxR mutant grown at 21°C. Mechanistically, our data revealed that the expressions of flhDC and rovA, the key regulators controlling the flagellar regulon and invasin respectively [22], [25], [33], were down-regulated in the lpxR mutant. Although we have reported that lipid A acylation status affects motility and invasion [14], the phenotypes were found in mutants lacking the late-acyltransferases and hence displaying major changes in the lipid A structure at 21°C [14]. This is in contrast to the lpxR mutant grown at 21°C, where the LpxR-dependent deacylation was hardly observed. The fact that YeLpxR is in latent stage at this growth temperature may suggest that, in the lpxR mutant background, the absence of the enzyme in the outer membrane, not the lipid A deacylation, acts as the regulatory signal underlying the reduced expressions of flhDC and rovA. Given experimental support to this hypothesis, the catalytically inactive mutants LpxR(N9A) and LpxR(S34A) restored the expressions of flhDC, ylpA, inv and rovA to wild-type levels. These results are in good agreement with the notion that membrane-intrinsinc β-barrel proteins, such as LpxR, may launch transmembrane signal transduction pathways upon sensing outer membrane perturbations [59], in our case, the absence of the protein itself. Therefore, it can be speculated that those systems sensing extracytoplasmatic stresses could underlie the regulatory connection between the absence of LpxR and the expression of Y. enterocolitica virulence factors. Giving indirect support to our speculation, it has been reported that lipid A deacylation induces σE-dependent responses in E. coli [60], the Cpx system senses changes in LPS O-polysaccharide [61]. Experiments are underway to test whether the activation status of the Cpx and/or σE systems is altered in the lpxR mutant background and whether any of these systems is responsible for the reduced expression of flhDC and rovA found in the mutant.
The LPS contains a molecular pattern recognized by the innate immune system thereby arousing several host defence responses. On one hand, CAMPs target this LPS pattern to bind to the bacterial surface, which is necessary for their microbicidal action. On the other hand, recognition of the LPS by the LPS receptor complex triggers the activation of host defence responses, chiefly the production of inflammatory markers. Not surprisingly, the modification of the LPS pattern is a virulence strategy of several pathogens to evade the innate immune system, and Y. enterocolitica is not an exception. Recently, we have demonstrated that the temperature-dependent lipid A modifications with aminoarabinose and palmitate help Y. enterocolitica to avoid the bactericidal action of CAMPs [15]. In this context, it was not totally unexpected to find out that the lpxR mutant was as susceptible as the wild type to polymyxin B, a model CAMP, since the mass spectrometry analysis indicated that the aforementioned lipid A modifications were not affected in the lpxR mutant background. Concerning the activation of inflammatory responses, several studies highlight the critical role of pYV-encoded Yops, chiefly YopP, to prevent the activation of inflammatory responses in a variety of cells, including macrophages. Nevertheless, Rebeil and co-workers [12] conclusively demonstrated that purified LPS from Y. enterocolitca grown at 37°C is less inflammatory than that purified from bacteria grown at 21°C. This is in agreement with the concept that underacylated LPSs are less inflammatory than hexa-acylated ones [52]. Therefore, it was plausible to speculate that the LpxR-dependent deacylation of LPS at 37°C was responsible for the reduced stimulatory potential of the LPS described by Rebeil and co-workers. To confirm this speculation we chose to challenge macrophages with alive bacteria instead of using purified LPS since there might be differences between the cellular recognition of purified LPS and the LPS expressed in the complex lipid environment of the bacterial outer membrane. To our initial surprise, we observed that the lpxR mutant elicited similar inflammatory response than the wild type when both strains were grown at 37°C. The fact that these responses were significantly lower than those elicited by bacteria grown at 21°C suggested that pYV-encoded factors were attenuating the inflammatory response. Therefore, we hypothesized that the arsenal of Yops injected to the cell were efficiently counteracting the activation of inflammatory responses evoked by the lpxR mutant LPS. In fact, our data demonstrated that the production and function of the pYV-encoded virulence factors were not affected in the lpxR mutant. Giving support to our hypothesis, the inflammatory response elicited by the lpxR mutant cured of the pYV virulence plasmid grown at 37°C was significantly higher than that induced by the virulence plasmid negative wild-type strain. Moreover, our findings suggest that, among all Yops, YopP plays a major role in counteracting the inflammation elicited by the lpxR mutant since the TNFα levels induced by the lpxR mutant cured of the pYV virulence plasmid grown at 37°C were not different than those triggered by YeO8-ΔyopP-ΔlpxR. On the whole, our results and those reported by Rebeil and co-workers [12] are consistent with a model in which the characteristic low inflammatory response associated to Y. enterocolitica infections might be the sum of the anti-inflammatory action exerted by YopP and the reduced activation of the LPS receptor complex due to the expression of a LpxR-dependent deacylated LPS. In this scenario, the latency of LpxR may facilitate a quick bacterial response upon entering the host to reduce the initial recognition of the pathogen by the LPS receptor complex. This will allow the pathogen to activate other host countermeasures, among others the pYV-encoded type III secretion system, which is a time consuming process.
Bacterial strains and plasmids used in this study are listed in Table 1. Unless otherwise indicated, Yersinia strains were grown in lysogeny broth (LB) medium at either 21°C or 37°C. When appropriate, antibiotics were added to the growth medium at the following concentrations: ampicillin (Amp), 100 µg/ml for Y. enterocolitica and 50 µg/ml for E. coli; kanamycin (Km), 100 µg/ml in agar plates for Y. enterocolitica, 50 µg/ml in agar plates for E. coli, and 20 µg/ml in broth; chloramphenicol (Cm), 20 µg/ml; trimethoprim (Tp), 100 µg/ml; tetracycline (Tet) 12.5 µg/ml; and streptomycin (Str), 100 µg/ml.
In silico analysis led to the identification of Y. enterocolitica 8081 homologue of lpxR (YE3039), yopP (YEP0083) and yopE (YEP0053) [accession number AM286415; [20]]. To obtain the lpxR, yopP, and yopE mutants two sets of primers (Table S1) were used for each gene to amplify two different fragments from each gene, LpxRUP and LpxRDOWN, YopPUP and YopPDOWN, YopEUP and YopEDOWN, respectively. Both fragments were BamHI-digested, purified, ligated, amplified as a single PCR fragment using a mixture of GoTaq Flexi polymerase (2.5 units/reaction; Promega) and Vent polymerse (2.5 units/reaction; New England Biolabs), gel purified and cloned into pGEMT-Easy (Promega) to obtain pGEMTΔlpxR, pGEMTΔyopP, and pGEMTΔyopE respectively. A kanamycin resistance cassette flanked by FRT recombination sites was obtained as a BamHI fragment from pGEMTFRTKm and it was cloned into BamHI-digested pGEMTΔlpxR and pGEMTΔyopP to generate pGEMTΔlpxRKm and pGEMTΔyopPKm respectively. ΔlpxR::Km, and ΔyopP::Km alleles were amplified using Vent polymerase (New England Biolabs) and cloned into SmaI-digested pKNG101 to obtain pKNGΔlpxRKm and pKNGΔyopPKm, respectively. ΔyopE allele was obtained by PvuII-digestion of pGEMTΔyopE, gel purified and cloned into SmaI-digested pKNG101 to obtain pKNGΔyopE. pKNG101 is a suicide vector that carries the defective pir-negative origin of replication of R6K, the RK2 origin of transfer, and an Str resistance marker [62]. It also carries the sacBR genes that mediate sucrose sensitivity as a positive selection marker for the excision of the vector after double crossover [62]. Plasmids were introduced into E. coli CC118-λpir from which they were mobilized into Y. enterocolitica 8081 by triparental conjugation using the helper strain E. coli HB101/pRK2013. Bacteria were diluted and aliquots spread on Yersinia selective agar medium plates (Oxoid) supplemented with Str. Bacteria from 5 individual colonies were pooled and allowed to grow in LB without any antibiotic overnight at RT. Bacterial cultures were serially diluted and aliquots spread in LB without NaCl containing 10% sucrose and plates were incubated at RT. The recombinants that survived 10% sucrose were checked for their antibiotic resistance. The appropriate replacement of the wild-type alleles by the mutant ones was confirmed by PCR and Southern blot (data not shown). In the case of YeO8-ΔlpxRKm and YeO8-ΔyopPKm mutants, the kanamycin cassette was excised by Flp-mediated recombination [63] using plasmid pFLP2Tp. This plasmid is a derivative from pFLP2 constructed by cloning a trimethoprim resistance cassette, obtained by SmaI digestion of p34S-Tp [64], into ScaI-digested pFLP2. The generated mutants were named YeO8-ΔlpxR and YeO8-ΔyopP, respectively. YeO8-ΔyopP-ΔlpxRKm and YeO8-ΔpmrF-ΔlpxRKm double mutants were obtained mobilizing the pKNGΔlpxRKm plasmid into YeO8-ΔyopP and YeO8-ΔpmrF, respectively. The replacement of the wild-type alleles by the mutant ones was done as described above and confirmed by PCR (data not shown).
To cure the pYV plasmid from YeO8-ΔlpxRKm, bacteria were grown at 37°C in Congo Red Magnesium oxalate agar plates [65]. Colony size and lack of uptake of Congo Red were used to detect loss of the virulence plasmid. This was further confirmed by testing the YadA-dependent autoagglutination ability [66].
A 443 bp DNA fragment containing the promoter region of lpxR was amplified by PCR using Vent polymerase (see Table S1 for primers used), EcoRI digested, gel purified and cloned into EcoRI-SmaI digested pGPL01Tp suicide vector [15]. This vector contains a promoterless firefly luciferase gene (lucFF) and a R6K origin of replication. A plasmid in which lucFF was under the control of the lpxR promoter was identified by restriction digestion analysis and named pGPL01TpYelpxR. This plasmid was introduced into E. coli DH5α-λpir from which it was mobilized into Y. enterocolitica by triparental conjugation using the helper strain E. coli HB101/pRK2013. Strains in which the suicide vectors were integrated into the genome by homologous recombination were selected. This was confirmed by PCR (data not shown).
To complement the lpxR mutant, a DNA fragment of 1.5 kb was PCR-amplified using TaKaRa polymerase (see Table S1 for primers used) gel purified, and cloned into pGEMT-Easy (Promega) to obtain pGEMTComlpxR. A fragment, containing the putative promoter and coding region of the deacylase, was obtained by PvuII digestion of pGEMTComlpxR, gel purified and cloned into the ScaI site of the medium copy plasmid pTM100 [40] to obtain pTMLpxR. For the construction of plasmid pTMLpxRFLAG, the lpxR coding region with its own promoter and a FLAG epitope sequence right before the stop codon was PCR amplified using Vent polymerase, primers LpxRtagging and LpxrFLAG (Table S1) and genomic DNA as template. The fragment was phosphorylated, gel purified and cloned into ScaI-digested pTM100. pTMLpxR and pTMLpxRFLAG were introduced into E. coli DH5α-λpir and then mobilized into Y. enterocolitica strains by triparental conjugation using the helper strain E. coli HB101/pRK2013.
Lipid As were extracted using an ammonium hydroxide/isobutyric acid method and subjected to negative ion matrix-assisted laser desorption ionization time-of-flight (MALDI-TOF) mass spectrometry analysis [14], [67]. Analyses were performed on a Bruker Autoflex II MALDI-TOF mass spectrometer (Bruker Daltonics, Incorporated) in negative reflective mode with delayed extraction. Each spectrum was an average of 300 shots. The ion-accelerating voltage was set at 20 kV. Dihydroxybenzoic acid (Sigma Chemical Co., St. Louis, MO) was used as a matrix. Further calibration for lipid A analysis was performed externally using lipid A extracted from E. coli strain MG1655 grown in LB at 37°C. Interpretation of the negative-ion spectra is based on earlier studies showing that ions with masses higher than 1000 gave signals proportional to the corresponding lipid A species present in the preparation [9], [12], [17], [68]. Important theoretical masses for the interpretation of peaks found in this study are: lipid IVA, 1405; C12, 182, C14, 210; C16∶1, 236.2; aminoarabinose (AraNH), 131.1; C16, 239.
Site-directed mutagenesis of the lpxR gene was performed by PCR [69]. Plasmid pTMLpxR, obtained with a minipreparation kit (Macherey-Nagel), was used as template and the desired mutations were introduced by the primer pairs described in Table S1. Amplifications were carried out in 50 µl reaction mixture using Vent DNA polymerase (New England BioLabs.). The PCR was started with initial 70 sec incubation at 95°C and then steps (95°C 50 sec, 60°C 75 sec and 72°C 6 min) were repeated 20 times followed by a 10 min extension time at 72°C. The obtained PCR products were gel purified, phosphorylated with T4 polynucleotide kinase, ligated, and digested with DpnI to break down any remaining template plasmid. The ligated PCR-product was transformed into E. coli C600. Plasmid DNA was isolated from transformants and the lpxR gene was completely sequenced to confirm the generated mutations and to ensure that no other changes were introduced. The name of each mutant construct includes the wild-type residue (single-letter amino acid designation) followed by the codon number and mutant residue (typically alanine).
For the construction of plasmids pTMLpxR(N9A)FLAG and pTMLpxR(S34A)FLAG, the lpxR alleles encoded into pTMLpxR(N9A) and pTMLpxR(S34A) were PCR amplified using Vent polymerase, and primers LpxRtagging and LpxrFLAG (Table S1). The fragments were phosphorylated, gel purified and cloned into ScaI-digested pTM100 [40]. Plasmids were introduced into E. coli DH5α-λpir and then mobilized into Y. enterocolitica strains by triparental conjugation using the helper strain E. coli HB101/pRK2013.
Overnight 5-ml cultures of Y. enterocolitica strains were diluted 1∶21 into 100 ml of LB in a 250-ml flask. Cultures were incubated with aeration at 21°C or 37°C until OD600 0.8. Bacteria were recovered by centrifugation (6500×g; 10 min, RT) and they were resuspended in 2 ml of 10 mM Tris/HCl (pH 7.4)-5 mM MgSO4 containing 2% Triton X-100 (v/v). Cells were broken by sonication (Branson digital sonifier; microtip 1/8″ diameter, amplitude 10%) for 15×1 min cycles, each cycle comprised 1 min sonication step separated by 1 min intervals. Unbroken cells were eliminated by centrifugation (2000×g, 20 min), and cell envelopes were recovered by ultracentrifugation (Beckman 70.1 Ti rotor; 45 000×g; 1 h, 4°C). The cell envelopes were resuspended in 500 µl of distilled water. The protein concentration was determined using the BCA Protein Assay Kit (Thermo Scientifc).
80 µg of proteins were separated on 4–12% SDS-PAGE, and semi-dry electrotransferred onto a nitrocellulose membrane using as transfer buffer SDS-PAGE-urea lysis buffer [a freshly prepared 1∶1 mix of 1× SDS running buffer (12 mM Tris, 96 mM glycine, 0.1% SDS] and urea lysis buffer (10 mM Na2HPO4, 1% β-mercaptoethanol, 1%SDS, 6 M urea)] [70]. Membrane was blocked with 4% skim milk in PBS. Membranes were stained using anti-Flag antibody (1∶2000; Sigma) following the instructions of the supplier.
A homology model of YeLpxR was constructed based on the crystal structure of StLpxR (PDB code 3FID; [21]. The YeLpxR sequence was used as bait to search Protein Data Bank with the Basic Local Alignment Search Tool (BLAST) at NCBI (http://blast.ncbi.nlm.nih.gov/). A pairwise sequence alignment was made using the program MALIGN [71] in the BODIL modeling environment [72], and a picture of the alignment was created using ESPript [73]. The essential water molecule and the zinc ion in the StLpxR crystal structure were also included in the YeLpxR model. A set of ten models was created with the program MODELLER [74], from which the model with the lowest value of the MODELLER objective function was analyzed and compared to the crystal structure of StLpxR by superimposing with the program VERTAA (Johnson & Lehtonen, 2004) in BODIL. Different rotamers for D10 and D31 were searched with the program Jackal (http://wiki.c2b2.columbia.edu/honiglab_public/index.php/Software:Jackal). D10 was changed to the same rotamer as in the crystal structure of StLpxR, while the rotamer used for D31 was the one with the lowest energy according to Jackal. SURFNET [75] was used to detect surface cavities, while PyMOL (Version 1.4, Schrödinger, LLC) was used for preparing pictures. For the SURFNET calculations, the minimum radius for gap spheres was set to 1.5 Å and the maximum radius was 4.0 Å.
For the docking studies, a Kdo2-lipid A, both with and without aminoarabinose, was modified from the coordinates for the LPS molecule in the crystal structure of FhuA [76]. The fatty acyl chains were removed from the Kdo2-lipid A molecule in order to reduce the number of rotatable bonds and make the docking more reliable. Aminoarabinose was added to the modified Kdo2-lipid A molecule with SYBYL (Version 8.0, Tripos Associates, Inc., St Louis, MO, USA), and the structure was minimized with the conjugate gradient method and Tripos force field. The modified Kdo2-lipid A, both with and without aminoarabinose, was docked to the YeLpxR model and the StLpxR crystal structure (PDB code 3FID) with GOLD via Discovery Studio (CSC IT Center for Science Ltd, Espoo, Finland), with default docking parameters and the receptor cavity defined to D10, Q16, T/S34, K67, and Y130.
The reporter strains were grown at 21°C or at 37°C on an orbital incubator shaker (180 r.p.m.) until OD540 1.6. The cultures were harvested (2500×g, 20 min, 24°C) and resuspended to an OD540 of 1.0 in PBS. A 100 µl aliquot of the bacterial suspension was mixed with 100 µl of luciferase assay reagent (1 mM D-luciferin [Synchem] in 100 mM citrate buffer pH 5). Luminescence was immediately measured with a Luminometer LB9507 (Berthold) and expressed as relative light units (RLU). All measurements were carried out in quintuplicate on at least three separate occasions.
Phenotypic assays for swimming motility were initiated by stabbing 2 µl of an overnight culture at the centre of agar plates containing 0.3% agar and 1% tryptone [25], [26]. Plates were analyzed after 24 h of incubation at RT and the diameters of the halos migrated by the strain from the inoculation point were compared. Experiments were run in quadruplicate in three independent occasions.
To measure flhDC expression, plasmid pRSFlhDC08 [26] encoding the transcriptional fusion flhDC::lucFF was integrated into the genomes of the strains by homologous recombination. This was confirmed by Southern blot (data not shown). Luminescence was determined as previously described.
β-galactosidase activity was determined as previously described with bacteria grown in 1% tryptone at RT [77]. Alkaline phosphatase activity was determined in permeabilized cells and the results are expressed in enzyme units per OD600 as previously described [78]. Experiments were run in duplicate in three independent occasions.
Bacteria were grown at 21°C or at 37°C in 5 ml of LB medium on an orbital incubator shaker (180 r.p.m.) until an OD600 of 0.3. 0.5 ml of ice-cold solution EtOH/phenol [19∶1 v/v (pH 4.3)] were added to the culture and the mixture was incubated on ice for 30 min to prevent RNA degradation. Total RNA was extracted using a commercial NucleoSpin RNA II kit as recommended by the manufacturer (Macherey-Nagel).
cDNA was obtained by retrotranscription of 2 µg of total RNA using a commercial M-MLV Reverse Transcriptase (Sigma), and random primers mixture (SABiosciences, Quiagen). 50 ng of cDNA were used as a template in a 25-µl reaction. RT-PCR analyses were performed with a Smart Cycler real-time PCR instrument (Cepheid, Sunnyvale, CA) and using a KapaSYBR Fast qPCR Kit as recommended by the manufacturer (Cultek). The thermocycling protocol was as follows; 95°C for 3 min for hot-start polymerase activation, followed by 45 cycles of 95°C for 15 s, and 60°C for 30 s. SYBR green dye fluorescence was measured at 521 nm.
cDNAs were obtained from three independent extractions of mRNA and each one amplified by RT-qPCR in two independent occasions. Relative quantities of lpxR mRNAs were obtained using the comparative threshold cycle (ΔΔCT) method by normalizing to rpoB and tonB genes (Table S1).
Overnight cultures of Y. enterocolitica strains were diluted 1∶50 into 25 ml of TSB supplemented with 20 mM MgCl2 and 20 mM sodium oxalate in a 100-ml flask. Cultures were incubated with aeration at 21°C for 2.5 h, and then transferred at 37°C for 3 h. The optical density at 540 nm of the culture was measured and the bacterial cells were collected by centrifugation at 1500×g for 30 min. Ammonium sulphate (final concentration 47.5% w/v) was used to precipitate proteins from 20 ml of the supernatant. After overnight incubation at 4°C, proteins were collected by centrifugation (3000×g, 30 min, 4°C) and washed twice with 1.5 ml of water. Dried protein pellets were resuspended in 50 to 80 µl of sample buffer and normalized according to the cell count. Samples were analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) using 12% polyacrylamide gels and proteins visualized by Coomassie brilliant blue staining.
Control experiments revealed that the secretion of Yops was not affected in yopE and yopP mutants except that each mutant did not produce either YopE or YopP, respectively (data not shown).
Bacteria were grown overnight in 2 ml RPMI 1640 medium lacking phenol red at 37°C without shaking. The OD540 of the culture was measured and CFUs were determined by plating serial dilutions. Bacteria from 1-ml aliquot were recovered by centrifugation (16 000×g, 10 min, 4°C) and resuspended in 200 µl of SDS-sample buffer. Samples were incubated for 4 h at 37°C and kept frozen at −20°C. Samples were analyzed by SDS-PAGE using 10% polyacrylamide gels and proteins visualized by Coomassie brilliant blue staining. Samples were normalized according to the cell count and they were not boiled before loading the gel.
Overnight cultures of Y. enterocolitica strains grown at 37°C were diluted 1∶10 into 5 ml of LB and grown with aeration at 37°C for 2.5 h. bacteria were pelleted, washed once with PBS and resuspended to an OD540 of 0.3 in PBS.
12 mm circular coverslips in 24-well tissue culture plates were coated overnight at 4°C with 10 µg/ml human collagen type I (Sigma) in PBS (final volume 100 µl). Coverslips were washed three times with TBS and later they were blocked for 1 h at 4°C with 2% BSA in TBS. Finally, coverslips were washed three times and were incubated at 37°C with 100 µl of the bacterial suspension. After 1 h incubation, the coverslips were washed three times with PBS and then bacteria fixed with 3.7% paraformaldehyde (PFA) in PBS pH 7.4 for 20 min at room temperature. PFA fixed cells were incubated with PBS containing 0.1% saponin, 10% horse serum and Hoechst 33342 (1∶25000) for 30 min in a wet dark chamber. Finally, coverslips were washed twice in 0.1% saponin in PBS, once in PBS and once in H2O, mounted on Aqua Poly/Mount (Polysciences) and analysed with a Leica CTR6000 fluorescence microscope. Bacteria were counted in images from three randomly selected fields of view obtained at a magnification of ×100 taken with a Leica DFC350FX camera. Wild-type adhesion was set to 100%.
Carcinomic human alveolar basal epithelial cells (A549, ATTC CCL-185) were maintained in RPMI 1640 tissue culture medium supplemented with 1% HEPES, 10% heat inactivated foetal calf serum (FCS) and antibiotics (penicillin and streptomycin) in 25 cm2 tissue culture flasks at 37°C in a humidified 5% CO2 atmosphere as previously described [79]. For infections, A549 cells were seeded on 12 mm circular coverslips in 24-well tissue culture plates to 70% confluence. Cells were serum starved 16 h before infection. Overnight cultures of Y. enterocolitica strains grown at 21°C were diluted 1∶10 into 5 ml of LB and grown with aeration at 21°C for 1.5 h and then 1 h at 37°C. Bacteria were pelleted, washed once with PBS and resuspended to an OD600 = 1 (approximately 109 CFU/ml) in PBS. Cells were infected with this suspension to get a multiplicity of infection of 25∶1. After 1 h incubation, the coverslips were washed three times with PBS and then cells fixed with 3.7% PFA in PBS pH 7.4 for 20 min at room temperature. PFA fixed cells were incubated with PBS containing 0.1% saponin, 10% horse serum, Hoechst 33342 (1∶2500), and OregonGreen 514-phalloidin (1∶100) (Invitrogen) for 30 min in a wet dark chamber. Finally, coverslips were washed twice in 0.1% saponin in PBS, once in PBS and once in H2O, mounted on Aqua Poly/Mount (Polysciences) and analysed with a Leica CTR6000 fluorescence microscope. Images were taken with a Leica DFC350FX camera.
YopE translocation into A549 cells was done as previously described [38]. Briefly, A549 cells were seeded in 12-well tissue culture plates to 80% confluence. Cells were serum starved 16 h before infection. Overnight cultures of Y. enterocolitica strains grown at 21°C were diluted 1∶10 into 5 ml of LB and grown with aeration at 21°C for 1.5 h and then 1 h at 37°C. Bacteria were pelleted, washed once with PBS and resuspended to an OD600 = 1 (approximately 109 CFU/ml) in PBS. Cells were infected with this suspension to get a multiplicity of infection of 25∶1. To synchronize infection, plates were centrifuged at 200×g during 5 min. After 1 h infection, cells were washed twice with PBS and resuspended in 400 µl of PBS with the help of a rubber policeman. Cell suspensions were transferred to a 1.5 ml microcentrifuge tube and cells pelleted (16 000×g; 12 sec). Supernatant was carefully removed and cells were resuspended in 100 µl of 1% digitonin (w/v) in PBS supplemented with a cocktail of protease inhibitors (Halt protease inhibitor single-use cocktail EDTA-free; Thermo). After 2-min incubation at RT, samples were centrifuged (16 000×g; 10 min, 4°C). 80 µl of the supernatant, containing cytosolic proteins, were collected to whom 20 µl of 5× SDS sample buffer were added. The pellet, containing intact bacteria and cell membranes, was resuspended in 100 µl 1× SDS sample buffer. Aliquots corresponding to approximately 6×104 infected A549 cells were analysed by SDS-polyacrylamide gel electrophoresis and Western blotting using rabbit polyclonal antiserum raised against YopE (1∶2000 dilution).
Strains were grown aerobically for 16 h at RT, pelleted and resuspended to an OD540 of 0.3 in PBS. Bacteria suspensions were added to subconfluent HeLa cells at a multiplicity of infection of ∼25∶1. After a 30 min infection, monolayers were washed twice with PBS and then incubated for an additional 90 min in medium containing gentamicin (100 µg/ml) to kill extracellular bacteria. This treatment was long enough to kill all extracellular bacteria. After this period, cells were washed three times with PBS and lysed with 0.5% saponin in PBS and bacteria were plated. Experiments were carried out in triplicate on three independent occasions. Invasion is expressed as CFUs per monolayer.
Bacteria were grown either at 21°C or 37°C in 5 ml LB in a 15-ml Falcon tube with shaking (180 rpm), and harvested (2500×g, 20 min, 24°C) in the exponential growth phase (OD540 0.8). Bacteria were washed once with PBS and a suspension containing approximately 1×105 CFU/ml was prepared in 10 mM PBS (pH 6.5), 1% Tryptone Soya Broth (TSB; Oxoid), and 100 mM NaCl. Aliquots (5 µl) of this suspension were mixed in 1.5 ml microcentrifuge tubes with various concentrations of polymyxin B (Sigma). In all cases the final volume was 30 µl. After 1 h incubation at the bacterial growth temperature, the contents of the tubes were plated on LB agar. Colony counts were determined and results were expressed as percentages of the colony count of bacteria not exposed to antibacterial agents. All experiments were done with duplicate samples on at least four independent occasions.
Murine macrophages RAW264.7 (ATCC, TIB71) were grown on DMEM tissue culture medium supplemented with 10% heat-inactivated foetal calf serum (FCS) and Hepes 10 mM at 37°C in an humidified 5% CO2 atmosphere. For bacterial infection, cells were seeded in 24-well tissue culture plates 15 h before the experiment at a density of 7×105 cells per well. Overnight cultures of Y. enterocolitica strains grown at 21°C were diluted 1∶10 into 5 ml of LB and grown with aeration at 37°C or 21°C for 3 h. Bacteria were pelleted, washed once with PBS and resuspended to an OD600 = 1 (approximately 109 CFU/ml) in PBS. Cells were infected with this suspension to get a multiplicity of infection of 25∶1. To synchronize infection, plates were centrifuged at 200×g during 5 min. After a 30 min infection, cells were washed twice with PBS and then incubated for an additional 180 min in medium containing gentamicin (100 µg/ml). Supernatants were removed from the wells, cell debris removed by centrifugation, and samples were frozen at −80°C. TNFα present in supernatants of culture cells was determined by ELISA (Bender MedSystems) with a sensitivity <4 pg/ml.
The results were analyzed by the one-sample t test using GraphPad Prism software (GraphPad Software Inc.). Results are given as means ± SD. A P value of <0.05 was considered to be statistically significant.
|
10.1371/journal.ppat.1006030 | T Cell Receptor Vβ Staining Identifies the Malignant Clone in Adult T cell Leukemia and Reveals Killing of Leukemia Cells by Autologous CD8+ T cells | There is growing evidence that CD8+ cytotoxic T lymphocyte (CTL) responses can contribute to long-term remission of many malignancies. The etiological agent of adult T-cell leukemia/lymphoma (ATL), human T lymphotropic virus type-1 (HTLV-1), contains highly immunogenic CTL epitopes, but ATL patients typically have low frequencies of cytokine-producing HTLV-1-specific CD8+ cells in the circulation. It remains unclear whether patients with ATL possess CTLs that can kill the malignant HTLV-1 infected clone. Here we used flow cytometric staining of TCRVβ and cell adhesion molecule-1 (CADM1) to identify monoclonal populations of HTLV-1-infected T cells in the peripheral blood of patients with ATL. Thus, we quantified the rate of CD8+-mediated killing of the putative malignant clone in ex vivo blood samples. We observed that CD8+ cells from ATL patients were unable to lyse autologous ATL clones when tested directly ex vivo. However, short in vitro culture restored the ability of CD8+ cells to kill ex vivo ATL clones in some donors. The capacity of CD8+ cells to lyse HTLV-1 infected cells which expressed the viral sense strand gene products was significantly enhanced after in vitro culture, and donors with an ATL clone that expressed the HTLV-1 Tax gene were most likely to make a detectable lytic CD8+ response to the ATL cells. We conclude that some patients with ATL possess functional tumour-specific CTLs which could be exploited to contribute to control of the disease.
| Human T lymphotropic virus-1 infects T cells, causing them to multiply. In some people, cellular replication is unchecked, resulting in an aggressive blood cancer called adult T-cell leukemia/lymphoma. The virus proteins are efficiently recognised as ‘foreign’ by the immune system in most infected individuals. People with cancer have weak immune responses to certain viral proteins, however it was not known whether the immune system can attack the malignant cells in this disease. In this paper, we developed a method which allows us to directly monitor malignant cells, and used it to test whether malignant and non-malignant infected cells are killed by immune cells from people with the cancer. We found that some people had immune cells which could kill the cancer cells. These observations are both new and important because they raise the possibility of boosting the immune response to malignant cells as a novel therapeutic strategy for this aggressive and hard-to-treat disease.
| Adult T cell leukemia/lymphoma is a mature T cell malignancy caused by the retrovirus human T lymphotropic virus-1 (HTLV-1). Four clinical subtypes exist: acute, lymphoma, chronic and smouldering, which range from highly aggressive to indolent in their clinical course [1,2]. Advances in chemotherapy protocols have contributed only a modest increase in overall survival of aggressive subtypes, and few patients receive potentially curative allogeneic hematopoietic stem cell transplantation (HSCT)[3]. Antiviral drugs (zidovudine and interferon alpha, AZT/IFN)[4–7] and molecular targeted therapy (anti-CCR4, Mogamulizumab)[8–10] have shown promising results, especially in chronic ATL, but their efficacy in the lymphoma and acute subtypes is limited. There is an urgent need for new therapies and strategies to consolidate existing treatments.
HTLV-1 establishes persistent infection by integration of the provirus into the genomic DNA of T lymphocytes, and propagates in the host by both clonal proliferation and cell-to-cell transmission[11,12]. Expression of structural genes on the sense strand of the 9kb genome is induced by the viral transcriptional transactivator protein Tax, triggering production of viral particles, cellular activation and proliferation. The antisense strand encodes HTLV-1 b-zip protein (HBZ), which opposes many of the actions of Tax[13]. HTLV-1+ individuals carry thousands of long-lived infected CD4+ clones in their peripheral blood, each of which has arisen from a single infection event[12,14]. Malignant cells in ATL are HTLV-1-infected clones: in 91% of ATL cases a single dominant proviral integration site makes up over 35% of the proviral load[15], circulating alongside subdominant populations of polyclonal infected and uninfected T cells. Although the genomic integration site influences clonal proliferation and proviral gene expression[16], it does not appear to explain clonal dominance in most cases of ATL[15]. Spontaneous mutations in the T cell receptor (TCR)/NF-kB[17], CCR4[18], p53[19] and, Notch-1[20] signalling pathways are frequently observed in malignant clones.
Several lines of evidence indicate that the outcome of HTLV-1 infection is determined by the equilibrium set between proliferation of infected cells and the activity of abundant, chronically activated, HTLV-1-specific cytotoxic T lymphocytes [21,22]. Major histocompatibility complex (MHC) class 1 alleles HLA-A*0201 and C*08 are associated with a low proviral load[23] in southern Japan. Tax protein is highly immunodominant in the HTLV-1-specific CD8+ response, and tax is silenced or deleted in the dominant clone in over 50% of patients with ATL, implying the presence of strong CTL selection pressure. Paradoxically, ectopic expression of Tax can be oncogenic in vivo[24]. The region of the viral genome which encodes HBZ is highly conserved in ATL[25], suggesting HBZ also has a role in oncogenesis. The ability of an individual to present peptides from HBZ to CD8+ cells is associated with a low proviral load[26], however, HBZ evades immune detection by means of low-level protein expression and weak immunogenicity[26]. In addition, biological actions are exerted by untranslated HBZ mRNA[25,27].
ATL patients are commonly immunosuppressed, and frequently present with opportunistic infections. Previous studies on samples from ATL patients have reported that the frequency and diversity of HTLV-1-specific CD8+ T cells is significantly lower in ATL patients than in non-malignant HTLV-1 infection[28,29]. In addition to the silencing of Tax expression, several mechanisms by which ATL cells might escape CTL have been proposed. The malignant clone in 5%-6% of ATL patients carries mutations in HLA-A or -B genes, and the MHC class 1-encoding region in ATL is frequently subject to hypermethylation and copy-number variation[17]. ATL cells frequently express the regulatory T-cell-associated transcription factor FoxP3[30] and the coinhibitory ligand PD-L1[31], but it remains unclear whether primary ATL clones directly suppress CD8+ responses. Indeed, the susceptibility of primary ATL clones to CD8+-mediated lysis is not known, though rare occurrences of spontaneous disease remission[32], and successful allogeneic HSCT[33,34] have been reported to involve induction and maintenance of HTLV-1-specific CTLs[35].
Measuring the rate at which ATL clones are killed by CD8+ cells requires a reliable method to distinguish ATL clones from both non-malignant HTLV-1 infected cells and uninfected T cells. We recently published that CADM1 expression identifies 60–70% of infected cells in HTLV-1 carriers [36]. ATL patients have high frequencies of CADM1+[37],CCR4+[38], CD25+[1] and CD7−[39] cells in their peripheral blood. These cells often express FoxP3[40] and low levels of CD3 epsilon [41]. However, this combination of markers is also expressed by a subset of CD4+ T cells in uninfected donors [42] and asymptomatic HTLV-1 carriers (ACs), particularly those with a high proviral load [39,43], thus may not be used to directly identify the ATL clone.
Here, we used TCRVβ subunit staining, immunophenotyping and high-throughput sequencing to identify clonally expanded populations in a well-characterised cohort of ATL patients. We show that in some individuals with ATL, the malignant clone is susceptible to lysis by cultured autologous CD8+ cells. Autologous CD8+ cells from ATL patients preferentially killed targets that expressed the viral sense strand: both Tax+ATL clones and Tax+non-malignant cells were killed. In all donors, cells which did not express Tax escaped killing by CD8+ cells.
Peripheral blood mononuclear cells from ATL patients, asymptomatic HTLV-1 carriers and patients with HTLV-1 associated myelopathy/tropical spastic paraparesis (HAM/TSP) were stained with a panel of antibodies specific for 24 TCRVβ subunits (S1 and S2 Tables) and CADM1.
The frequency distributions of TCRVβ subunits in CADM1+ (which typically carry one proviral copy per cell[36]) and CADM1− (low proviral load) T cells (both CD4+ and CD8+) were ascertained by dividing live CD3+ cells into 50 possible groups on the basis of TCRVβ staining (see Materials and Methods). Linker-mediated PCR (LM-PCR) followed by high-throughput sequencing (HTS) were performed to corroborate the observed frequency distributions (Fig 1A).
We used an oligoclonality index (OCI, Gini index)[12] to compare the frequency distribution of TCRVβ subunits (OCI-flow) with the frequency distribution of unique proviral integration sites (UIS, OCI-UIS). The frequency distribution of TCRVβ subunits in ACs and patients with HAM/TSP resembled that in healthy donors [44] (S1A and S2 Figs), with no significant difference in the OCI-flow of infected CADM1+CD3+ cells and the OCI-flow of predominantly uninfected CADM1−CD3+ cells (Fig 1B). By contrast, the OCI-flow of CADM1+CD3+ cells in ATL patients was significantly higher than that of CADM1−CD3+ cells from the same donor, and CADM1+/−CD3+ cells from donors without malignancy (Fig 1B). These results indicate that an OCI-flow>0.7 is associated with ATL (see below). In ATL patients, the OCI-flow for CADM1+CD3+ T cells measured by flow cytometry was significantly correlated with the OCI-UIS measured by HTS (Fig 1C). In addition, the absolute frequency of the most abundant UIS detected by HTS was significantly correlated with the frequency of the most abundant population of T cells which shared a single Vβ subunit (Fig 1C). We therefore refer henceforth to the dominant TCRVβ-expressing population of CD4+ cells, in individuals with an OCI-flow (CADM1+CD3+) > 0.7, as the ‘ATL clone’.
We detected putatively malignant expansions in patients with chronic (n = 12) or acute (n = 6) leukemia (Fig 1B, S1B and S2 Figs); in 16 cases by direct identification of the TCRVβ and two cases in which the TCRVβ subunit was not represented in the TCRVβ antibody panel. Each case had a population of T cells which shared a Vβ subunit comprising >35% of CADM1+ cells[15], and an OCI-flow of CADM1+CD3+ cells > 0.7. There was no evidence of preferential transformation of cells expressing particular TCRVβ subunits (S1 Table). Two out of five lymphoma patients also had CADM1+CD3+ PBMC with an OCI-flow > 0.7. Patients with leukemic type ATL who had an OCI-flow (CADM1+CD3+) <0.7 were in remission, and did not have a dominant proviral integration site (>35% of the PVL) detectable by HTS.
Direct flow-cytometric identification of clonal HTLV-1-infected populations in ATL permitted detailed assessment of the sensitivity and specificity of other established immunophenotypic markers of ATL. Using multicolour flow cytometry we evaluated co-expression of phenotypic ATL markers (CD3, CD25, CD7, CCR4 and CADM1) on cells which carried the respective dominant TCRVβ (designated TCRVβX+) with those which did not (TCRVβX−). As described in the literature, CD3 epsilon was significantly downregulated on expanded clones in ATL (Fig 2A and 2B), compared with cells from the same individual which expressed other TCRVβ subunits. CCR4 was expressed by a median of 98% cells within the malignant clone; CADM1 by 93%; and CD7 was downregulated on 96% (Fig 2B). CD25 had the poorest sensitivity of all the markers: a median of 66% of malignant cells were CD25+.
We tested the ability of CD7 and CD25 to discriminate between malignant and non-malignant infected cells by comparing the frequency of expression on CCR4+CADM1+CD4+cells from individuals with and without malignancy. Although expression of CD7 was significantly downregulated on infected CCR4+CADM1+ cells versus other CD4+ T cells in all HTLV-1-infected subjects, CD7 expression was lowest on CCR4+CADM1+ cells from ATL patients (S3 Fig). In contrast, the frequency of CD25 expression on CCR4+CADM1+ cells did not differ between the three disease states: ATL, HAM/TSP and AC (S3 Fig). After TCRVβ positivity, CD7 downregulation was the most specific marker of ATL clones. Thus, the combination of markers of clonal expansion (TCRVβ+ and CD7low) and infection (CADM1+) allows sensitive detection and accurate quantification of expanded ATL clones.
Within our cohort of 24 age-matched HTLV-1-infected individuals without malignancy (in whom the PVL ranged from undetectable to 79 copies per 100 PBMC, S1 Table), the OCI-flow of CADM1+CD3+ cells did not exceed 0.7. We plotted receiver operator curves (ROC) to evaluate the sensitivity and specificity by which the OCI-flow of CADM1+CD3+ cells could identify individuals with clinically evident ATL (Fig 3), compared with the common diagnostic investigations: enumerating CD7−CD4+ cells and CD25+CD4+ cells. Five ATL patients within the original cohort who were in clinical remission were excluded from this analysis on the basis of clinical observations (not on the basis of oligoclonality). Area under the curve (AUC) analysis rated the diagnostic power of the OCI-flow (CADM1+CD3+)and CD7−CD4+ frequency as ‘excellent’ (AUC 0.9–1), and CD25+CD4+ frequency as ‘good’ (AUC 0.8–0.9), and both tests had significantly higher diagnostic power than the frequency of CD25+CD4+ T cells (Fig 3, p = 0.001, OCI-flow (CADM1+CD3+) vs. CD25; p = 0.03, CD7 vs. CD25, one-tailed test [45]).
In 16 individuals with a known dominant TCRVβ, all T cells (including ATL clones) expressed MHC class 1 at a similar intensity (Fig 4A and S4 Fig). CADM1 expression was significantly higher on ATL clones than on non-malignant infected cells within the same individual, or CADM1+ cells from ACs (Fig 4B and S4 Fig). In polyclonal infected populations (CD4+CADM1+ cells in ACs, CADM1+VβX−cells in ATL patients), a median of 11–15% of CADM1+ cells expressed Tax after overnight culture ([36]; Fig 4C and S4 Fig). By contrast, ATL clones fell into two distinct groups: those in which <5% of cells expressed Tax (TaxlowATL) and those in which >5% expressed Tax (TaxhighATL). FoxP3 and PD-L1 were highly expressed in some cases, by both Taxhigh and Taxlow ATL clones (S5 Fig).
We tested the ability of autologous CD8+ cells to kill malignant clones using an ex vivo cell survival assay [46]. In order to mimic in vivo CD8+ cell:target cell frequencies, we incubated CD4+ PBMCs from ATL patients for 18h with a range of ratios of autologous CD8+ T cells and quantified the absolute number of surviving cells in the following populations: CADM1−CD4+ cells (which have a low proviral load [36]), malignant HTLV-1 infected CADM1+VβX+CD4+ cells, and non-malignant CADM1+VβX−CD4+ cells (Fig 5), which typically carry a single proviral copy per cell. This strategy permitted us to estimate the efficacy by which each subset was targeted by CD8+ cells in vivo, in the presence of other potential CTL targets. Because previous reports indicated that ex vivo CD8+ cells from ATL patients had negligible lytic function [28], purified CD8+ cells were expanded in culture for 2 weeks, both to increase the effector: target ratio, and to allow potential reactivation of lytic function.
At the effector: target ratios tested, no significant lysis of the ATL clone by autologous ex vivo CD8+ T cells was detected (Fig 6A). When compared with ACs, CD8+ from ATL patients also had a markedly reduced ability to lyse non-malignant Tax-expressing CADM1+CD4+ cells (Fig 6). Tax−CADM1+ and CADM1− CD4+ cells were not killed by ex-vivo CD8+ cells in either cohort. By contrast, after expansion in vitro, cultured CD8+ cells from 3 of 9 donors with ATL killed a proportion of their respective ATL clone (Fig 6B). Addition of 20nM concanamycin A blocked killing of ATL cells, indicating that the observed effect is perforin-dependent (S6 Fig), as previously reported [47].
We observed that the ATL clone was not completely eliminated at any CD4+:CD8+ ratio, even supraphysiological ratios (Fig 6). All donors (3/3) whose CD8+ cells regained the ability to lyse the malignant cells had an ATL clone which strongly expressed the proviral sense strand genes, as detected by intracellular expression of Tax protein (Taxhigh) (Fig 7A). In contrast, the malignant clones of all other donors in the cohort (6/6) were Taxlow. Within the ATL clone, we observed a strong preferential lysis of Tax-expressing malignant cells; only one donor lysed Tax-negative malignant cells (Fig 7A). Between 20–60% of malignant Tax expressing cells were cleared in each donor (Fig 7A). To quantify the preferential CD8+ targeting of cells that express the viral plus strand, we calculated for each donor the rate at which Tax-expressing and non-expressing cells were killed after in vitro culture (Fig 7B). Cultured CD8+ cells killed Tax-expressing ATL clones at a higher rate than ex vivo CD8+ cells in 3 of 3 cases. In addition, cultured CD8+ cells from patients with Taxlow clones also had enhanced ability to kill non-malignant HTLV-1 infected cells which expressed the viral plus strand (Fig 7B).
An array of novel anti-cancer immune therapies are currently in clinical trials, which potentiate existing immune responses, and induce tumour-specific immunity by vaccination, or infusion of engineered tumour-specific T cells. Might these approaches be effective in ATL?
We demonstrate that abnormal clonal expansions of HTLV-1-infected T cells are readily detectable in individuals with ATL by TCRVβ flow cytometry, which is faster, cheaper and less labour-intensive than the current gold-standard technique of high-throughput sequencing of proviral integration sites. In this cohort, an oligoclonality index of > 0.7 within CADM1+CD3+cells reliably identified individuals who had a dominant ATL clone as validated by high-throughput sequencing. Whilst CD25 expression is frequently high in ATL, we show that in most individuals, ~40% of cells in the ATL clone are CD25 negative. Over 94% of ATL clones were CCR4+CADM1+CD7−; the exceptions were one CD7dim ATL clone and one CADM1− clone. Analysis of TCRVβ expression within the heavily infected CADM1+CD3+ population allows direct flow cytometric analysis of the clonal structure of HTLV-1 infected cells, and can distinguish ATL patients from age-matched HTLV-1 carriers with high specificity and sensitivity. Whilst we did not have sufficient cases to independently test the diagnostic power of OCI-flow of CADM1+CD3+ cells in an unrelated cohort of cases and controls, our observations indicate that this measure could be useful in the diagnosis of ATL: particularly in detecting the presence of monoclonal/oligoclonal populations of HTLV-1 infected cells.
We exploited this technique to measure the rate at which ATL clones are lysed by ex vivo, autologous CTLs. Ex vivo CD8+ cells were unable to kill autologous malignant ATL cells, even at supraphysiological E:T ratios. In addition, CD8+ mediated killing of non-malignant cells which express viral proteins was less efficient in ATL patients than in asymptomatic carriers. In certain patients, in vitro culture of CD8+ cells revealed a population of CD8+ cells which could kill the ATL clone. This ability was associated with expression of the viral plus-strand genes by the ATL clone: TaxhighATL clones and Tax expressing non-malignant infected cells were preferentially targeted by cultured autologous CD8+ cells.
In most subjects in the present study, Tax was expressed on <5% of cells within the ATL clone after overnight in vitro culture, and Tax-negative cells consistently escaped CD8+-mediated killing. We observed no defect in MHC class 1 expression by ATL clones, and the level of expression of CADM1 by ATL clones was significantly higher than that in non-malignant HTLV-1-infected cells. CADM1 expression on the target cell enhances its susceptibility to CTL killing [36,48]; thus CADM1 could contribute to CD8+ lysis of ATL clones. The level of expression of CADM1 did not significantly differ between Taxhigh and Taxlow ATL clones: so while CADM1 is likely to facilitate killing of ATL clones which present epitopes which are recognised by CD8+ cells, CADM1 expression alone does not appear to expose the ATL clone to lysis by CD8+ cells. Likewise, the expression levels of PD-L1 and FoxP3 did not differ between Taxhigh and Taxlow clones in our cohort, so we could not make any inferences on the role of FoxP3/PD-L1 in the escape of ATL cells in this study.
ATL clones have the potential to present a range of non-self antigens to CTL: for example, the consistently expressed HTLV-1 antigen HBZ[49], or neoepitopes generated by the frequent somatic mutations observed in ATL: a recent study detected 6404 somatic mutations in 81 ATL cases by exome sequencing [17]. Apart from the frequent loss of expression of the dominant CTL target antigen Tax by ATL clones, there was no evidence that these clones evaded the immune response by downregulation of MHC class 1; nevertheless, the bulk of ATL cells escaped CTL lysis in most individuals. We conclude that the CTL response to antigens presented by ATL clones is insufficient or suppressed in established disease.
Although the CD8+ response in patients with ATL appears insufficient to maintain control of ATL cell expansion in vivo, the capacity of autologous CD8+ cells to lyse the malignant clone that we report here indicates an opportunity for therapeutic intervention by boosting the CD8+ response, particularly in patients where the ATL clone expressed the viral plus strand. Immunisation strategies have focused on Tax, and more recently HBZ. Because Tax expression is intermittent or low in vivo, and frequently deleted in ATL clones, and HBZ is only weakly immunogenic [50], other HTLV-1 antigens or neoantigens may be more effective CD8+ epitopes. Whilst we observed CD8+ killing of Tax expressing cells within ATL clones, boosting Tax-specific CD8+ responses alone is likely to strongly select for deletion of Tax in malignant clones: clearly a CTL response which targets multiple antigens would reduce the likelihood of immune escape. While certain somatic mutations are observed in a high proportion of primary ATL clones[17], the combination of neoepitopes and the individual’s HLA type is unique for most individuals with ATL. Immunisation with epitopes from autologous ATL clones could elicit a broader cellular immune response to the malignancy in comparison with immunisation with HTLV-1 epitopes alone. Finally, the ability to maintain long-term populations of effector cells will be a critical factor determining the efficacy of the ATL-specific CD8+ response.
Donors attended the National Centre for Human Retrovirology (Imperial College Healthcare NHS Trust, St Mary's Hospital, London). Written informed consent was obtained and research was conducted under the governance of the Communicable Diseases Research Group Tissue Bank, approved by the UK National Research Ethics Service (09/H0606/106, 15/SC/0089).
All ATL subtypes were included (S1 Table). PBMC were isolated from whole blood by density-gradient centrifugation using histopaque-1077 (Sigma-Aldrich, Poole) from EDTA-anticoagulated blood. Isolated PBMCs were washed twice in PBS then cryopreserved in FCS (Life technologies, Paisley) with 10% dimethysulfoxide (Sigma-Aldrich).
Genomic DNA was extracted using a DNeasy kit (Qiagen, Manchester), according to the manufacturer’s instructions, and proviral load was estimated as described in Manivannan et al, 2016 [36]. Genomic DNA (20 ng, 6.7 ng or 2.2 ng in 4 μl H2O) was subjected to thermal cycling in the presence of FastSYBR (Life Technologies) master mix and the following primer pairs: SK43/SK44- 5'CGGATACCCAGTCTACGTGT3' /5'GAGCCGATAACGCGTCCATCG3' (tax gene) or GAPDHF/GAPDHR- 5’AACAGCGACACCCATCCTC3’/5’ CATACCAGGAAATGAGCTTGACAA3’ (gapdh gene). DNA amplification was monitored in real time with a QuantStudio7 thermal cycler (Life technologies). DNA from a naturally-infected primary T cell clone which contained a single-copy of tax and two copies of gapdh as used as a standard. The proportion of PBMC which carry the provirus was estimated as follows: (copies of tax)/(2*copy number of gapdh)*100. Where > 1 copy of Tax is detected per 2 copies of GAPDH, the value exceeds 100%. Linker-mediated (LM)-PCR, high-throughput sequencing, data extraction and analysis of viral integration sites were carried out as described in Gillet et al[12]. Random fragments of genomic DNA (1 μg) generated by sonication were ligated to a partially double-stranded DNA adaptor. Nested PCR (two rounds) was used to amplify the region between the HTLV-1 LTR and the adaptor. Amplicons generated from adaptors with unique 6bp barcodes were combined into libraries; following which, sequence data from paired-end 50 bp reads and a 6 bp index (barcode) read were acquired on an Illumina HiSeq/MiSeq platform. Paired reads were then aligned to a human genome reference (Hg18). The number of individual cells which were sequenced within a given HTLV-1 infected clone were estimated by quantifying the number of distinct genomic shear sites generated by sonication (read2) for each paired unique integration site (junction between the provirus and human genome- read 1), and correcting to a calibration curve. The absolute abundance of unique integration sites per 100 PBMC was estimated by combining the proviral load and relative abundance of each clone.
The oligoclonality index (OCI) was used as a metric to compare the clone frequency distribution between samples. This was based directly on the Gini index [51], which calculates the relative inequality within a given distribution. The OCI was computed using the reldist package (http://CRAN.R-project.org/package=reldist) in R (http://www.R-project.org/). Values range between 0 and 1, with 0 indicating that all clones make up an equal proportion of the load, and 1 indicating that a single clone dominates completely [12].
Flow cytometric staining was performed as previously described [50] using panels of antibodies and stains outlined in S2 Table. Cells (3x 105-2x106) were stained for 5 min with 1 μl/ml fixable Live/Dead blue viability stain (Life technologies). After incubation cells were washed once with FACS buffer (PBS containing 7% normal goat serum). Surface molecules were stained for 20 min at room temperature (RT) with the antibodies listed in S2 Table. In order to quantify the frequency of T cells utilising each TCRVβ subunit, eight PBMC samples were stained with three anti-TCRVβ antibodies in parallel using the Beckman Coulter IOTest Beta mark kit. For the CD8+ killing experiments, PBMC were stained with an anti-TCRVβ antibody specific for the subunit most frequently utilised in that donor. Biotinylated antibodies were detected by staining with streptavidin-PeCy7 or -BV421 (Biolegend), 10 min at RT in FACS buffer. To stain intracellular antigens, cells were fixed and permeabilised using FoxP3 staining buffers (eBioscience, SanDiego), and stained with anti-Tax AF488 or anti-FoxP3 for 25 min at RT. Cells which were surface stained only were fixed with 2% paraformaldehyde in PBS for 20 min at RT. Data was acquired using a BD LSRFortessa, and analysed using Kaluza software. Gating strategy is outlined in S7 and S8 Figs.
The frequency of live CADM1+CD4+CD3+ cells which bound each anti-TCRVβ antibody was expressed as a percentage of total live CD3+ T cells. In order to estimate the frequency of T cells expressing Vβ subunits which were not recognised by antibodies in the panel, the sum of all positively identified TCRVβ subunits was subtracted from the total frequency of live CADM1+CD4+ within CD3+ cells. The frequencies of TCRVβ-expressing live CADM1+CD8+CD3+ cells were also calculated in the same manner, as CADM1+CD8+ cells are also heavily infected with HTLV-1 [36]. The resulting 50 frequencies (including instances where a particular population was undetectable within CD3+ cells), were used to compute the oligoclonality index as described for the proviral integration site data. To avoid introducing sampling error in the case of low PVL (and thus low frequencies of CADM1+ cells) flow cytometric data from donors for which <500 CADM1+ events were acquired were excluded from this analysis.
CD8+ cells were isolated from cryopreserved PBMC by positive selection using magnetic beads (Miltenyi Biotech) following the manufacturer’s protocol. The CD8+ fraction was placed in culture at 5 x105 cells/ml for 13 days in the presence of 1 μg/ml phytohemagluttinin-L (Sigma Aldrich) and 100 IU/ml IL-2 (Promocell). At three day intervals, 50% of the culture medium was replaced and supplemented with 100 IU/ml IL-2. Cells were split as required. Flow cytometric analysis indicated that the mean frequency of live CD8+ cells in each culture was 99.6%; with a mean residual contamination of on average 0.28% of the ATL clone from that donor. On day 13, CD8+ cells were depleted from a second vial of cryopreserved PBMC. Cultured or freshly isolated ex-vivo CD8+ cells were added to between 3x105 and 5x105 CD8− PBMCs at a range of effector: target (E:T) ratios in duplicate: CD8-depleted), the natural CD8+:CD4+ ratio (median 1:23), 1:4 and 1:2 as permitted by the number of CD8+ cells recovered. As significantly greater numbers of cultured CD8+ cells were recovered (versus ex vivo CD8+ cells) ratios of 1:1, 2:1 and 4:1 were tested where possible. Cells were co-cultured 1ml RPMI containing 10% FCS, 2 mM L-glutamine, 50 U/ml penicillin, 50 μg/ml streptomycin (Gibco) and 20 μg/ml DNAse (Sigma). After 18h, a 100 μl sample of each culture was harvested in order to count the absolute numbers of CD3+, CD4+ and CD8+ cells present. This was performed by adding 50 μl of an antibody master mix containing 1 μl anti CD8 AF700, 0.5 μl anti-CD3 BV510 and 0.25 μl anti-CD4 BV605 to each sample. Samples were incubated at RT for 30 min, after which 150 μl 2% paraformaldehyde in PBS was added, without any centrifugation/washing steps. Prior to flow cytometric analysis, 10 μl of CountBright absolute counting beads (Life Technologies) were added to each tube. The number of cells surviving was calculated as follows: # cells in tube = (# cells collected / # beads collected) × total # beads added to the tube. The remaining portion of each sample was analysed by flow cytometry as described above using the panel of antibodies outlined in S2 table.
The relative frequency of cells in each subset was obtained using the gating strategy outlined in S8 Fig. If the total number of CD4+ cells in the tube changed during the course of experiment, frequencies were normalised to the absolute count of CD4+ cells in the CD8-depleted culture condition. In addition, the exact E:T ratio (total CD3+CD8+ cells: total CD3+CD4+) which was achieved in the co-culture was quantified in each case.
The rate at which cells in a given subpopulation of cells were cleared (% target cells killed/%CD8/day) was estimated in each subject as described in Asquith et al [46] using the following equation: dy/dt = c- εyz; where y is the percentage of targets within total CD4+ cells, c is the rate of antigen presentation (assumed to be constant during the short-term culture), ε is the CD8+ cell-mediated lytic efficiency, and z is the proportion of CD3+ cells that are CD8+. This model was solved analytically and fitted to the data using nonlinear least-squares regression (SPSS v22).
|
10.1371/journal.ppat.1005850 | Consecutive Inhibition of ISG15 Expression and ISGylation by Cytomegalovirus Regulators | Interferon-stimulated gene 15 (ISG15) encodes an ubiquitin-like protein that covalently conjugates protein. Protein modification by ISG15 (ISGylation) is known to inhibit the replication of many viruses. However, studies on the viral targets and viral strategies to regulate ISGylation-mediated antiviral responses are limited. In this study, we show that human cytomegalovirus (HCMV) replication is inhibited by ISGylation, but the virus has evolved multiple countermeasures. HCMV-induced ISG15 expression was mitigated by IE1, a viral inhibitor of interferon signaling, however, ISGylation was still strongly upregulated during virus infection. RNA interference of UBE1L (E1), UbcH8 (E2), Herc5 (E3), and UBP43 (ISG15 protease) revealed that ISGylation inhibits HCMV growth by downregulating viral gene expression and virion release in a manner that is more prominent at low multiplicity of infection. A viral regulator pUL26 was found to interact with ISG15, UBE1L, and Herc5, and be ISGylated. ISGylation of pUL26 regulated its stability and inhibited its activities to suppress NF-κB signaling and complement the growth of UL26-null mutant virus. Moreover, pUL26 reciprocally suppressed virus-induced ISGylation independent of its own ISGylation. Consistently, ISGylation was more pronounced in infections with the UL26-deleted mutant virus, whose growth was more sensitive to IFNβ treatment than that of the wild-type virus. Therefore, pUL26 is a viral ISG15 target that also counteracts ISGylation. Our results demonstrate that ISGylation inhibits HCMV growth at multiple steps and that HCMV has evolved countermeasures to suppress ISG15 transcription and protein ISGylation, highlighting the importance of the interplay between virus and ISGylation in productive viral infection.
| Type I IFN response is a front-line defense against virus infection. Activation of type I IFN signaling leads to expression of a subset of cellular proteins encoded by interferon-stimulated genes (ISGs). ISG15 encodes an ubiquitin-like protein that is covalently conjugated to protein lysine residues. ISG15 modification (ISGylation) of a protein causes changes of protein function. ISGylation is known to inhibit the replication of many viruses, although pro-viral effects of ISGylation are also reported. Given that ISG15 and the enzymes involved in ISGylation are strongly induced upon virus infection, understanding the interplay between virus and ISGylation is an important issue in virus-host interaction. Nevertheless, viral substrates of ISG15 and viral strategies to regulate ISGylation-mediated antiviral responses are limited to only a few examples. In this study we demonstrate that ISGylation suppresses human cytomegalovirus (HCMV) infection but the virus is armed with countermeasures that consecutively reduce ISG15 transcription and protein ISGylation. Interestingly, a viral ISG15 target is found to inhibit ISGylation. This study highlights that ISGylation is a critical innate immune response against HCMV infection and interfering with ISG15-mediated anti-viral immunity is critical for productive viral infection.
| Type I interferons (IFNs) are multifunctional cytokines that represent crucial components of the innate immune response to viral infection. Recognition of viral infection by host cells induces the synthesis of type I IFNs. Secreted IFNs interact with IFN receptors on target cells, triggering a signaling cascade that involves Janus kinase (JAK) and signal transducer and activator of transcription (STAT) families. Activated STAT1 and STAT2 heterodimerize and bind to IFN regulatory factor 9 (IRF9) to form a complex called IFN-stimulated gene factor 3 (ISGF3). This complex translocates into the nucleus and induces ISGs with diverse antiviral activities by binding to IFN-stimulated response elements (ISREs) in their promoters (for review [1]).
ISG15 was identified as an IFN-inducible ubiquitin homolog. Like ubiquitin, its carboxy-terminal LRLRGG motif is required both for recognition by processing enzymes and covalent conjugation to lysine residues of target proteins. ISG15 modification (also termed ISGylation) is an IFN-stimulated and -regulated process that is found only in higher vertebrate animals and appears to modulate the function of target proteins (for review [2]). UBE1L is the E1 activating enzyme for ISG15 [3], and UbcH8, an ubiquitin E2 conjugating enzyme, also acts as the ISG15 E2 conjugating enzyme [4, 5]. HERC domain and RCC1-like domain containing protein 5 (Herc5), estrogen-responsive finger protein (EFP), and human homolog of Drosophila ariadne (HHARI) have been identified as E3 ligases for ISGylation in human cells [6–9]. ISG15, UBE1L, UbcH8, and Herc5 are IFN-inducible [4, 6, 10, 11]. Conjugated ISG15s are removed by an ISG15-specific protease, UBP43 (also known as USP18) [12]. Interestingly, UBP43 is also IFN-inducible [13, 14] and acts as a negative regulator of innate immune responses independent of its protease activity but dependent on its direct interaction with IFNAR2, a subunit of the type I IFN receptor [15].
Antiviral responses involving protein ISGylation have been reported against diverse viruses. Several cellular proteins involved in antiviral signaling, including RIG-I, MDA-5, STAT1, JAK1, IRF3, PKR, Mx1, and RNase L, were also identified or suggested as substrates for ISGylation [16–21]. ISGylation suppressed replication of diverse viruses, such as influenza virus (type A and B) [22–25], human immune deficiency virus (HIV) [26, 27], hepatitis C virus (HCV) [28–30], Japanese encephalitis virus [31], Sindbis virus [23, 32, 33], Ebola VP40 virus-like particle [34, 35], herpes simplex virus type-1 [23], murine γ-herpesvirus 68 [23], vaccinia virus [36], dengue and West Nile viruses [37], porcine reproductive and respiratory syndrome virus [38], Kaposi’s sarcoma-associated herpesvirus (KSHV) [39], and respiratory syncytial virus [40]. However, the antiviral mechanism of ISGylation against specific viruses is poorly understood. Herc5 associates with polyribosomes, and ISGylation appears to be restricted largely to newly synthesized proteins, suggesting that newly synthesized viral proteins may be primary targets of ISG15 [41]. ISGylation of NS1A in influenza A virus disrupted its association with importin-α, which mediates the nuclear import of NS1A, thus inhibiting viral replication [42]. ISGylation also suppressed the release of retrovirus particles by disrupting the budding process-related protein complex (for review [43]). An ISGylation-independent antiviral effect of ISG15 was also demonstrated in Chikungunya virus infection [44].
Although several studies have suggested a general role for ISG15 as an antiviral molecule, proviral effects of ISGylation have also been reported for certain viruses. In Newcastle disease virus (NDV) infection, ISGylation of RIG-I reduced IFN responses as a negative feedback regulation mechanism [45], whereas ISGylation of IRF3 stabilized IRF3 [21]. In addition, enhanced ISGylation by ISG15 overexpression promoted HCV production, while reduced ISGylation by UBE1L- or ISG15-knockdown inhibited HCV production [46, 47]. Therefore, the effects of global protein ISGylation appear to vary among the different viruses. In addition, ISG15 is secreted to the extracellular space as a free unconjugated form [48, 49]. While secretion of ISG15 in granulocytes is shown to activate T cells and natural killer cells to produce IFNγ in mycobacterial infection [50, 51], the role of secreted ISG15 in viral infection is not clear. Furthermore, a role of free ISG15 as a negative regulator that prevents IFNα/β overamplification and auto-inflammation by sustaining UBP43 levels has been suggested in humans but not in mice [52, 53].
Human cytomegalovirus (HCMV) is an opportunistic pathogen that causes severe disease complications and pathologies in newborns and immunocompromised individuals [54]. During productive infection, HCMV gene expression occurs sequentially in three phases: immediate-early (IE), early, and late. IE proteins and virion-associated tegument proteins play key roles in initiating viral gene expression and modulating host cell functions. HCMV employs several mechanisms to counteract IFN production and subsequent ISG activation. IE2 and pp65 inhibit IFN production [55–57], whereas IE1 suppresses the IFN response by directly binding to STAT2 [58–60] and PML [61, 62]. HCMV infection results in a decrease in the levels of JAK1 and p48, two components of the type I IFN signaling pathway [63, 64]. Modulation of the stability and phosphorylation of STAT proteins during HCMV infection was also reported [65, 66].
ISG15 transcription is induced in HCMV infection; however, its regulation during infection, the role of ISGylation in viral growth, and viral targets of ISG15 have not been characterized. In this study, we show that ISG15 expression and ISGylation are initially induced after HCMV infection but later suppressed by viral responses, and that IE1, a viral inhibitor of STAT signaling, plays an important role in reducing ISG15 transcription. By silencing the expression of E1, E2, and E3 ISGylation enzymes and of an ISG15 protease, we also demonstrate that ISGylation inhibits HCMV growth at multiple steps, including viral gene expression and virion release. Furthermore, we show that pUL26, a viral tegument protein, interacts with ISG15, E1, and E3 and is modified by ISG15, which inhibits pUL26 activity to promote viral growth. Moreover, we reveal that expression of pUL26 is able to suppress ISGylation induced by virus infection. Our results indicate that HCMV has evolved countermeasures to suppress ISG15 transcription and protein ISGylation, highlighting the important of the interplay between virus and ISG15 signaling during virus infection.
The time course of ISG15 expression and protein ISGylation during HCMV infection were investigated with different multiplicity of infections (MOIs). In human fibroblast (HF) cells infected with HCMV (Towne), the levels of ISG15 and protein ISGylation were elevated by 24 h at all MOIs tested (MOIs of 0.2 to 10) (Fig 1A, lanes 2–6). At 48 and 72 h after infection, even greater levels of ISG15 expression and protein ISGylation were observed at relatively low MOIs (0.2, 0.5, and 1) (Fig 1A, lanes 13–15 and 24–26); however, the levels of free ISG15 and ISG15 conjugates at high MOIs (3 and 10) were much lower than those at low MOIs (Fig 1A, lanes 16–17 and 27–28). The time course of ISG15 expression and protein ISGylation was also examined in cells infected with UV-inactivated virus (UV-HCMV). In UV-HCMV infection, the levels of ISG15 and ISG15 conjugates were elevated at 24 and 48 h and correlated proportionally with MOI (Fig 1A, lanes 7–11 and 18–22). Levels of free ISG15 at 72 h induced by UV-HCMV were lower than those at 48 h, probably due to the termination of signaling (Fig 1A, compare lanes 29–33 and 18–22). The lack of viral gene expression in UV-HCMV infection was verified by the absence of viral IE protein expression. Collectively, these results comparing HCMV and UV-HCMV infection demonstrate that ISG15 expression and protein ISGylation are initially induced by HCMV infection, but are subsequently suppressed in a manner dependent on viral gene expression.
Notably, we observed a greater induction of ISG15 and protein ISGylation with HCMV than with UV-HCMV at low MOIs (MOIs of 0.2, 0.5, and 1) (Fig 1A, compare lanes 2–4 and 7–9). We hypothesized that this was due to the induction of ISG15 at different level in uninfected cells that surround infected cells at low MOIs. To test this hypothesis and to examine the effect of HCMV infection on ISG15 expression at a single cell level, HF cells were infected with HCMV or UV-HCMV at a low MOI (0.5) and stained for ISG15 and viral UL112-113 proteins. We found that ISG15 expression was reduced in HCMV-infected cells, in which UL112-113 viral replication proteins were expressed at high level; however, it was markedly increased in neighboring uninfected cells compared to mock-infected cells. We also found that the levels of ISG15 in uninfected neighboring cells were higher in HCMV infection than in UV-HCMV infection (Fig 1B). This result suggests that an indirect effect of virus infection on neighboring uninfected cells is responsible for the greater induction of ISG15 and protein ISGylation by HCMV than by UV-HCMV at low MOIs.
Since HCMV IE1 inhibits the activation of ISRE-containing promoters by sequestering STAT2 [58–60] and PML [61], IE1 expression may be responsible for the suppression of free ISG15 and ISG15 conjugate levels during HCMV infection. To test this hypothesis, we first compared the effects of wild-type HCMV, UV-HCMV, and IE1-deleted mutant virus (CR208) infection (MOI of 3) on ISG15 transcription by RT-PCR. The CR208 virus exhibited an MOI-dependent growth pattern with showing a severe growth defect in HF cells at low MOIs but normal growth at high MOIs [67]. All viruses increased ISG15 mRNA levels 12 h after infection; however, ISG15 induction was terminated earlier for HCMV than UV-HCMV, and not appreciably terminated for CR208, which continued to produce high levels of ISG15 transcripts even at a late stage of infection (72 h) (Fig 2A). ISG15 transcription induced by UV-HCMV infection might be gradually decreased due to the termination of the IFN signaling through several negative regulatory mechanisms. However, when cells were infected with CR208 at this high MOI, the replication of virus without IE1 appeared to lead to a robust activation of IFN signaling. This result indicates that IE1 indeed plays an important role in reducing ISG15 transcription during HCMV infection.
We also compared the levels of ISG15 and ISG15 conjugates in HCMV, UV-HCMV, and CR208 infection. When cells were infected with viruses at an MOI of 3, free ISG15 levels were elevated at 12 h by all viruses, whereas the level of ISG15 conjugates markedly increased after 24 h (Fig 2B). The delayed induction of ISGylation is similar to what was observed in IFNβ-treated cells [68] and may result from the delayed induction of the ISGylation machinery. Consistent with the results shown in Fig 1A, UV-HCMV induced more ISG15 conjugates than wild-type virus at this MOI (Fig 2B, compare lanes 3–5 and 8–10). Importantly, CR208 also resulted in greater ISGylation than wild-type virus (Fig 2C. compare lanes 3–5 and 13–15), indicating that IE1 is required for the suppression of ISG15 expression and ISGylation during HCMV infection. Immunoblot analysis demonstrated that IE1 and IE2 failed to be expressed in UV-HCMV infection and that IE1 failed to be expressed and IE2 levels were reduced in CR208 infection under these experimental conditions (Fig 2B and 2C). Although our results show that IE1 is largely responsible for the suppression of ISG15 transcription, it is notable that CR208 infection resulted in slightly lower levels of ISGylation compared to UV-HCMV infection (Fig 2B, compare lanes 10 and 15). Considering that ISG15 transcript levels remained high up to 72 h in CR208-infected cells (Fig 2A), this finding suggests that other viral processes, besides those mediated by IE1, may also be implicated in the downregulation of protein ISGylation.
The inhibition of ISG15 expression by IE1 was further investigated using IE1-overexpressing HF cells generated using retroviral vectors. Control and IE1-overexpressing HF cells were infected with HCMV, UV-HCMV, or CR208. The results of immunoblot analysis showed that IE1 overexpression suppressed the induction of ISG15 and ISG15 conjugates by both virus infection and IFNβ treatment (as a control) (Fig 2C, compare lanes 2–5 and 7–10), further supporting the critical role of IE1 in reducing ISG15 expression.
Although reduction of protein ISGylation during HCMV infection may be largely attributed to suppression of ISG15 transcription by IE1, it cannot be ruled out that IE1 also affects the ISGylation reaction. Therefore, we studied whether IE1 directly affects ISG15 conjugation reactions using co-transfection/ISGylation assays. To set up co-transfection/ISGylation assays, two different ISG15 forms, an active form with a stop codon immediately after the C-terminal double glycine residues (ISG15GG) and an inactive form with the double glycine residues substituted with alanine residues (ISG15AA), were employed. In cells transiently co-transfected with ISG15, UBE1L (E1), UbcH8 (E2), and Herc5 (E3), intact ISG15 and ISG15GG (an active form), but not ISG15AA (an inactive form), were conjugated to proteins (Fig 2D). When co-transfection/ISGylation assays were performed with or without IE1 overexpression, IE1 expression by co-transfection or retroviral transduction did not inhibit levels of ISG15 conjugates, indicating that IE1 does not inhibit the enzymatic cascade of reactions required for protein ISGylation (Fig 2E and 2F).
The role of protein ISGylation in HCMV infection was investigated by silencing expression of ISGylation enzymes using RNA interference. HF cells expressing control shRNA or shRNA for UBP43, an ISG15-specific protease, were generated using lentiviral vectors. IFNα treatment of normal and control shRNA (shC)-expressing HF cells induced the expression of UBP43 proteins; however, UBP43-specific shRNAs (shUBP43-1 and shUBP43-2) efficiently suppressed this induction (Fig 3A). UBP43 knockdown enhanced protein ISGylation in cells infected with UV-HCMV that stimulates interferon signaling (Fig 3B), but reduced the production of progeny virions in HCMV-infected cells to nearly 10% of that in control cells (Fig 3C), suggesting that enhanced ISGylation by UBP43 knockdown attenuates HCMV growth. However, UBP43 is also known to negatively regulate IFN signaling by downregulating JAK/STAT signaling [15, 69]. Consistently, we found that UBP43 knockdown enhanced STAT2 phosphorylation in HCMV-infected cells (Fig 3D). Therefore, it is possible that the reduction of HCMV growth in UBP43 knockdown cells is also a result of other aspects of the IFN response.
The effect of ISGylation on HCMV growth was further assessed by depleting Herc5, a major ISG15 E3 ligase in human [6, 7]. Herc5 knockdown HF cells were generated using lentiviral vectors expressing shRNAs. qRT-PCR assays confirmed the efficient reduction of Herc5 transcript levels in cells expressing Herc5-specific shRNAs (shHerc5-1 and shHerc5-2) compared to in cells expressing control shRNA (shC) after UV-HCMV infection (Fig 3E). We found that Herc5 knockdown markedly reduced protein ISGylation as expected (Fig 3F), but increased virus titers by 8- to 11-fold compared to in control cells (Fig 3G), demonstrating that the reduction of ISGylation by Herc5 knockdown facilitates HCMV growth. Similar enhancement of HCMV growth was observed in HF cells depleted of UBE1L (E1) or UbcH8 (E2) (S1 Fig). Collectively, our results with UBP43, Herc5, UBE1L, and UbcH8 knockdown cells demonstrate an inverse relationship between ISGylation and HCMV growth, indicating a general inhibitory role of ISGylation in HCMV infection.
We also investigated whether enhanced ISGylation affects HCMV growth by ectopically expressing ISG15 together with ISGylation enzymes. HF cells co-transfected with plasmids expressing myc-ISG15GG or myc-ISG15AA and ISGylation enzymes were infected with HCMV and the production of progeny virions was compared. The results showed that the expression of myc-ISG15GG and ISGylation enzymes induced higher levels of ISG15 conjugates and reduced progeny virus titers to 50% of control, whereas the expression of ISG15AA did not significantly affect the levels of ISG15 conjugates or progeny virions (S2 Fig). These results are also consistent with our results using shRNA-expressing cells in which we showed an inverse relationship between levels of ISG15 conjugates and HCMV growth.
To investigate the mechanisms by which ISGylation inhibits HCMV growth, we first compared the expression profile of viral proteins in control and Herc5 knockdown cells. When cells were infected with an MOI of 0.2, levels of major IE (IE1 and IE2), early (p52), and late (pp28) viral proteins produced at 24, 48, 72, and 96 h were higher in Herc5 knockdown cells than in control cells (Fig 4A, left panels). This effect of knocking down Herc5 diminished when cells were infected at higher MOIs. At an MOI of 1, IE1 expression was similar between control and Herc5 knockdown cells, although the levels of IE2, p52, and pp28 were slightly increased in Herc5 knockdown cells (Fig 4A, center panels). At an MOI of 5, levels of viral proteins were comparable between control and Herc5 knockdown cells (Fig 4A, right panels). Progeny virion titers measured in the culture supernatant correlated with the levels of expressed viral proteins (Fig 4B). These results demonstrate that reduced ISGylation by silencing Herc5 promotes viral gene expression, an effect that is more evident at lower MOIs.
We further investigated whether ISGylation affects the activation of viral promoters using reporter assays (S3 Fig). HF cells co-transfected with ISG15, UBE1L, UbcH8, and Herc5 exhibited substantially increased protein ISGylation, whereas cells co-transfected without Herc5 did not. In HF cells exhibiting enhanced ISGylation by co-transfection, the activity of viral MIE promoter was repressed to 50% of that observed in control cells. Similarly, IE2-mediated activation of viral early [UL112-113 and UL54 (POL)] and late [UL99 (pp28)] promoters was also suppressed in cells showing enhanced ISGylation. In control experiments, ectopic expression of Herc5 alone was not sufficient to increase ISGylation and therefore did not affect viral promoter activation. These results demonstrate that enhanced ISGylation inhibits the activation of viral promoters.
We also assessed whether ISGylation inhibits HCMV virion release as is the case in HIV infection [43]. Control and Herc5 knockdown cells were infected with HCMV at an MOI of 1 or 5 and the levels of cell-associated and extracellular progeny virions produced were compared. The results showed that although more progeny virions were produced in Herc5 knockdown cells than in control cells (Fig 4B), the percentages of cell-associated virions to total virions were lower in Herc5-knockdown cells, 3.5% (MOI 1) and 3.8% (MOI 5), than in control cells, 19.7% (MOI 1) and 17% (MOI 5) (Fig 4C). These results indicate that, as observed in HIV infection, ISGylation inhibits HCMV virion release.
To identify potential HCMV proteins that interact with the ISG15 pathway, we screened the HCMV ORF library [70] for ISG15- or UBE1L-interacting proteins using yeast two-hybrid assays. Twenty HCMV proteins were identified as potential ISG15-interacting proteins, and five viral proteins (encoded from RL1, UL19, UL21A, UL26, and UL30) were found to interact with both ISG15 and UBE1L. Most of these ISG15 binding and all of UBE1L binding in yeast assays were also detected by co-immunoprecipitation (co-IP) assays (S4 Fig and S2 Table). Among them, the UL26 gene encodes the tegument proteins, p27 and p21, which are produced using two in-frame start codons and are shown to regulate viral gene expression, NF-κB signaling, and virion stability [71–74]. Since UL26 interacted with both ISG15 and UBE1L and its role in viral growth was relatively well reported compared to others, we further investigated the interaction of pUL26 with the ISG15 pathway. In co-IP assays, UL26-p21 interacted with ISG15AA, demonstrating that ISG15 can non-covalently interact with UL26 (Fig 5A). In similar co-IP assays, UL26-p21 also interacted with UBE1L and Herc5, but not UbcH8 (Fig 5B–5D). When co-IP assays were performed using HCMV (Towne)-infected cell lysates, immunoprecipitation with an anti-UL26 antibody co-precipitated unconjugated free ISG15 (Fig 5E), and in a reciprocal experiment, immunoprecipitation with an anti-ISG15 antibody co-precipitated UL26-p21 (Fig 5F), indicating that UL26-p21 interacts with ISG15 during virus infection.
It has been suggested that newly synthesized viral proteins may be broadly modified by ISG15 during virus infection [41]. Therefore, we also investigated whether UL26 is covalently conjugated by ISG15 during HCMV infection. Cell lysates prepared from HCMV (Towne)-infected cells were boiled in SDS-containing buffer and then immunoprecipitated with an anti-UL26 antibody. Immunoblotting of the sample with anti-ISG15 antibody revealed a band that is consistent with an ISG15-modified form of UL26-p21 (Fig 5G). To further investigate these non-covalent and covalent interaction of UL26 proteins with ISG15, we generated a recombinant HCMV (Toledo) expressing UL26 proteins tagged at their carboxyl termini with an HA tag (S5 Fig). Compared to the wild-type virus, the recombinant virus produced equivalent amounts of viral major IE (IE1 and IE2) and UL26 proteins, and progeny virions (Fig 5H and 5I). When co-IP assays were performed with lysates from UL26-HA virus infected cells, immunoprecipitation of UL26 proteins with an anti-HA antibody co-precipitated free ISG15 (Fig 5J). When HA-UL26 virus-infected cells were subjected to co-IP assays to detect ISGylated proteins, bands that are consistent with ISG15-modified forms of UL26 proteins were detected (Fig 5K). Since ISG15 can modify ubiquitin, forming ISG15-ubiquitin mixed chains [75] and a single lysine reside can be poly-ISGylated [20, 76], the smear bands of ISGylated UL26 appear to be UL26 proteins that contain ISG15-ubiquitin mixed chains or poly-ISG15 chains. Taken together, our data suggest that the UL26-encoded proteins non-covalently interact with ISG15 and are also covalently modified by ISG15. In a control experiment, we found that some viral proteins, which did not interact with ISG15, UBE1L, and Herc5 in co-IP assays, could be ISGylated in co-transfection/ISGylation assays (S6 Fig), supporting the concept that viral proteins may be broadly ISGylated during infection due to IFN-upregulated expression of Herc5 (E3) in polyribosomes [41].
The UL26 ORF from the Towne strain contains three lysine residues (K54, K136, and K169) and K54 and K169 are conserved in human CMVs (Fig 6A). To determine the ISGylation sites of UL26 proteins, we performed co-transfection/ISGylation assays using wild-type UL26-p21 or its mutants in which lysine residues were replaced with arginines. The results showed that the K54R, K136R, and K169R mutants were still ISGylated; however, the K136/169R double mutant was not ISGylated, indicating that K136 and K169 are the major ISGylation sites (Fig 6B). When HF cells expressing wild-type UL26-p21 or K136/169R mutant were generated by retroviral vectors, expression level of the K136/169R mutant protein was higher than that of the wild-type protein (Fig 6C), suggesting that ISGylation of UL26 proteins may regulate protein stability.
The SUMO fusion proteins were often used to study the function of SUMO modification of proteins [77–81]. To investigate the effect of UL26 ISGylation on its activity, we used the UL26-p21(K136/169R)-ISG15AA fusion protein as a surrogate for the ISG15-modified form of UL26-p21. We produced the ISG15 fusion to the lysine mutant form of UL26 to compare the activities of ISGylation-defective UL26 and its ISG15 fusion form. HF cells expressing the K136/169R mutant or the UL26-p21(K136/169R)-ISG15AA fusion protein were generated by retroviral vectors (Fig 6D). UL26 proteins have been shown to inhibit TNFα-induced NF-κB activation [74]. UL26-p21(K136/169R) moderately inhibited TNFα-induced NF-κB activation but UL26-p21(K136/169R)-ISG15AA did not, suggesting that ISGylation of UL26 inhibits its activity to downregulate NF-κB signaling (Fig 6E). We further investigated the effect of UL26 ISGylation on its role in viral growth. When control and UL26-expressing cells were infected with the UL26-deleted mutant HCMV (AD169) [72] and the levels of progeny virus titers were compared, pre-expression of the K136/169R mutant significantly increased the growth of mutant virus but its ISG15AA fusion protein did not (Fig 6F). Together, these results indicate that ISGylation of UL26 inhibits its activities to suppress NF-κB signaling and promote the growth of UL26-deleted mutant virus.
Since influenza virus NS1B antagonizes ISGylation in human cells via direct binding to ISG15 [3, 82], we tested whether UL26 can regulate protein ISGylation. We found that expression of UL26-p21 reduced the levels of ISG15 conjugates in cells co-transfected with UBE1L (E1), UbcH8 (E2), Herc5 (E3), and ISG15GG, an active form of ISG15 (Fig 7A). In a similar assay, ISGylation of charged multivesicular body protein (CHMP) 5, a component of the endosomal sorting complex required for transport (ESCRT) machinery, was inhibited by UL26-p21 expression (Fig 7B). When control and UL26-p21-expressing HF cells were infected with UV-HCMV, the levels of ISG15 conjugates were significantly reduced in UL26-p21-expressing cells compared to control cells (Fig 7C). This effect of UL26 was also found with the K136/169R mutant protein (Fig 7D), which still interacted with ISG15AA in co-IP assays (Fig 7E), suggesting that the inhibitory effect of UL26 on ISGylation is not dependent on its own ISGylation. In control experiments, viral proteins such as pUL85, pUL71, and IE2 did not inhibit ISGylation of cellular proteins in co-transfection/ISGylation assays (S6 and S7 Figs).
To further investigate the inhibitory effect of UL26 on ISGylation during HCMV infection, we compared ISGylation levels between cells infected with wild-type and UL26-deleted mutant viruses. With an MOI of 0.2, we found that the ISGylation levels were initially increased until 48 h but gradually decreased at 72 h and 96 h, and that the UL26-deleted virus infection showed only minimally increased levels of ISGylation at late times of infection compared to wild-type virus infection (S8 Fig, compare lanes 5–7 and 11–13). We reasoned this minimal effect of UL26 to the suppression of ISG15 transcription by IE1. Therefore, to minimize the effect of IE1 expression we performed a similar experiment in cells expressing shRNA for IE1 (shIE1). Control shRNA (shC) and shIE1-expressing HF cells, which were produced by retroviral vectors, were infected with wild-type and UL26-deleted mutant viruses at an MOI of 5. The results of immunoblotting showed that the expression of shIE1 substantially reduced the IE1 protein accumulation compared to control cells (Fig 7F, compare lanes 2–5 and 7–10), and that in shIE1-expressing cells the ISGylation levels were substantially increased at 24 h after infection and gradually decreased at 48 h and 72 h under these experimental conditions (Fig 7F, lanes 8–10). Notably, we found that the UL26-deleted virus less effectively reduced the ISGylation levels than wild-type virus at the late phase of infection (Fig 7F, compare lanes 8–10 and 13–15). These results demonstrate an inhibitory effect of UL26 on ISGylation during virus infection.
Since HCMV growth was suppressed by IFN-induced ISGylation and UL26 inhibited ISGylation, we tested whether UL26-deleted mutant virus is more susceptible to type I IFN treatment than wild-type virus. We found that while the titers of wild-type virus were reduced by 12-fold by IFNβ pre-treatment, those of UL26-deleted virus were reduced by 40-fold, indicating that UL26-deleted virus less effectively overcomes the IFNβ-mediated anti-viral responses than wild-type virus (Fig 7G). Overall, our results demonstrate that, in addition to IE1 that suppresses ISG15 transcription, UL26 plays an important role in evading the ISG15-associated antiviral responses by inhibiting protein ISGylation (Fig 7H).
Our analysis with UV-HCMV and IE1-deleted mutant virus demonstrates that ISGylation is induced by HCMV infection and that IE1 plays a central role in downregulating ISGylation by reducing ISG15 transcription. The latter is consistent with previous findings that IE1 represses transcription of ISGs by sequestrating STAT2 [58–60] and PML [61, 62]. Notably, although IE1 effectively reduced ISG expression, the level of ISGylation during HCMV infection largely depended on the MOI.
Our data provide evidence for the antiviral roles of ISGylation during HCMV infection. To discern the effects of free ISG15 expression and protein ISGylation on virus replication, we used HF cells in which a specific ISGylation enzyme was depleted by shRNA. The data consistently showed an inverse relationship between the level of ISG15 conjugates and HCMV growth; i.e., enhanced ISGylation by UBP43 knockdown decreased viral growth, while reduced ISGylation by depletion of UBE1L, UbcH8, or Herc5 increased viral growth. Therefore, we conclude that protein ISGylation in general inhibits HCMV growth. Free ISG15 has also been shown to inhibit the replication of certain viruses [34, 35, 44]. In our analysis, however, overexpression of ISG15GG with UBE1L, UbcH8, and Herc5 led to a mild reduction of HCMV growth, whereas ISG15AA, an inactive form, did not significantly affect viral growth, suggesting that expression of free ISG15 prior to HCMV infection may minimally affect viral growth in cultured cells. It should be noted that given the role of free ISG15 in stabilizing UBP43 (or USP18), a negative regulator of IFN signaling [53], overexpression or depletion of ISG15 might affect both positive and negative activities of ISG15 to viral growth.
HCMV replication was inhibited by ISGylation at multiple steps. First, the expression of viral genes was inhibited under conditions where the level of ISG15 conjugates was increased. Viral regulators that promote viral gene expression may be a direct target of ISGylation. However, we could not observe ISGylation of IE1, which is responsible for the activation of MIE promoters, or IE2, a strong transactivator of viral early and late genes. ISGylation of cellular proteins that in particular play a role in innate immune responses may affect viral gene expression. Notably, ISGylation of IRF3 increases its stability and enhances IRF3-mediated transcriptional activation during Sendai virus infection [16, 21] and ISGylation of 4EHP, an mRNA 5' cap structure-binding translation suppressor, plays a role in IFN-induced innate immune response [9]. Second, HCMV virion release was inhibited by ISG15 conjugation. The inhibitory role of ISG15 expression in the budding process of enveloped viruses has been demonstrated in retrovirus infection. ISG15 expression inhibits ubiquitination of HIV Gag and tumor susceptibility gene-101 (Tsg101) proteins, leading to disruption of their interaction [26]. ISG15 conjugation of CHMP5, 2A, and 6 in the ESCRT machinery causes release of vacuolar protein-sorting 4 (VPS4) from the membrane, leading to inhibition of virion release [43]. The ESCRT machinery seems to be involved in the process of HSV-1 and HCMV maturation [83–86]. Therefore, ISGylation may affect the HCMV maturation process by targeting components involved in the ESCRT machinery.
Although the antiviral role of ISG15 expression and ISGylation has been demonstrated in several viruses, studies on viral targets for ISGylation and viral strategies that interfere with the ISG15-mediated antiviral functions are limited to a few examples. ISGylation of NS1A of influenza A virus inhibits virus replication by interfering with NS1A nuclear import [42]. KSHV vIRF1 is ISGylated but it role on vIRF1 function is not known [39]. In the present study, we demonstrated that HCMV pUL26 is a target for ISGylation. UL26 ISGylation appears to regulate protein stability by competing with ubiquitination. Our analysis using the UL26-p21-ISG15 fusion protein demonstrated that ISGylation inhibits the activities of UL26-p21 to downregulate the TNFα-mediated NF-κB activation and to complement the growth defect of UL26-deleted virus. Therefore, ISGylation of pUL26 is thought to inactivate its function, suppressing HCMV growth. Given the notion that newly synthesized proteins are targeted extremely broadly, perhaps stochastically, by Herc5 in polyribosomes [41], several other HCMV proteins may be ISGylated during infection. We indeed observed that some viral proteins, which did not interact with ISG15, UBE1L, and Herc5 in co-IP assays, were ISGylated in our co-transfection/ISGylation assays. However, we think that ISGylation occurs in a manner dependent on the context of each protein. Furthermore, since several HCMV proteins were bound to ISG15 and/or UBE1L in yeast and co-IP assays, it is also likely that other viral proteins besides pUL26 may affect ISGylation. Identification of more HCMV proteins that interact with the ISGylation system and studies on their functional relevance are warranted.
A few examples for viral regulation of ISGylation have been described. The NS1 protein of influenza B virus non-covalently binds to ISG15 and inhibits ISGylation [3]. Similarly, the vaccinia virus E3 protein interacts with ISG15 and disrupts its antiviral activity [36]. Nairoviruses and arteriviruses have shown to encode ovarian tumor domain (OTU)-containing proteases that hydrolyze ISG15 from target proteins [87] and severe acute respiratory syndrome (SARS) coronavirus encoded the papain-like protease that cleaves both ubiquitin and ISG15 [88]. Recently, KSHV-encoded vIRF1 was shown to interact with Herc5 [39]. In this study, we demonstrated that HCMV UL26-p21 is able to non-covalently bind to ISG15, UBE1L, and Herc5, and inhibit ISGylation. Whether binding all of ISG15, UBE1L, and Herc5 is critical for pUL26 to inhibit ISGylation is not clear and needs to be further investigated. Comparative analysis of wild-type and UL26-deleted mutant viruses demonstrate that de novo expression of UL26 in virus-infected cells correlated with reduced accumulation of ISG15 conjugates. In addition to IE1, the presence of additional viral functions that downregulate protein ISGylation was prompted by our observation that UV-HCMV infection resulted in a higher level of ISG15 conjugates than IE1-deleted mutant virus at late times (Fig 2B). UL26 deletion mutant virus shows a moderate growth defect at low MOIs [72, 89]. Notably, the UL26 virus (in AD169 strain) showed a reduced plaque formation at low MOIs and this defect could be rescued by IE1 overexpression [72]. More importantly, like IE1-deleted virus [58–60], the growth of UL26-deleted virus was more sensitive to pre-treatment of type I IFNs [74] (and this study), suggesting a role of UL26 in antagonizing type I IFN response. It is likely that a common ISG15-targeting mechanism is shared by influenza B virus, vaccinia virus, and HCMV. HCMV also encodes a tegument protein pUL48 that contains a deubiquitinating protease domain; however, its activity was specific for ubiquitin and did not cleave ISG15 [90].
In this study, we demonstrate that the cellular ISGylation system is a critical part of cellular innate immune response against HCMV infection. We also provide evidence that HCMV has developed several strategies to disarm the ISGylation-mediated antiviral activity; IE1 reduces ISG15 transcription and UL26 inhibits protein ISGylation. This interplay of HCMV with the cellular ISGylation system may be critical for the virus to successfully establish a persistent infection.
Human foreskin fibroblast (HF) (ATCC) and human embryonic kidney (HEK) 293T cells (ATCC) were grown in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum, penicillin (100 U/ml), and streptomycin (100 μg/ml). DNA transfection of 293T cells was performed using the N,N-bis-(2-hydroxyethyl)-2-aminoethanesulfonic acid-buffered saline (BBS) version of the calcium phosphate procedure. Electroporation of HF cells was conducted using a Microporator MP-100 (Digital Bio), as described previously [59]. Stocks for the parent Towne virus and the CR208 mutant virus with IE1-deleted were prepared in ihfie1.3 cells as described previously [59]. The HCMV (Towne strain)-GFP virus was grown in HF cells after electroporation with the bacmid DNAs. Wild-type and UL26-deleted HCMVs (AD169 strain) were previously described [72] and grown in UL26-expressing HF cells. Wild-type and UL26-HA-expressing HCMVs (Toledo strain) generated in this study were grown in HF cells. To produce UV-inactivated HCMV (UV-HCMV), the virus stock was irradiated with UV light three times at 0.72 J/cm2 using a CL-1000 Crosslinker (UVP).
Mammalian expression plasmids for HA-IE1 (pDJK170) and HA-IE2 (pDJK171) were cloned using the pSG5 vector and plasmid for myc-ISG15 (pOK20) was cloned using the pCS3-MT (with a hexa-myc tag) vector using Gateway technology as previously described [59]. Plasmids for myc-ISG15GG (pYJ12), an active form of ISG15 with a termination codon added immediately after the double glycine residues, or myc-ISG15AA (pYJ14), a conjugation-defective mutant in which the double glycine residues are replaced with alanine residues, were produced using the Stratagene QuickChange site-directed mutagenesis protocol. The pSG5-driven plasmids expressing Flag-UbcH8 (pYJ23) and HA-Herc5 (pYJ29) were produced using Gateway technology. Plasmids for SRT-UL26-p21 (pSE124) and Myc-UL26-p21 (pSE107) were cloned using the pcDNA6 (Life Technologies) vector and pCS3-MT vector, respectively, using Gateway technology. For the SRT-UL26-p21 expression plasmid used in ISGylation assays, two lysine residues on the linker between the SRT tag and the UL26 ORF were changed to alanines to block their possible ISGylation, resulting in pYJ178. Site-directed mutagenesis was performed on the pYJ178 background to produce plasmids expressing the lysine to arginine mutant versions of SRT-UL26-p21; K54R (pYJ179), K136R (pYJ180), K136/169R (pYJ182), and K54/136/169R (pYJ183). Plasmids for HA-hUBE1L (pCAGGS-HA-hUBE1L) and Flag-hUbcH8 (pFlagCMV2-UbcH8) and plasmids for S-tagged Herc5 (pCI-neo-S2-Herc5) were kindly provided by Dong-Er Zhang (Moores Cancer Center, University of California, San Diego, La Jolla, CA 92093, USA).
Yeast AH109 (MATa) cells were transformed with plasmid expressing the GAL4-DNA-binding (DB)-ISG15 (TRP+) or GAL4-DB-UBE1L (TRP+) fusion protein. Y187 (MATα) cells were transformed with plasmid expressing the GAL4-activation domain (A)-HCMV ORF (LEU2+) fusion proteins. Each transformant was selected on plates lacking tryptophan (SC-Trp) or leucine (SC-Leu). Trp+ and Leu+ transformants were mated with each other on YPD plates. Diploid cells (a/α) were selected on plates lacking both tryptophan and leucine (SC-TrpLeu). Trp+Leu+ colonies were tested for their growth on plates that lack tryptophan, leucine and histidine (SC-TrpLeuHis). Cells expressing bait and prey that interact with each other grow on SC-TrpLeuHis. Cells expressing both GAL4- DB-ISG15 and GAL4-A-UBE1L were used as a positive control, whereas cells expressing GAL4-DB-ISG15 and GAL4-A only were used as a negative control.
The Toledo-BAC clone encoding UL26 proteins with a C-terminal HA tag was produced by using a counter-selection BAC modification kit (Gene Bridges). The scheme for bacmid mutagenesis is described in S5 Fig and the LMV primers used for mutagenesis are listed in S1 Table.
Retroviral vectors expressing IE1 (pYH38) or UL26 (pYJ104), UL26(K136/169R) (pYJ176), and UL26(K54/136/169R) (pYJ177) were produced on the background of pMIN (murine leukemia virus-based retroviral vector) using Gateway technology as previously described [59]. Retroviral vectors expressing shRNA for UbcH8 (pMSCVpuro-shUbcH8) was previously described [5]. To produce retroviral vectors expressing shRNA for UBE1L (pMSCVpuro-shUBE1L-1), short hairpin RNA (shRNA) for UBE1L was amplified with U6 promoter by PCR with primers 5′- TTTGGATCCCAAGGTCGGGCAGGAAGAGGGCCTATTTCC-3′ and 5′-TTTGAATTCAAAAAGGATGATGACAGCAACTTCTCTCTTGAAGAAGTTGCTGTCATCATCCGGTGTTTCGTCCTTTCCACAAGATATATAA-3′ (target sequence underlined). The PCR product was digested with BamHI and EcoRI and ligated to MSCV-PGKpuro (BD Biosciences Clontech) digested with BglII and EcoRI. Recombinant retroviruses were prepared in 293T cells after co-transfection with retroviral vectors together with the packaging plasmids pHIT60 (Gag-Pol) and pMD-G expressing the envelope G protein of vesicular stomatitis virus (VSV) [59] using Metafectene reagents (Biotex). Viral supernatants were collected at 48 h after transfection. HF cells were transduced by retroviruses in the presence of polybrene (7.5 μg/ml). Cells were selected with G418 (0.5 mg/ml) (Calbiochem) and maintained in a medium containing G418 (0.1 mg/ml).
Lentiviral vector pLKO.1-TRC control expressing a non-hairpin control RNA (shC) was purchased from Addgene. pLKO.1-based lentiviral vectors expressing shRNA for UBP43 (shUBP43-1: TRCN0000004194 and shUBP43-2: TRCN0000004195) and Herc5 (shHerc5-1: TRCN0000004171 and shHerc5-2: TRCN0000004169) were purchased from Open Biosystems. To produce lentiviruses, 293T cells were transfected with lentiviral vectors together with plasmids pCMV-DR8.91 expressing the gag-pol, tat, and rev proteins of human immunodeficiency virus (HIV) and pMD-G. At 48 h, the viral supernatants were collected and used to transduce HF cells in the presence of polybrene (7.5 μg/ml). The transduced cells were selected with puromycin (1 μg/ml) and maintained in a medium containing puromycin (0.5 μg/ml).
Mouse monoclonal antibody (MAb) 810R, which detects epitopes present in both IE1 and IE2, was purchased from Chemicon. Mouse MAbs against UL44 (p52) and UL99 (pp28) were obtained from Virusys. Anti-β-actin and anti-α-tubulin mouse MAbs were purchased from Sigma. Anti-HA rat MAb 3F10 and anti-myc mouse MAb 9E10, conjugated with peroxidase or labeled with fluorescein isothiocyanate (FITC), were purchased from Roche. Anti-ISG15 (F-9) and anti-STAT2 mouse MAbs were obtained from Santa Cruz. Mouse MAb against SRT epitope was previously described [91]. UBP43 antibody was previously described [92]. Rabbit polyclonal Ab (PAb) for STAT2 (C-20) and STAT2 phosphorylated at Tyr689 were purchased from Santa Cruz and Upstate, respectively. Rabbit PAb for ISG15 was kindly provided by Chin Ha Chung (Seoul National University, Seoul, Republic of Korea).
For immunoblot analysis, cells were washed with phosphate-buffered saline (PBS) and total cell lysates were prepared by boiling the cell pellets in sodium dodecyl sulfate (SDS) loading buffer. Equal amounts of the clarified cell extracts were separated on a SDS-polyacrylamide gel or Gradi-Gel II (Elpis biotech, Republic of Korea) and electroblotted onto nitrocellulose membranes. The blots were blocked by incubation for 30 min at room temperature with PBS plus 0.1% Tween 20 (PBST) containing 5% nonfat dry milk. After being washed with PBST three times, the blots were incubated with the appropriate antibodies in PBST for 1 h at room temperature. After three 5 min washes with PBST, the blots were incubated with horseradish peroxidase-conjugated goat anti-mouse IgG or anti-rabbit IgG (Amersham) for 1 h at room temperature. The blots were then washed three times with PBST, and the protein bands were visualized with enhanced chemiluminescence system (Amersham).
For IFA, cells were fixed in ice-cold methanol for 5 min and rehydrated in cold PBS. Then, the cells were incubated with appropriate primary antibodies in PBS at 37°C for 1 h, followed by incubation with appropriate secondary antibodies at 37°C for 1 h. The mounting solution containing Hoechst and anti-fade reagent (Molecular Probes) was used. For double-labeling, two different antibodies were incubated together. Slides were examined and with a Carl Zeiss LSM710Meta confocal microscope system.
HF cells (2 × 105) were collected and incubated with 200 μl of lysis buffer [40 mM Tris-HCl (pH 7.8), 50 mM NaCl, 2 mM EDTA, 1 mM MgSO4, and 1% Triton X-100 plus 5 mM dithiothreitol] for 20 min on ice. The extracts were clarified in a microcentrifuge and 20 μl of extracts were incubated with 350 μl of reaction buffer A (25 mM Gly-Gly pH 7.8, 15 mM ATP and 4 mM EGTA) and then mixed with 100 μl of 1 mM luciferin (Sigma). A TD-20/20 luminometer (Turner Designs) was used for the 10-s assay of the photons produced (measured in relative light units).
The diluted samples were used to inoculate a monolayer of 4 × 104 HF cells in a 24-well plate. At 24 h post infection, cells were fixed with 500 μl of cold methanol for 10 min. The cells were then washed three times in phosphate-buffered saline (PBS) and incubated with anti-IE1 rabbit polyclonal antibody in PBS at 37°C for 1 h, followed by incubation with phosphatase-conjugated anti-rabbit immunoglobulin G (IgG) antibody in phosphate-buffered saline (PBS) at 37°C for 1 h. Finally, the cells were gently washed in PBS and treated with 200 μl of developing solution (nitroblue tetrazolium/5-bromo-4-chloro-3-indolylphosphate) at room temperature for 1 h. The positively stained cells were counted for at least three to five separate fields per well under a light microscope (× 200 magnification).
Co-transfected 293T cells (8 × 105) or virus-infected HF cells were harvested and sonicated in 0.7 ml co-IP buffer [50 mM Tris-Cl (pH 7.4), 50 mM NaF, 5 mM sodium phosphate, 0.1% Triton X-100, containing protease inhibitors (Sigma)] by a microtip probe (Vibra-Cell; Sonics and Materials, Inc., USA) for 10 s (pulse on: l s, pulse off: 3 s). Cell lysates were incubated with appropriate antibodies. After incubation for 16 h at 4°C, 30 μl of a 50% slurry of protein A- and G-Sepharose (Amersham) were added and then the mixture was incubated for 2 h at 4°C to allow adsorption. The mixture was then pelleted and washed 7 times with co-IP buffer. The beads were resuspended and boiled for 5 min in loading buffer. Each sample was analyzed by SDS-PAGE and immunoblotting with appropriate antibodies.
For co-transfection/ISGylation assays, 293T cells were co-transfected with plasmids expressing target protein and ISGylation enzymes. Co-transfected or virus-infected cells were treated with 0.5 mM NEM (N-ethylmaleimide) for 30 min before they were harvested. Cell pellets were resuspended with 10% SDS lysis buffer containing protease inhibitors and boiled for 10 min. Cell lysates were diluted 10-fold with co-IP buffer (50 Mm Tris-Cl [pH 7.4], 50 mM NaF, 5 mM sodium pyrophosphate, containing protease inhibitors) and sonicated by using a Microtip probe (Vibra cell; Sonics and Materials, Inc.). The clarified cell lysates were incubated with appropriate antibody for 16 h and then with 30 μl of a 50% slurry of protein G for 2 h. The mixture was pelleted and washed seven times with co-IP buffer. The bound proteins were boiled and analyzed by SDS-PAGE followed by immunoblot assays
Total RNAs were isolated from 2 × 105 cells using TRIzol reagent (Invitrogen) and MaXtract High Density (Qiagen). First-strand cDNA was synthesized by using the random hexamer primers in the SuperScript III system (Invitrogen). Quantitative real-time TR-PCR (qRT-PCR) was performed using the Applied Biosystems ABI Prism SDS software and the following primers: for ISG15, 5ʹ-GCTGGGACCTGACGGTG-3ʹ (sense) and 5ʹ-TTAGCTCCGCCCGCCAG-3ʹ (anti-sense); for UBE1L, 5ʹ-AGGTGGCCAAGAACTTGGTT-3ʹ (sense) and 5ʹ-CACCACCTGGAAGTCCAACA-3ʹ (anti-sense); for UbcH8, 5ʹ-AACCTGTCCAGCGATGATGC-3ʹ (sense) and 5ʹ-TGGTGCAAGGCTTCCAGTTC-3ʹ (anti-sense); for Herc5, 5ʹ-GGGATGAAAGTGCTGAGGAG-3ʹ (sense) and 5ʹ-CATTTTCTGAAGCGTCCACA-3ʹ (anti-sense); for β-actin, 5ʹ-AGCGGGAAATCGTGCGTG-3ʹ (sense) and 5ʹ-CAGGGTACATGGTGGTGCC-3ʹ (anti-sense).
Statistical significances were determined using the Student’s t-test and are indicated by *P<0.05, **P<0.01, or ***P<0.001.
|
10.1371/journal.pgen.1002414 | Target Site Recognition by a Diversity-Generating Retroelement | Diversity-generating retroelements (DGRs) are in vivo sequence diversification machines that are widely distributed in bacterial, phage, and plasmid genomes. They function to introduce vast amounts of targeted diversity into protein-encoding DNA sequences via mutagenic homing. Adenine residues are converted to random nucleotides in a retrotransposition process from a donor template repeat (TR) to a recipient variable repeat (VR). Using the Bordetella bacteriophage BPP-1 element as a prototype, we have characterized requirements for DGR target site function. Although sequences upstream of VR are dispensable, a 24 bp sequence immediately downstream of VR, which contains short inverted repeats, is required for efficient retrohoming. The inverted repeats form a hairpin or cruciform structure and mutational analysis demonstrated that, while the structure of the stem is important, its sequence can vary. In contrast, the loop has a sequence-dependent function. Structure-specific nuclease digestion confirmed the existence of a DNA hairpin/cruciform, and marker coconversion assays demonstrated that it influences the efficiency, but not the site of cDNA integration. Comparisons with other phage DGRs suggested that similar structures are a conserved feature of target sequences. Using a kanamycin resistance determinant as a reporter, we found that transplantation of the IMH and hairpin/cruciform-forming region was sufficient to target the DGR diversification machinery to a heterologous gene. In addition to furthering our understanding of DGR retrohoming, our results suggest that DGRs may provide unique tools for directed protein evolution via in vivo DNA diversification.
| Diversity-generating retroelements function through a unique, reverse transcriptase–mediated “copy and replace” mechanism that enables repeated rounds of protein diversification, selection, and optimization. The ability of DGRs to introduce targeted diversity into protein-coding DNA sequences has the potential to dramatically accelerate the evolution of adaptive traits. The utility of these elements in nature is underscored by their widespread distribution throughout the bacterial domain. Here we define DNA sequences and structures that are necessary and sufficient to direct the diversification machinery to specified target sequences. In addition to providing mechanistic insights into conserved features of DGR activity, our results provide a blueprint for the use of DGRs for a broad range of protein engineering applications.
| Diversity-generating retroelements (DGRs) have been identified in numerous bacterial phyla [1], [2]. Although most DGRs are bacterial chromosomal elements, they are prevalent in phage and plasmid genomes as well. The prototype DGR was identified in a temperate bacteriophage, BPP-1, on the basis of its ability to switch tropism for different receptor molecules on host Bordetella species [3]. Tropism switching is mediated by a phage-encoded DGR which introduces nucleotide substitutions in a gene that specifies a host cell-binding protein, Mtd (major tropism determinant), positioned at the distal tips of phage tail fibers. This allows phage adaptation to the dynamic changes in cell surface molecules that occur during the infectious cycle of its bacterial host [3]. Comparative bioinformatics predicts that all DGRs function by a fundamentally similar mechanism using conserved components ([1]; Gingery et al., unpublished data). These include unique reverse transcriptase (RT) genes (brt for BPP-1), accessory loci (avd or HRDC), short DNA repeats, and target genes that are specifically diversified [1]–[4].
As illustrated by the BPP-1 DGR shown in Figure 1A, diversity results from the introduction of nucleotide substitutions in a variable repeat (VR) located at the 3′ end of the mtd gene [1]–[4]. Variable sites in VR correspond to adenine residues in a homologous template repeat (TR), which remains unchanged throughout the process [1]–[4]. Transcription of TR provides an essential RNA intermediate that is reverse transcribed by Brt, creating a cDNA product which ultimately replaces the parental VR [4]. During this unidirectional retrotransposition process of mutagenic homing, TR adenines are converted to random nucleotides which subsequently appear at corresponding positions in VR [1]–[4]. Adenine mutagenesis appears to occur during cDNA synthesis and is likely to be an intrinsic property of the DGR-encoded RT [4].
Located at the 3′ end of VR is the IMH (initiation of mutagenic homing) region, which consists of at least two functional elements: a 14 bp GC-only sequence [(GC)14] which is identical to the corresponding segment of TR, and a 21 bp sequence containing 5 mismatches with TR that determines the directionality of information transfer [1]. Using a saturating co-conversion assay, we have precisely mapped a marker transition boundary that appears to represent the point at which 3′ cDNA integration occurs and information transfer begins [4]. This maps within the (GC)14 element and we previously postulated that it represents the site of a nick or double-strand break in the target DNA [4]. If true, the resulting 3′ hydroxyl could serve to prime reverse transcription of the TR-derived RNA intermediate in a target DNA-primed reverse transcription (TPRT) mechanism [4]–[7]. cDNA integration at the 5′ end of VR requires TR/VR homology and may occur via template switching during cDNA synthesis [4].
There are 23 adenines upstream of the (GC)14 element in the BPP-1 TR, each of which is capable of variation [3]. The theoretical maximum DNA sequence diversity is ∼1014, which translates to a maximum protein diversity of nearly 10 trillion distinct polypeptides at the C-terminus of Mtd. For Mtd and other DGR-diversified proteins, co-evolution has resulted in the precise positioning of TR adenines to correspond to solvent exposed residues in the ligand binding pockets of variable proteins [8], [9]. As implicated in Figure 1A, mutagenic homing occurs through a “copy and replace” mechanism that precisely regenerates all cis-acting components required for further rounds of diversification [4]. This allows the system to operate over and over again to optimize ligand-receptor interactions.
The goal of this study was to characterize requirements for target site recognition by the BPP-1 DGR. Along with insights into the mechanism of mutagenic homing, our results reveal engineering principles that allow DGRs to be exploited to diversify heterologous genes through a process that is entirely contained within bacterial cells.
5′ and 3′ boundaries of the BPP-1 DGR target sequence were delineated using a PCR-based assay that specifically detects VR sequences that have been modified by DGR-mediated retrohoming [4]. The system consists of a donor plasmid (pMX-ΔTR23-96, Figure 1B) carrying avd, a modified TR containing a 30 bp tag (TG2), and brt co-expressed from a BvgAS-regulated promoter [4], and a recipient prophage genome deleted for avd, TR, and brt (BPP-1ΔATR, Figure 1C). TR retrotransposition from the donor plasmid to the recipient prophage VR creates a “tagged” VR that can be detected using primer pairs specific for the tag and VR-flanking sequences (P1/P4 and P2/P3 in Figure 1B; Table 1). Controls include the demonstration that homing products are Brt-dependent and contain mutagenized adenines. An advantage of this assay is that it does not require infectious phage particle formation and consequently allows manipulation of sequences that are required for Mtd function.
Deletions were introduced into VR and adjacent sequences in BPP-1ΔATR lysogens (Figure 1C and Figure S1) and the abilities of mutated prophages to serve as recipients in retrohoming assays were measured (see Materials and Methods). As shown in Figure 1D, sequences upstream of VR were dispensable for DGR homing (lanes 4&13). A deletion mutation that truncates the first 20 bp of VR still supported homing, although at a decreased level (lanes 5&14). Sequence analysis of homing products for this mutant suggested that 5′ cDNA integration occurred at cryptic sites within the truncated VR, although 3′ cDNA integration occurred in a normal manner (Figure 1D, lanes 5&14; Figure S2). At the 3′ end, homing was highly dependent on a 35 bp region located downstream of VR (lanes 6&15 vs. lanes 7&16). This implicated sequences with 8 bp inverted repeats that could potentially form a hairpin structure in ssDNA or a cruciform structure in dsDNA as a possible determinant of DGR target function (Figure 1C). Additional analysis showed that deletion of sequences immediately downstream of the stem was well tolerated (3′Δ58, Figure 1C and 1E), while further deletions at the 3′ end (3′Δ68) reduced target function to essentially non-detectable levels in homing assays.
In the experiments in Figure 1, homing products were not detected using a donor plasmid expressing enzymatically inactive Brt (BrtSMAA, in which the active site motif YADD is replaced by SMAA; [1], [3], [4]), and sequence analysis of products generated with primer sets P1/P4 and P2/P3 demonstrated transfer of the TG2 tag from TR to VR. Adenine mutagenesis was observed in ∼53% of clones containing P1/P4 products and ∼32% of clones containing P2/P3 products (data not shown), which had 3 and 2 TR adenine residues available for mutagenesis, respectively. These observations indicated that true DGR homing products were being detected. Equivalent amounts of template phage DNA, as measured by quantitative PCR, were included in each experiment (lanes 19–27, Figure 1D; lanes 17–24, Figure 1E).
We next determined whether the primary sequence or the secondary structure of the putative hairpin/cruciform located downstream of VR is important for function. To disrupt the structure, 7 consecutive residues proximal to the loop on the 3′ half of the stem were changed to their complementary residues (StMut, Figure 2A). The resulting mutant was essentially unable to support DGR homing at a level that could be detected in PCR-based assays (lanes 3&9, Figure 2B). Complementary substitutions were subsequently introduced to the 5′ half of the stem to generate StRev (Figure 2A). If the primary sequence is important, the StRev recipient should remain non-functional. Alternatively, if the structure of the stem is the critical element, restoring base pairing interactions might restore DGR target function. As shown in Figure 2B (lanes 5&11), this appears to be the case, as the StRev mutant regained DGR homing activity. Homing products were verified by sequencing and adenine mutagenesis was observed (Figure S3).
Phage tropism switching assays provide a quantitative measure of DGR function [1], [3], [4]. Although the evolution of new ligand specificities is an inherently stochastic process, the frequency at which it occurs reflects the combined efficiencies of retrohoming and adenine mutagenesis. In Figure 2C, tropism switching was measured using BPP-1ΔATR or mutant derivatives complemented with plasmid pMX1, which provides avd, TR and brt in trans (see Materials and Methods). The StMut mutation resulted in over a 1000-fold decrease in tropism switching, which was restored to near WT levels by the StRev allele. Sequence analysis of VR regions in phages with switched tropisms (5 random clones each) confirmed adenine mutagenesis in every case (Figures S4, S5, S6).
Taken together, these data argue that the ability to form a hairpin or cruciform structure, as opposed to the primary sequence of the inverted repeats, is a critical determinant of target site recognition. The residual tropism switching activity of StMut phage suggests that hairpin/cruciform-independent pathways may exist, although they operate at a much lower efficiency.
To determine if the hairpin/cruciform structure can form in vitro, supercoiled plasmids carrying WT or mutant BPP-1 DGR target sequences were isolated and treated with phage T7 DNA endonuclease I, followed by primer extension with 5′ end-labeled primers to identify specific cleavage sites [10], [11]. T7 DNA endonuclease I is a structure-specific enzyme that resolves DNA four-way (Holiday) junctions and has previously been used to identify DNA hairpin or cruciform formation [10], [11]. As shown in Figure 3, cleavage sites were detected on both DNA strands in the hairpin/cruciform structure, with major cleavage sites at or near the four-way junction. Minor cleavage sites were also detected at or near the loop, as T7 DNA endonuclease I also has some activity on single-stranded DNA [12]. T7 endonuclease I cleavage at the hairpin/cruciform region requires structure formation, as plasmids containing a disrupted stem (StMut) were not cleaved in the corresponding region. Linearization of plasmids containing the WT sequence eliminated cleavage, suggesting that negative supercoiling is required for hairpin/cruciform formation [13], [14]. These results demonstrate that hairpins can form on either strand of the target DNA. Although it is likely that they form simultaneously on both strands to create cruciforms, this is not directly addressed by enzyme cleavage assays, hence the hairpin/cruciform designation.
We next determined whether the orientation of the target sequence relative to the phage genome is important for DGR retrohoming. In the experiment in Figure 4A, a segment of the BPP-1ΔATR prophage that includes VR and its flanking sequences was inverted, and PCR-based DGR homing assays were performed with donor plasmid pMX-ΔTR23-96. DGR homing into the inverted target occurred at a level comparable to that of the WT control (Figure 4B), and sequence analysis indicated that normal homing products were produced (Figure S7). These results show that the polarity of phage replication is not important for DGR homing, and that the hairpin/cruciform structure functions in a manner that is independent of its orientation relative to the leading or lagging strands formed during DNA replication.
Inverted repeats are nearly always found downstream of VR sequences in target genes [Gingery et al., unpublished data], as illustrated by the phage DGR sequences shown in Figure 5. These elements display a striking pattern of similarity, suggesting they have conserved and important functions. In each case, hairpin/cruciform structures with 7–10 bp GC-rich stems and 4 nt loops can potentially be formed. Although stems are always GC-rich, their sequences differ, while loops are more conserved with the consensus sequence (5′GRNA3′, with R = A or G, N = any nucleotide) in the sense strand. The exact distance between the hairpin/cruciform structures and the 3′ ends of their respective VRs appears to be quite flexible. We took advantage of the BPP-1 DGR system to test the relevance of these patterns of conservation, with the goal of generating a more comprehensive understanding of parameters important for target site recognition.
We first studied requirements for stem length and sequence and found that although minor changes are tolerated, the WT configuration appears to be optimized for BPP-1 DGR function. Of the stem length variants in Figure 6A, extensions are better tolerated than deletions. Removal of 2, 4 or 6 bp proximal to the loop results in markedly decreased activity in both PCR-based homing (Figure 6B) and phage tropism switching assays (Figure 6C), to levels similar to those observed with the StMut allele in which the stem is completely abolished (Figure 2A). Insertion of 2 bp next to the loop had little effect on activity, while longer insertions gradually decreased target site function. Keeping the length of the stem constant, a sequence change in the middle of the stem that converts 4 GC base pairs to AT base pairs (StAT, Figure 6A) greatly reduced, but did not eliminate function. We next tested the effects of altering the sequence and size of the loop using the mutant constructs shown in Figure 7A. Substituting CTTT for the consensus loop sequence GAAA, or increasing the size of the loop by as little as 2 nt, decreased activity in PCR-homing (Figure 7B) and tropism switching assays (Figure 7C) to near background levels. Based on these experiments, it appears that an 8–10 bp GC-rich stem is optimal for BPP-1 DGR homing, and that both the size and sequence of the 4 bp loop are critical for function. Our results correlate with the patterns of conservation shown in Figure 5.
In the experiments in Figure 8, we tested the effects of altering the position of the hairpin/cruciform with respect to the 3′ boundary of VR and probed sequence requirements for the intervening region. SpM4 (Figure 8A), in which the 4 residues in the spacer were switched to the complementary nucleotides, retained WT activity (Figure 8B). In contrast, deletion of the spacer (SpD4) resulted in a significant decrease in target function. The SpM4 and SpD4 mutations eliminate the mtd stop codon and generate non-infective phages, obviating the ability to measure tropism switching. Nonetheless, their relative levels of activity were readily apparent in PCR-homing assays. Expansion of the spacer was tolerated to a greater extent than deletion. SpI3, which has a 3 bp insertion in the spacer (Figure 8A), showed no significant defect in PCR-homing or phage tropism switching assays (Figure 8B and 8C), but longer insertions gradually decreased target site function.
The SpI6 insertion, which increases the distance between the hairpin/cruciform structure and the 3′ end of VR by 6 bp, retains a measurable level of activity. We took advantage of this and used a marker coconversion assay (Figure S8; [4]) to determine the relationship between the position of the hairpin/cruciform structure and the site at which information transfer initiates. As summarized in Figure 8D, our coconversion assay measured transfer of nucleotide polymorphisms from tagged TR donors to a recipient VR carrying the SpI6 mutation using PCR-based homing assays (data not shown). With the WT recipient, a coconversion boundary occurs between positions 107 and 112, and this was interpreted as representing the site at which TR-derived cDNA synthesis initiates [4]. As shown in Figure 8D, the coconversion boundary remains essentially unchanged in the SpI6 mutant. Although the position of the hairpin structure affects the efficiency of DGR homing, it does not determine the site at which cDNA is integrated at the 3′ end of VR.
To determine if the results presented here complete our understanding of DGR-encoded requirements for retrohoming to a target gene, we applied them as engineering principles in an attempt to construct a functional, synthetic, TR/VR system. For a DNA sequence to serve as a recipient VR, three conditions must be met. First, it must be adjacent to an IMH region with functional (GC)14 and 21 bp elements at its 3′ end [1], [4]. Second, the IMH region must be followed by inverted repeats capable of forming a hairpin/cruciform structure of appropriate size, composition and distance from IMH. And finally, sufficient VR/TR sequence homology must be provided to allow efficient upstream (5′) cDNA integration. In recent studies we have shown that although short stretches of nucleotide identity (≥8 bp) between the TR-derived cDNA and VR target sequences are sufficient to complete the homing reaction, homing efficiency is increased with longer (≥19 bp) stretches of homology [4]. With these parameters in mind, we tested our ability to engineer the BPP-1 DGR to target a heterologous reporter gene (aph3′Ia; [15]) which provides facile detection of targeting events by antibiotic selection.
The recipient VR-KanS cassette shown in Figure 9A contains an aph3′Ia kanamycin resistance (KanR) allele with a 3′ deletion that renders it nonfunctional by removing coding sequences for 6 essential C-terminal residues. The truncated gene was placed immediately upstream of IMH, followed by the hairpin/cruciform-forming inverted repeats from the BPP-1 DGR. Transcription is directed by the native aph3′Ia promoter. The donor plasmid expresses avd, brt, and one of two engineered TRs (TR-Km1, TR-Km2) from the Pfha promoter. Both TRs contain the intact 3′ end of the aph3′Ia open reading frame, followed by two consecutive stop codons and sequences 97–134 from the 3′ end of the BPP-1 TR. For TR-Km2, the aph3′Ia fragment is also flanked, at its 5′ end, by the first 22 residues of the BPP-1 TR. DGR-mediated retrotransposition from the donor TR constructs to the VR-KanS recipient should regenerate a full-length aph3′Ia gene conferring KanR.
We first tested whether targeting can occur in the context of a replicating phage. BPP-1ΔATR*KanS carries the VR-KanS cassette inserted between attL and bbp1 on the left arm of the prophage genome [16], along with a deletion of avd, TR and brt and a series of synonymous substitutions in IMH to inactivate the mtd VR (Figure 7A). B. bronchiseptica RB50 carrying the TR-Km1 or -Km2 donor plasmid, or derivatives with a null mutation in brt, were infected with BPP-1ΔATR*KanS and targeting efficiencies were determined by infecting RB50 with progeny phages and measuring relative numbers of KanR lysogens. KanR lysogens were readily detected when targeting occurred from Brt+ TR donors, but not Brt− donors (Figure 9B), and sequence analysis showed that KanR resulted from the regeneration of full-length aph3′Ia alleles which often contained mutations at positions corresponding to adenines in donor TRs (Figures S9 and S10). It is interesting to note that the TR-Km1 donor was significantly more efficient than TR-Km2. This suggests that the majority of cDNAs are extended to the 5′ termini of these short synthetic TRs, and target (VR) homology to the extreme 3′ ends of the extension products may be advantageous for cDNA integration.
We also tested the ability to target the VR-KanS cassette when present on a resident prophage in the bacterial chromosome or on a plasmid. In the experiment in Figure 9C, RB50/BPP-1ΔATR*KanS lysogens were transformed with donor plasmids under conditions that suppress Pfha promoter activity. Following a 6 hr pulse of Pfha induction, cells were plated under promoter-suppressing conditions on media with or without kanamycin. In Figure 9D, a similar protocol was used to target a VR-KanS cassette carried on a medium copy number plasmid in RB50 cells containing a TR donor plasmid, but no other phage sequences. In both experiments, KanR colonies were readily detected when targeting occurred from Brt+, but not Brt− TR donors, and sequence analysis showed characteristic patterns of adenine mutagenesis (Figures S11, S12, S13, S14). Taken together, our results demonstrate the ability to engineer a VR/TR system that targets a heterologous reporter gene on a phage, plasmid or bacterial genome. The data in Figure 9D show that no BPP-1 phage products, other than those encoded in the DGR, are required for mutagenic retrohoming.
Understanding DGR target site recognition requires a precise definition of cis-acting sequences important for retrohoming. Our analysis of the boundaries of the BPP-1 DGR target showed that sequences upstream of VR are dispensable, as predicted by previous results [4]. More importantly, we show that homing is facilitated by an element downstream of VR, beyond the point at which TR/VR homology ends. Sequence analysis, mutagenesis, and structure-specific nuclease assays demonstrated that GC-rich inverted repeats directly following VR form a hairpin/cruciform structure that plays a critical role in retrohoming. Highly similar elements are present in analogous locations in many phage- or prophage-related DGRs (Figure 5), and hairpin/cruciform structures are predicted for the majority of DGRs that naturally reside on bacterial chromosomes and plasmids as well [Gingery et al., unpublished data]. We propose that DNA hairpin formation near the 3′ end of VR is a conserved requirement for DGR-mediated retrohoming.
For the BPP-1 DGR target, the 8 bp stem appears to function as a structure that is dependent on nucleotide composition but not sequence. In contrast, the loop of the hairpin/cruciform structure is constrained in size and sequence and conforms to the consensus, 5′-GRNA, derived from comparisons with other phage-related DGRs. This suggests that loop sequence and size may be important for stabilizing the hairpin/cruciform structure [17], or for creating a strand bias in DNA cleavage by a host-encoded endonuclease. It is also possible that the loop is in direct physical contact with a critical component, such as Brt, Avd, a TR-containing RNA transcript, or other parts of the DGR target. By testing the effects of length and sequence variations between the hairpin/cruciform and VR, we found that distance is an important parameter, although some flexibility exists. Extending the spacer by 6 bp did not shift the marker coconversion boundary in the (GC)14 region during DGR homing [4], showing that the position of the hairpin/cruciform does not determine the site at which 3′ cDNA integration occurs.
DGRs are evolutionarily related to group II introns [1] and it is interesting to note that a subset of these retroelements, the group IIC introns, also target motifs with stem-loop structures [18]–[20]. In nature, group IIC introns are often found to be located short distances downstream of sequences encoding known or predicted factor-independent transcription terminators, which are composed of GC-rich stems with loops of varying sizes followed by poly-uridine stretches [18]–[20]. Using an in vitro mobility assay, Robart et al. [19] have shown that reconstituted ribonucleoprotein particles from the Bacillus halodurans B.h.I1 group IIC intron recognize structures in ssDNA that correspond to RNA hairpins formed during transcription termination. As observed with the BPP-1 DGR, the B.h.I1 mobility reaction was highly dependent on stem formation but not absolute sequence [19]. Stems shorter than 9 bp had significantly reduced activities in in vitro mobility assays, a longer stem (14 bp) retained function, and the efficiency of targeting correlated with GC content and predicted stem stability [19]. In contrast to our observations with the BPP-1 DGR, alterations in loop sequence had little effect on B.h.I1 mobility in vitro [19]. The adaptation of group IIC introns to recognize and insert downstream of factor-independent transcriptional terminators was proposed to provide a selective advantage by limiting their expression, avoiding the interruption of essential coding sequences, and facilitating horizontal spread as intrinisic terminators are common and conserved in bacteria [19]. For DGRs, we speculate that the ability to target sequences upstream of terminator-like stem-loop structures may have played a role in directing their sequence diversification capabilities to the 3′ coding regions of target genes.
The TPRT model for DGR homing postulates that cDNA synthesis initiates with a nick or double-strand break in the IMH (GC)14 sequence, providing a primer for reverse transcription of a TR-containing RNA transcript [4]. Analogous to target recognition by group IIC introns, the hairpin/cruciform structure may serve as a recognition element for a retrohoming complex that includes trans-acting DGR-encoded factors. A DNA endonuclease that might be responsible for cleavage awaits identification, and possibilities include Avd, Brt, a TR-derived catalytic RNA, or an unidentified host factor. It is also possible that the DNA hairpin/cruciform actively promotes single- or double-strand breaks. If DNA repair synthesis extends to the (GC)14 region, the elongating antisense strand could then be used for cDNA priming. DNA breaks at the hairpin/cruciform structure could be created by an endonuclease that cleaves the single-stranded loop, or by a structure-specific enzyme similar to T7 endonuclease I [21]. Since DNA cruciforms are structurally similar to Holiday junctions, host-encoded recombination proteins that function in resolving recombination intermediates could be involved [22]. The cDNA priming mechanism of the BPP-1 DGR appears to be different from that of mobile group II introns that lack a DNA endonuclease activity in their intron-encoded proteins [23]–[25]. Reverse transcription in retrohoming and ectopic transposition of these elements is proposed to be primed by either the leading or lagging strand during DNA replication, and strong strand-specific biases are observed [23]–[25]. Our observation that the BPP-1 DGR target sequence is orientation-independent suggests that DNA replication polarity does not play a significant role in cDNA priming. Although our results to date are consistent with TPRT, further studies are required to definitively characterize the mechanism of cDNA initiation and integration at the 3′ end of VR and to determine the precise role of the hairpin/cruciform structure in the retrohoming process.
The broad distribution of DGRs in nature attests to their utility, and prospects for adapting these elements for protein engineering applications are compelling. Our results demonstrate that the region containing the (GC)14 and 21 bp sequences in IMH, and an adjacent hairpin/cruciform, is sufficient to direct the DGR mutagenic homing machinery to a heterologous target gene through appropriate engineering of a cognate TR. Using similar design principles we have successfully targeted a tetracycline resistance determinant as well (HG and JFM, unpublished data). For DGRs to be useful tools, it will be necessary to engineer their activity to allow efficient and controlled diversification. Having defined the DGR-encoded cis- and trans-acting factors required to diversify heterologous sequences, efforts to optimize their activities can now proceed in an informed and comprehensive way. It will also be important to determine the effects of TR/VR size, composition, and position relative to cis-acting DGR elements, on the efficiency of diversifying heterologous sequences. In preliminary experiments, insertions of moderate size (up to ∼200 bp) at position 84 in the BPP-1 TR (134 bp) are transferred to VR and mutagenized at adenines, suggesting that sequences of >300 bp could be diversified by an engineered system (LVT, HG and JFM, unpublished data).
In addition to providing prodigious levels of diversity, mutagenic homing is a regenerative process that allows DGRs to operate through unlimited rounds to optimize variable protein functions [4]. This may be particularly advantageous for directed protein evolution since desired traits can be selected and continuously evolved in iterative cycles, without the need for library construction or other interventions, through a process that takes place entirely within bacterial cells.
B. bronchiseptica strains RB50, RB53Cm, RB54 and ML6401 have been described [16]. The BPP-1ΔATR lysogen was constructed from ML6401, an RB50 strain lysogenized with phage BPP-1, by deleting sequences from avd position 48 to position 882 of brt. Target region deletions/insertions and hairpin/cruciform modifications were introduced into the BPP-1ΔATR lysogen through allelic exchange [1], [4] and are diagramed in the figures. The BPP-1ΔATR* lysogen contains multiple silent mutations at both the 5′ and 3′ ends of VR to inactivate it as a DGR target. It was used as the parental strain to create the BPP-1ΔATR*KanS lysogen, in which the KanR gene aph3′Ia has sequences encoding the C-terminal 6 amino acid residues truncated and is placed upstream of IMH and the hairpin/cruciform structure as a reporter for heterologous gene targeting. The aph3′Ia allele also contains an AAA to CGC substitution resulting in K260R. The VR-KanS reporter cassette was inserted between attL and bbp1 of the phage genome. Phage BPP-1ΔATR and its various derivatives were produced from the above lysogens.
Plasmid pMX-ΔTR23–96 has TR positions 23–96 deleted and replaced by a 30 bp PCR tag as in pMX-ΔTR23–84 [4]. Its RT-deficient derivative contains the YMDD to SMAA mutation at Brt positions 213–216 [3], [4]. Plasmids pMX1 and pMX1SMAA were used for phage tropism switching assays and have previously been described [4].
pUC-StWT is a pUC18-based plasmid containing the WT BPP-1 DGR target from position −6 upstream of VR to position +82 downstream of VR. pUC-StMut is its derivative with 7 residues in the 3′ half of the stem, proximal to the loop, mutated to their complementary nucleotides.
Plasmids pMX-TRC85T, pMX-TRC91T, pMX-TRC97T, pMX-TRC100T, pMX-TRC105T, pMX-TRC107T, pMX-TRC109T, pMX-TRC112T, pMX-TRC115T, pMX-TRC120T and pMX-TRC125T have been previously described [4].
Plasmids pMX-Km1 and pMX-Km2 were constructed from pMX-ΔTR23–96 for KanR gene targeting, both containing the last 36 bp of aph3′Ia. The 36 bp sequence and its following two stop codons replace TR positions 1–96 in pMX-Km1 and TR positions 23–96 in pMX-Km2.
Plasmid pHGT-KanS contains the VR-KanS cassette described above and was used as the recipient plasmid for KanR targeting. The plasmid also carries a tetracycline resistance gene.
Phage production for DGR functional assays was carried out by either single-cycle lytic infection or mitomycin C induction from lysogens as previously described [4], except for minor modifications as noted. For single-cycle lytic infection, B. bronchiseptica RB50 cells transformed with appropriate donor plasmids were grown overnight at 37°C in Luria-Bertani (LB) media containing 25 µg/ml of chloramphenicol (Cam), 20 µg/ml streptomycin (Str), and 10 mM nicotinic acid to modulate to the Bvg− phase and prevent transcription from the Pfha promoter. An amount of cells equal to 1 ml of culture (OD600 = 1.0) was pelleted, rinsed, and resuspended in 2.5 ml Stainer Scholte (SS) medium [26] containing 25 µg/ml Cam and 20 µg/ml Str (SS+Cam+Str). Cultures were grown for 3 hr at 37°C to modulate bacteria to the Bvg+ phase and activate Pfha promoter expression. An aliquot of 500 µl from each culture was used for OD600 measurement and cell number calculation. Phage particles were added to the rest of the culture at a multiplicity of infection of ∼2.0. Following 1 hr incubation at 37°C for phage absorption, infected cells were pelleted and resuspended in 1 ml of fresh, prewarmed SS+Cam+Str media and incubated at 37°C for 3 hr post phage addition to allow completion of a single cycle of phage development. Progeny phages were harvested following chloroform extraction.
For phage production from lysogens, RB50 derivatives carrying appropriate prophages and donor plasmids were grown and modulated to the Bvg+ phase as in single-cycle lytic infections. Phage production was induced with 0.2 µg/ml mitomycin C for 3 hr at 37°C. Progeny phages were harvested by chloroform extraction.
Phage tropism switching and DGR homing assays have been previously described [4].
Plasmids containing the WT BPP-1 DGR target and the StMut mutation were isolated from E. coli DH5αλpir cells using the QIAprep Spin miniprep kit (Qiagen). Plasmids were linearized by digestion with BglI as indicated. To analyze hairpin/cruciform structure formation in supercoiled or relaxed DNAs, 0.5 µg of supercoiled or linearized plasmids were treated with 10 units of T7 DNA endonuclease I (New England Biolabs, Ipswich, MA) for 40 minutes as in Miller et al. [11]. The reactions were terminated by phenol-chloroform-isoamyl alcohol (25∶24∶1) extraction and DNAs were precipitated with ethanol. T7 DNA endonuclease I cleavage sites were determined by primer extension with 5′-end 32P-labeled primers using Vent (exo-) DNA polymerase (New England Biolabs, Ipswich, MA) as in Miller et al. [11], except that 5% DMSO was added for GC-rich templates. Primer extension products were resolved on 6% polyacrylamide/8 M urea gels, alongside Sanger sequencing ladders generated with the same labeled primers and a plasmid template containing the WT target.
To target the KanR gene on a replicating phage, BPP-1ΔATR*KanS phage particles were used for single-cycle lytic infection of RB50 cells transformed with appropriate donor plasmids, similar to phage production by single-cycle lytic infection described above. Progeny phages were titered and ∼1011 pfu of different phages were added to 25 ml RB50 cells (OD600 = 1.2) in SS+Str media for 8.0 hr to reestablish lysogens. Cells were pelleted and resuspended in 5 ml LB and serial dilutions were plated on LB+NA+Str and LB+NA+Str+Kan (50 µg/ml) to determine KanR gene targeting frequencies. Lysogen reestablishment efficiencies ranged from 60% to 100% based on PCR analysis of 10 colonies each picked on LB+NA+Str plates using phage specific primers. KanR targeting efficiency for each donor plasmid was determined as the ratio of colony forming units (cfu) on LB+NA+Str+Kan plates to those on LB+NA+Str, calibrated with the lysogen reestablishment efficiency for that sample.
To target the KanR gene on a prophage in the bacterial chromosome, RB50 cells lysogenized with phage BPP-1ΔATR*KanS were transformed with appropriate donor plasmids. Starting cultures were grown overnight in LB+NA+Str+Cam as described above. An amount of cells equal to 1 ml of culture (OD600 = 1.0) was pelleted, rinsed, and resuspended in 2.5 ml SS+Cam+Str and grown at 37°C for 6 hours. Serial dilutions were plated on LB+NA+Str and LB+NA+Str+Kan (50 µg/ml) to determine KanR gene targeting frequencies. KanR targeting efficiencies were determined as relative numbers of KanR cells as above. To target the KanR gene on a plasmid, the recipient plasmid pHGT-KanS and appropriate donors were transformed into RB50 cells and analyzed similarly. Tetracycline was added to 5.0 µg/ml for recipient plasmid maintenance.
|
10.1371/journal.pgen.1006291 | Multisite Phosphorylation of NuMA-Related LIN-5 Controls Mitotic Spindle Positioning in C. elegans | During cell division, the mitotic spindle segregates replicated chromosomes to opposite poles of the cell, while the position of the spindle determines the plane of cleavage. Spindle positioning and chromosome segregation depend on pulling forces on microtubules extending from the centrosomes to the cell cortex. Critical in pulling force generation is the cortical anchoring of cytoplasmic dynein by a conserved ternary complex of Gα, GPR-1/2, and LIN-5 proteins in C. elegans (Gα–LGN–NuMA in mammals). Previously, we showed that the polarity kinase PKC-3 phosphorylates LIN-5 to control spindle positioning in early C. elegans embryos. Here, we investigate whether additional LIN-5 phosphorylations regulate cortical pulling forces, making use of targeted alteration of in vivo phosphorylated residues by CRISPR/Cas9-mediated genetic engineering. Four distinct in vivo phosphorylated LIN-5 residues were found to have critical functions in spindle positioning. Two of these residues form part of a 30 amino acid binding site for GPR-1, which we identified by reverse two-hybrid screening. We provide evidence for a dual-kinase mechanism, involving GSK3 phosphorylation of S659 followed by phosphorylation of S662 by casein kinase 1. These LIN-5 phosphorylations promote LIN-5–GPR-1/2 interaction and contribute to cortical pulling forces. The other two critical residues, T168 and T181, form part of a cyclin-dependent kinase consensus site and are phosphorylated by CDK1-cyclin B in vitro. We applied a novel strategy to characterize early embryonic defects in lethal T168,T181 knockin substitution mutants, and provide evidence for sequential LIN-5 N-terminal phosphorylation and dephosphorylation in dynein recruitment. Our data support that phosphorylation of multiple LIN-5 domains by different kinases contributes to a mechanism for spatiotemporal control of spindle positioning and chromosome segregation.
| Protein kinases control biological processes by phosphorylating specific amino acids of substrate proteins. It remains a major challenge to identify which phosphorylation events are critical in vivo and how phosphorylation affects protein function. Recent developments in CRISPR/Cas9-based genetic engineering make it possible to substitute individual amino acids, which allows investigating the role of single and multi-site phosphorylation of substrates in vivo. Here, we focus on an intensively phosphorylated cell division protein, LIN-5NuMA. C. elegans LIN-5 participates in chromosome segregation and is essential for positioning the spindle and cell cleavage plane during asymmetric cell division. Previously, we demonstrated that the polarity kinase PKC-3 phosphorylates LIN-5 to inhibit its function. Our current analysis reveals four additional phosphorylated residues that are critical for LIN-5 function. Two of these residues contribute to the interaction of LIN-5 with its binding partner GPR-1/2, whereas the other two residues are critical in dynein motor recruitment by LIN-5. Together, our results reveal that multisite phosphorylation of LIN-5 is essential to ensure proper chromosome segregation and spindle positioning.
| Animal development and tissue homeostasis depend critically on cell divisions that create cells with specific shapes and functions, in the right numbers and at the proper positions. The spindle apparatus plays a central role in the cell division process, as it segregates the chromosomes in mitosis and determines the plane of cell cleavage during cytokinesis [1–3]. Placement of the spindle in the cell center during division results in the formation of daughter cells of equal size, whereas off-center migration and spindle rotation allows the creation of differently sized daughter cells at specific locations. Moreover, the plane of cell cleavage determines whether polarized cells undergo symmetric or asymmetric cell division. Asymmetric cell divisions create cell diversity and allow maintenance of tissue-specific stem cells, by combining self-renewal with the generation of differentiating daughter cells (Reviews: [4,5]). Thus, tight control of the spindle function and position is needed to coordinate chromosome segregation with cleavage plane determination, which is essential for genetic stability, tissue integrity and stem cell maintenance in a wide variety of evolutionary contexts.
Pioneering studies in Caenorhabditis elegans and Drosophila melanogaster revealed that the position of the spindle responds to polarity cues during asymmetric cell division [1,2,4,5]. In C. elegans, anterior-posterior (A-P) polarity is established after fertilization of the oocyte. This involves re-distribution of specific partitioning-defective (PAR) proteins into two opposing domains of the cell cortex. The PDZ-domain proteins PAR-3 and PAR-6 form a complex with the PKC-3 aPKC polarity kinase and become restricted to the anterior half of the zygote, while the PAR-2 ring-finger protein and PAR-1 kinase occupy the posterior domain [6]. This A-P polarity guides the asymmetric localization of cytoplasmic determinants as well as the position of the mitotic spindle. During the first mitotic division, the spindle is positioned off-center, to instruct an asymmetric cell division that creates a larger anterior blastomere (AB) and smaller germline precursor cell (P1). Next, the spindle rotates by 90 degrees in P1, to instruct another asymmetric division with a cleavage plane perpendicular to the one of AB. These early divisions of the C. elegans embryo have served as an important model for studies of the coordinated regulation of cell polarity, fate determinant localization, and spindle positioning during asymmetric cell division.
In addition, studies in C. elegans and Drosophila uncovered an evolutionarily conserved protein complex that mediates spindle positioning. This complex consists of the alpha subunit of a heterotrimeric G protein in association with the TPR/GoLoco protein GPR-1/2 and coiled-coil protein LIN-5 in C. elegans (Gα–Pins–Mud in Drosophila, Gα–LGN–NuMA in mammals) (Reviews: [1–6]). The GPR-1/2 GoLoco motifs interact with Gα-GDP [7], while the tetratricopeptide repeats (TPR) associate with the C-terminus of LIN-5 (Fig 1A). The ternary protein complex acts at the cell cortex in conjunction with cytoplasmic dynein and microtubule plus ends to generate microtubule pulling forces that promote chromosome segregation and position the spindle [8–12]. Based on results obtained for NuMA, an extended N-terminal domain of LIN-5 likely mediates interaction with the dynein motor complex [9]. It remains unclear how Gα–GPR-1/2–LIN-5 engages dynein and microtubule depolymerization in the generation of cortical pulling forces, and how pulling forces are temporally and spatially restricted.
Asymmetric positioning and rotation of the spindle result from imbalance in the pulling forces. It has long been known that the cortical polarity of the C. elegans zygote is fundamental for the spatial organization of pulling forces, creating a higher net force in the posterior than the anterior, which causes the spindle to move off center [13,14]. This is in part achieved through PKC-3 mediated phosphorylation of LIN-5, which inhibits anteriorly directed pulling forces [15]. Phosphorylation also appears to regulate cortical pulling forces in other systems. For example, phosphorylation by aPKC inhibits Pins/LGN localization to the apical cell membrane and promotes planar cell division of MDCK canine kidney cells during cyst formation [16]. Moreover, phosphorylation of NuMA by PLK1 and CDK1 has been implicated in the timing of chromosome segregation and positioning of the mitotic spindle in human cells [17,18]. In addition to spindle positioning, the Gα–GPR-1/2–LIN-5 complex is essential for chromosome segregation, in all cell divisions except for the first few embryonic divisions in C. elegans [19–21]. Phosphorylation is likely to play a key role in coordinating chromosome segregation and spindle positioning through spatiotemporal regulation of Gα–GPR-1/2–LIN-5 function.
Our previous studies identified extensive in vivo phosphorylation of LIN-5 in C. elegans embryos [15]. The function of the majority of these phosphorylations remained unknown. Here we apply a combination of techniques to determine which phosphorylations are critical for LIN-5 function. CRISPR/Cas9-mediated genetic engineering allowed us to introduce single codon alterations in the C. elegans genome, and to compare non-phosphorylatable and potentially phosphomimetic LIN-5 mutants. In addition to PKC-3, we found that the PAR-1 polarity kinase likely phosphorylates LIN-5 in vivo, but physiological consequences of this phosphorylation were not detected. Alanine substitution mutagenesis of lin-5 transgenes pointed to four phosphorylated residues with critical functional contributions. Two of these residues form part of a 30 amino-acid domain of LIN-5 required for binding GPR-1/2. Phosphorylation of these residues promotes cortical pulling forces and GPR-1/2 localization in vivo, and appears to occur sequentially by GSK3 and casein kinase 1 (CK1). Moreover, we identified essential residues in the LIN-5 N-terminus that are phosphorylated by CDK1. Our data from extensive knockin replacement mutants are consistent with a mechanism involving sequential phosphorylation and dephosphorylation of the LIN-5 N-terminus in dynein recruitment to the meiotic spindle and cell cortex. Thus, a combination of phosphorylations by cell-cycle and polarity associated kinases likely underlies the spatiotemporal control of pulling forces in chromosome segregation and asymmetric cell division.
Previously, we described that at least 25 residues of LIN-5 are phosphorylated in vivo (Fig 1A) [15]. To acquire insight in which phosphorylations are functionally relevant, we replaced each phosphorylated serine or threonine with an alanine residue that cannot be phosphorylated. The relevant codon alterations were introduced in a cloned genomic lin-5 DNA fragment and subsequently tested for functionally complementing the lin-5(e1348) null mutation in vivo [19]. In the presence of maternal product, lin-5(e1348) mutants fail to undergo chromosome segregation during postembryonic divisions and continue abortive mitoses [19–21]. Transgenes containing wild type lin-5 or gfp::lin-5 coding sequences restored post-embryonic cell divisions in lin-5(e1348) null mutants (Fig 1B). However, these lin-5 transgenes appeared susceptible to germline and somatic silencing, as reliable rescue and GFP-LIN-5 expression was observed only in the F1 generation. Hence, we examined transgenic F1 animals, focusing on vulval development and nuclear divisions in the intestine as a quantitative measure for LIN-5 function (Fig 1B).
Alanine substitutions of threonine 168, serine 659, and serine 662 were the only single amino acid changes that significantly compromised LIN-5 function in vivo. The T168A mutation had the strongest effect and almost completely eliminated the ability to restore intestinal divisions in lin-5(e1348) null mutants (Fig 1B). Interestingly, this strong effect was specific for the intestine: LIN-5T168A expression allowed lin-5 mutants to develop a normal vulva (S1 Fig). T168 forms part of an ideal consensus phosphorylation site (S/T*-P-x-K/R) for the mitotic cyclin-dependent kinase 1 (CDK-1) [22]. CDK-1 is likely to regulate LIN-5, as multiple CDK-1 consensus sites are present in the LIN-5 N- and C-terminus, and CDK1 phosphoregulation of the NuMA C-terminus has been reported [18,23]. We generated double alanine substitutions of T168 in combination with T181 or S199, two nearby candidate residues for CDK-1 phosphorylation. Strikingly, the transgene encoding LIN-5[T168A,T181A], but not LIN-5[T168A,S199A], completely failed to rescue intestinal mitoses and vulva formation in lin-5(e1348) mutants (Fig 1B and S1 Fig). Because phosphorylation of T181 by itself was not essential for post-embryonic divisions, T168 and T181 phosphorylations likely cooperate to control LIN-5 function.
The individual and combined S659A and S662A substitutions (LIN-5[S659A,S662A]) also reduced lin-5(e1348) complementation. By contrast, simultaneous alanine substitutions of serines 729, 734, 737, and 739 did not prevent LIN-5 function (Fig 1B and S1 Fig). In agreement with the latter result, PKC-3 (aPKC) phosphorylation of these residues inhibits LIN-5 function and is not required for cell division [15]. Our alanine-substitution experiments indicate that in addition to spatiotemporal regulation of LIN-5 by PKC-3, phosphorylation of LIN-5 residues in the dynein-interacting N-terminus and GPR-1/2 binding C-terminus may contribute to LIN-5 regulation in vivo.
To determine whether CDK1 is indeed able to phosphorylate T168 and T181 of LIN-5, we performed in vitro kinase assays with recombinant GST-LIN-5 expressed in E. coli as a substrate. Indeed, immunopurified human CDK1/cyclin B phosphorylated GST-LIN-5, but not GST alone (S2A and S2B Fig). Analysis of in vitro phosphorylated GST-LIN-5 by mass spectrometry revealed extensive phosphorylation of T168, T181, and S744 of LIN-5 (S2B Fig). Additionally, peptides containing phosphorylated T704 and S756 were also found, and some other phosphopeptides less frequently. Taken together, CDK1/cyclin B phosphorylates LIN-5 in vitro at multiple sites including T168 and T181, and phosphorylation of T168 and T181 in vivo appears to be required for LIN-5 function.
In contrast to T168 and T181, residues S659 and S662 are not part of apparent consensus phosphorylation sites. In our previous in vivo mass-spectrometry data, the S659,S662 double phosphorylated peptides were abundant, while the corresponding unphosphorylated peptides were not detected [15]. This may indicate that S659 and S662 are constitutively phosphorylated in early embryos. To gain insight in which kinases may be involved, we examined LIN-5 phosphorylation in vitro with a series of polarity and cell cycle kinases, followed by mass spectrometry analyses. This revealed several residues that were phosphorylated by multiple kinases in vitro (Fig 2A). In striking contrast, S659 was only phosphorylated by GSK3, and none of the tested kinases phosphorylated S662 (Fig 2A). We considered several potential explanations for this lack of phosphorylation: the responsible kinase(s) may not have been included in the assays, residue S662 may not be accessible in the recombinant protein, or S662 phosphorylation may require a priming event. To test the latter possibility, we performed in vitro kinase assays with synthetic peptides that contain the S659 and S662 residues, either unphosphorylated or phosphorylated at one of the positions. Testing several kinases, we found that casein kinase 1 (CK1) efficiently phosphorylates S662, but only when the peptide contained a phosphorylated S659 residue (Fig 2B). As for the full length protein, only GSK3 phosphorylated S659 in the unphosphorylated peptide. Based on the combined in vitro data, we propose that GSK3 phosphorylation of residue S659 is a priming reaction for CK1 phosphorylation of S662. Highly similar phosphorylation has been reported for the Wnt/Frizzled co-receptor LRP6, with GSK3 priming for CK1 phosphorylation at similar sites [24,25].
In addition to CDK1, GSK3 and CK1 phosphorylation, our analyses revealed phosphorylation of LIN-5 by the polarity kinase PAR-1. While several phosphopeptides were detected, some were rare and the quantitative software program MaxQuant only recognized the S397 and S739 LIN-5 residues as in vitro phosphorylated by PAR-1 (Fig 2A). S397 is located in the LIN-5 coiled coil region and its phosphorylation was previously observed in embryos (Fig 2A) [15]. However, our previous in vivo analysis failed to identify LIN-5 phosphorylations that were diminished after par-1 RNAi [15]. Re-evaluation of the quantitative mass spectrometry data revealed that, although masked by an abundant unrelated peptide, the ratio between the phosphorylated and unphosphorylated S397 peptide was severely reduced in par-1(RNAi) embryos compared to control RNAi embryos (S3 Fig). In contrast, S739 phosphorylation was not significantly affected by par-1 knockdown in vivo [15]. Taken together, we identified multiple phosphorylated LIN-5 residues as well as candidate kinases that could be important in the regulation of LIN-5 function. In addition to four adjoining residues phosphorylated by PKC-3 in the C-terminus, T168 and T181 may be phosphorylated by CDK-1, S397 by PAR-1, and S659 by GSK-3, to prime phosphorylation of S662 by CK1.
Both S659,S662 and the four residues phosphorylated by PKC-3 are located in the LIN-5 C-terminus which mediates GPR-1/2 binding [15,26]. As phosphorylation could affect GPR-1/2 association, we wanted to define which LIN-5 residues are critical for GPR-1/2 binding. Testing deletion constructs in yeast two-hybrid assays confirmed that the LIN-5 C-terminal region is sufficient for GPR-1 association. GPR-1 interaction was observed for all truncated LIN-5 proteins except for those with deletions in the 609–671 amino acid region (Fig 3A). At the same time, including only the 609–671 LIN-5 fragment did not allow growth in this assay, possibly due to an inability of this short fragment to fold properly in yeast (Fig 3A). The essential 609–671 region does not contain serine 729, 734, 737, and 739 phosphorylated by PKC-3 in vivo, in agreement with our previous conclusion that PKC-3 phosphorylation of LIN-5 does not prevent interaction with GPR-1/2 [15].
To identify specific LIN-5 amino acids required for GPR-1/2 interaction, we performed “reverse yeast two-hybrid screening”. This method selects mutations that disrupt bait-prey protein interactions, making use of URA3-mediated conversion of 5-fluoroorotic acid (5-FOA) to a toxic product [27]. The normal interaction between LIN-5 and GPR-1 leads to GAL4-controlled URA3 expression in yeast two-hybrid assays, and causes cell death in the presence of 5-FOA. Thus, following mutagenesis of one of the binding partners, interaction-deficient alleles can be recovered from 5-FOA-resistant colonies [28]. We used PCR-based random mutagenesis of LIN-5 prey fragments (amino acids 609–821), and isolated 163 5-FOA resistant yeast colonies in a reverse yeast two-hybrid screen (for details see Materials and Methods, S4A Fig). 89 colonies contained a single missense mutation in the LIN-5 coding sequences, together changing 15 different amino acids. Substitutions of 12 of these 15 individual amino acids caused loss of GPR-1 interaction again in the re-test (Fig 3B and 3C). The 12 affected residues were all located between amino acids 638–667 of LIN-5. Importantly, the interaction-defective alleles included missense mutations of the phosphorylated residues S659 and S662. In fact, S662 was found altered to glycine, cysteine and asparagine (S4B Fig). These data indicate that a 30 amino acid stretch in the LIN-5 C-terminal region, which includes the in vivo phosphorylated S659 and S662 residues, mediates the interaction with GPR-1.
Following up on the interaction defective alleles, we noticed that the effect of missense mutations was substantially reduced when tested in the context of full length LIN-5, compared to the C-terminus only. Western blot analysis did not reveal substantial differences in protein levels compared to wild type (S5A Fig). The LIN-5 coiled-coil region promotes dimerization and is thereby expected to increase GPR-1 binding avidity. Only one of the four most frequently identified mutations, L663S, also interfered with full length LIN-5 binding to GPR-1 (Fig 3D, left panel). However, at a reduced temperature (20°C), this leucine 663 to serine (LIN-5[L663S]) mutation still allowed growth on selective media, indicating that GPR-1 interaction is not completely abolished. We also tested S659 and S662 phosphorylation-site mutants in the context of full length LIN-5. While the single mutations had little effect on GPR-1 binding, replacement of both serine 659 and 662 by alanine reduced GPR-1 interaction in yeast, as detected by lack of growth on -His plates at 30°C (Fig 3D, right panel). Phosphomimetic substitutions (S to D or E) of S659, S662, or both, did not reduce interaction (Fig 3D, right panel). These results are consistent with phosphorylation of S659 and S662 contributing to GPR-1/2 binding, and taking place in yeast as well as C. elegans. Taken together, our forward and reverse yeast two-hybrid assays identified LIN-5 residues that appear to mediate interaction with GPR-1/2, which are located within a 30 amino acid C-terminal domain. This includes S659 and S662, of which the phosphorylation in vivo likely contributes to GPR-1/2 binding.
We used CRISPR/Cas9-mediated gene targeting to engineer lin-5 alleles and examine the effects of amino acid substitutions in vivo [29–32]. First, we created the lin-5[L663S] mutation by introducing a single nucleotide alteration in the endogenous lin-5 locus. This resulted in a typical lin-5 loss-of-function phenotype, with homozygous sterile, thin and uncoordinated larvae that fail to undergo chromosome segregation but continue abortive mitoses [19,20]. We determined the number of nuclei in the intestine and ventral cord, following fixation and staining of DNA. In lin-5[L663S] mutants, both tissues contained severely reduced numbers of nuclei compared to the wild type, consistent with a failure to undergo chromosome segregation in most post-embryonic divisions (Fig 4A and 4B, S6A Fig). Thus, a single change of amino acid L663 in the GPR-1-binding motif of LIN-5 results in strong loss-of-lin-5 function. This result confirms the power of reverse yeast two-hybrid screening in identifying amino acids that affect protein-protein interactions in vivo [27].
Next, we used genome engineering to alter the in vivo phosphorylated residues T168, T181, S397, S659 and S662. For each residue, we created a non-phosphorylatable alanine substitution allele, as well as one or more potentially phosphomimetic alleles that contain aspartic acid or glutamic acid at the relevant positions. Alteration of the PAR-1 phosphorylated S397 residue had no apparent effect. Homozygous S397A and S397E animals were viable and showed normal development. Even close examination of LIN-5-mediated processes did not reveal abnormalities (See below; Fig 4C and 4D, S6A and S6B Fig and S1 Video). Thus, although this phosphorylation occurs in vivo, it is by itself not a major determinant of LIN-5 function.
Compared to our transgene rescue experiments (Fig 1B and S1 Fig), the effect of S659 and S662 alanine substitution mutations in endogenous lin-5 was quite mild. The lin-5[S659A,S662A] double mutant animals were viable, with only a slight reduction in intestinal nuclei number (S6A Fig), but displayed a significant increase in embryonic lethality (3.6±1.0% at 25°C, wild type 0.9±0.4%). The phosphomimetic lin-5[S659E,S662D] mutation did not cause embryonic lethality or larval defects, consistent with constitutive phosphorylation of these residues in early embryos (S6A Fig).
In stark contrast, alteration of the candidate CDK-1 phosphorylated residues in the N-terminus, threonine 168 and 181 to alanine (lin-5[T168A,T181A]), aspartic acid (lin-5[T168D,T181D]) or glutamic acid (lin-5[T168E,T181E]), all resulted in typical lin-5 mutant offspring. Regardless of the mutant combination, homozygous animals derived from heterozygous parents developed into sterile, thin and uncoordinated larvae, and showed severely impaired cell division during larval development (Fig 4A and 4B). Importantly, substitution of threonine 168 and 181 with serine residues (lin-5[T168S, T181S]) did not lead to any detectable phenotype or defects in cell division (Fig 4A and 4B). These observations and the in vivo phosphorylation of T168 and T181 indicates that phosphoregulation of T168 and T181 is critical for LIN-5 function, in agreement with the results of the transgene rescue experiments (Fig 1B and S1 Fig). Together, our targeted genome alterations identified several individual amino acids that are required for the in vivo function of LIN-5, including phosphorylated residues in the N-terminus and residues in the GPR-1 binding domain.
Additional characterizations of the non-phosphorylatable and phosphomimetic mutants revealed insight in the functional contribution of LIN-5 phosphorylation. As the contribution of S659 and S662 phosphorylation appeared quite subtle, we examined the spindle in early embryos with substitutions of these residues in detail. In the wild type, meiosis completes after fertilization and results in the formation of a haploid maternal pronucleus, which migrates to meet the paternal pronucleus in the posterior, after which the adjoined pronuclei and centrosomes migrate to the center, rotate and form a spindle along the long axis of the zygote [1,6] (S1 Video). Observations with differential interference contrast (DIC) microscopy showed that these events all occur normally in lin-5[S659A,S662A] and lin-5[S659E,S662D] mutants. Subsequently, in wild type embryos, the chromosomes become aligned at the metaphase plate and are segregated to opposite poles during anaphase. During spindle elongation, the posterior spindle pole oscillates extensively, while the anterior pole remains relatively steady. This coincides with spindle movement towards the posterior, and is followed by flattening of the posterior pole (S1 Video). Starting in anaphase, mutant embryos with non-phosphorylatable lin-5[S659A,S662A] deviated from the wild type, while lin-5[S659E,S662D] mutants showed no phenotype. Specifically, lin-5[S659A,S662A] mutants showed significantly dampened oscillation of both the anterior and posterior pole, reduced spindle elongation, and nearly absent flattening of the posterior spindle pole (Fig 4C and S6B Fig). Nevertheless, both non-phosphorylatable and phosphomimetic S659,S662 mutants underwent asymmetric division of the zygote as normal, which resulted in the formation of a larger anterior blastomere (AB) and smaller germline precursor cell (P1). The spindle normally rotates by 90 degrees prior to division of the P1 blastomere (S1 Video). This failed to occur or was incomplete in 47.1% of the lin-5[S659A,S662A] two-cell embryos, compared to 6.3% and 7.3% incomplete rotation scored in wild type and lin-5[S659E,S662D] mutant embryos, respectively (Fig 4C and S6B Fig). As protein levels were comparable to wild type (S5B Fig), these results suggest that cortical pulling forces are reduced in lin-5[S659A,S662A] mutants. Interestingly, this does not disrupt the asymmetry of the first division and has only a small effect on viability.
To determine cortical pulling forces more directly, we performed spindle severing assays with a UV laser beam [13]. Confirming our DIC analyses, the peak velocities of spindle pole movements were significantly reduced in lin-5[S659A,S662A] embryos (anterior pole 20.5%, posterior pole 18.4% reduced compared to wild type) (Fig 4D, S2 and S5 Videos). Similar experiments performed with lin-5[S659E,S662D] mutant embryos and PAR-1 phosphorylation site mutants (S397A and S397E) did not reveal significant divergence from the wild type (Fig 4D, S3, S4 and S6 Videos). These data support the conclusion that phosphorylation of S659 and S662 contributes to cortical pulling forces, both in the anterior and posterior, and thereby to spindle pole oscillation, spindle elongation, posterior pole flattening and spindle rotation in P1. Moreover, the finding that pulling forces, albeit reduced, remained asymmetric in lin-5[S659A, S662A] mutants explains why these mutants show normal asymmetry of the first division, and normal sizes of the AB and P1 blastomeres.
We wondered whether Wnt-signaling could locally control GSK-3 kinase activity to affect LIN-5 S659, S662 phosphorylation and asymmetric cell division. In the EMS blastomere of the 4-cell embryo, the spindle rotates from a left/right to anterior/posterior position to correctly specify and position the E and MS daughter cells [33]. This rotation is redundantly controlled by MES-1/SRC-1 and MOM-2/MOM-5 Wnt/Frizzled signaling pathways [34]. We examined whether the Wnt pathway contributes to EMS spindle rotation through phosphorylation of LIN-5[S659,S662]. Making use of a mes-1(bn74ts) mutant strain expressing GFP-β-tubulin, we observed normal spindle rotation in lin-5[S659A,S662A] mutant embryos, with only one of 13 embryos showing a tilted spindle angle in the EMS blastomere (S6C Fig). mes-1(bn74ts); lin-5[S659E,S662D] mutant embryos showed an occasional failed rotation or tilted spindle angle. In control mes-1(bn74ts); gsk-3(RNAi) mutants, the EMS spindle failed to rotate in 9/11 embryos (S6C Fig). This clear difference in phenotype shows that LIN-5 S659 phosphorylation is not the major contribution of GSK-3 in EMS spindle rotation. Asymmetric divisions of epithelial seam cells in the C. elegans epidermis also depend on a Wnt-β- catenin asymmetry pathway [35,36], and remained normal in lin-5[S659A,S662A] and lin-5[S659E,S662D] mutants. Thus, evidence for developmental regulation of LIN-5–GPR-1/2 interaction through Wnt-signaling was not obtained. Instead, absence of unphosphorylated S659,S662 peptides in our mass spectrometry analyses, and the wild type appearance of phosphomimetic mutants point to constitutive phosphorylation of the S659,S662 residues.
Our yeast two-hybrid data showed reduced interaction between LIN-5[S659A,S662A] and GPR-1 compared to wild type, which likely explains the reduced pulling forces observed in vivo. We examined whether the colocalization between LIN-5 and GPR-1 in vivo depends on LIN-5 phosphorylation. Hereto, we generated strains with lin-5[S659,S662] double phosphorylation-site alterations in combination with egfp::gpr-1, a CRISPR/Cas9-mediated knockin allele of the endogenous gpr-1 locus. Immunohistochemical detection of eGFP and LIN-5 showed normal colocalization of LIN-5 and GPR-1 in phosphomimetic lin-5[S659E,S662D] mutants at the centrosomes and cell cortex (Fig 5 and S7 Fig; note that LIN-5 becomes clearly visible at the cortex only after the one-cell stage). In contrast, in lin-5[S659A,S662A] mutant embryos, GPR-1 localized to the cortex but no longer accumulated at the centrosomes (Fig 5). Notably, GPR-1/2 localization at the cortex primarily depends on association with the GOA-1 and GPA-16 Gα proteins and is required for pulling forces, whereas ASPM-1–LIN-5 anchors GPR-1/2 at the centrosome without early embryonic requirement [21,37]. Thus, while the loss of centrosomal GPR-1 appears to confirm a reduced binding affinity for LIN-5[S659A,S662A] compared to wild type LIN-5, the reduced pulling forces likely result from a similarly reduced affinity between these proteins at the cortex. Nevertheless, LIN-5 still localized to the cortex in lin-5[S659A,S662A] mutants (S7 Fig). This likely reflects different dynamics of the two complexes; with rapid exchange of LIN-5 at the cortex while centrosomal GPR-1/2 accumulation likely depends on prolonged LIN-5 association. The combined observations in yeast two-hybrid assays, phenotypic analyses, and protein localization studies strongly indicate that phosphorylation of LIN-5 residues S659 and S662 contributes to the affinity of the LIN-5/GPR-1/2 interaction in vivo.
Characterization of the CDK-1 phosphorylation site mutants required a different strategy, as homozygous lin-5[T168A,T181A] and lin-5[T168D,T181D] mutants are fully sterile. To be able to examine the effects of these mutations in early embryos, we created trans-heterozygotes carrying these mutations and egfp::lin-5, a functional CRISPR/Cas9-generated knockin allele of endogenous lin-5. The egfp::lin-5 allele served both as a visible balancer for the lin-5 phosphorylation site mutations, and allowed selective knockdown of functional lin-5 by RNAi against egfp. This strategy allowed us to obtain and characterize early embryos with CDK1-phosphorylation site alterations in LIN-5.
Control immunohistochemical staining experiments confirmed that egfp RNAi treatment of homozygous egfp::lin-5 adults completely removed LIN-5 and eGFP from the offspring (S8 Fig). Following egfp RNAi treatment of heterozygous animals with wild type lin-5 (lin-5(+) / egfp::lin-5), LIN-5 localized normally, but the eGFP staining was lost (Fig 6). These results demonstrate that the RNAi effect remains specific for egfp::lin-5 and does not carry over to the untagged lin-5 allele. Testing balanced lin-5[T168A,T181A] and lin-5[T168D,T181D] animals the same way, we observed that the mutant LIN-5 proteins are expressed and localize as normal to the cortex and centrosomes, while the early embryonic divisions were clearly defective (Fig 6 and S5C Fig). Interestingly, lin-5[T168A,T181A] and lin-5[T168D,T181D] showed similar abnormalities, emphasizing the critical role for the in vivo phosphorylated threonine residues at these positions.
Using the above-described method, we also performed live imaging by time-lapse DIC microscopy and spindle severing experiments with lin-5[T168A,T181A] and lin-5[T168D,T181D] mutant embryos. Again, the defects observed in both mutants resembled lin-5 strong loss-of-function [19,26], and cortical pulling forces were greatly reduced in both mutants (Fig 4E and S7 and S8 Videos). In contrast, homozygous lin-5[T168S,T181S] mutants showed normal spindle pulling forces (Fig 4E and S9 Video). This indicates that the two threonine residues are not essential per se, but phosphorylation and de-phosphorylation at these sites is likely critical. The lin-5[T168S,T181S] mutants did show dampened spindle oscillation, which might result from somewhat different kinetics of threonine versus serine phosphorylation and dephosphorylation in CDK1 substrates [38].
Because the N-terminus of LIN-5 is implicated in the recruitment of dynein [9], we crossed both mutants with an mCherry::dhc-1 strain, in which the mCherry tag was introduced into the endogenous dynein heavy chain gene by CRISPR/Cas9-mediated knockin. This homozygous mCherry::dhc-1 strain is viable and develops as normal. mCherry::DHC-1 was diffusely detected in the cytoplasm, and distinctly localized at the nuclear envelope, kinetochores, astral microtubules, spindle poles and cell cortex. Localization of dynein was dynamic during all stages of mitosis, but cortical dynein was barely detectable at the one-cell stage. However, following treatment of permeabilized embryos with nocodazole to depolymerize microtubules, mCherry::DHC-1 accumulated on the cell cortex of one-cell embryos in metaphase and anaphase (Fig 7A and S9 Fig). Strikingly, this cortical dynein localization was abolished by lin-5 RNAi, and did not occur in lin-5[T168A,T181A] and lin-5[T168D,T181D] mutant embryos (Fig 7A). Since these mutant LIN-5 forms localize to the cell cortex, T168 and T181 are critical for the function of LIN-5 as a cortical dynein anchor.
In addition to cortical localization of dynein in mitosis, LIN-5 is also required for dynein recruitment to the poles of the meiotic spindle [37]. Accumulation of dynein at the spindle poles, as well as the cell cortex, occurs coincident with anaphase onset of meiosis I and II, and is needed for spindle rotation and expulsion of chromosomes into a polar body [37,39,40]. While homozygous lin-5[T168S,T181S] mutants showed normal meiosis, we observed polar body absence and abnormally large polar bodies in eGFP::LIN-5-depleted lin-5[T168A,T181A] and lin-5[T168D,T181A] embryos, consistent with lin-5 loss of function. To examine meiotic spindle rotation and dynein localization in such embryos, we combined the lin-5 mutations, balanced by egfp::lin-5, with homozygous gfp::tbb-2 β—tubulin and mCherry::DHC-1 dynein reporters (Fig 7B and S10–S21 Videos). In a control strain with wild type LIN-5, spindle rotation and dynein accumulation occurred in 10 of 11 embryos (the one exception showed rotation but only weak mCherry::DHC-1 accumulation) (Fig 7B left and S10–S12 Videos). Examination of egfp RNAi treated egfp::lin-5 embryos with combined DIC and fluorescence microscopy revealed normal diffuse association of DHC-1 with the meiotic spindle in meiotic prophase, followed by gradual loss of mCherry::DHC-1 from the anaphase spindle, rather than accumulation of dynein at the poles. The failure in dynein localization coincided with failure to rotate the meiotic spindle (Fig 7B). These results agree with our previously reported meiotic lin-5 RNAi phenotype [21,37], although this time we also observed abnormally elongated meiotic spindles in meiosis II in a subset of the embryos, as has been reported for dynein complex subunits [40]. eGFP::LIN-5-depleted lin-5[T168A,T181A] and lin-5[T168D,T181D] embryos were indistinguishable from lin-5 knockdown mutants (Fig 7B and S13–S21 Videos).
In conclusion, substitution of LIN-5 T168 and T181 with non-phosphorylatable alanine or phosphomimetic aspartic acid residues creates a severe defect in LIN-5-mediated dynein recruitment. In contrast, replacement of the same residues with phosphorylatable serine residues did not compromise LIN-5 function (Fig 4A, 4B and 4E). Combined with the available literature [37,39,41], these data point to CDK-1-mediated phosphorylation and subsequent dephosphorylation of the LIN-5 N-terminus as a critical step in dynein recruitment to the meiotic spindle and cell cortex (see below).
In this study, we investigated whether the extensive in vivo phosphorylation of the LIN-5NuMA protein is important for chromosome segregation and cell cleavage plane determination. We combined in vivo and in vitro phosphorylation analysis, identified critical phosphorylated LIN-5 residues by complementation, and defined the LIN-5–GPR-1 interaction domain by reverse yeast two-hybrid screening. Using this information, we created phosphosite mutants and tagged alleles by genetic engineering, and determined the in vivo contribution of individual phosphorylated residues by protein localization studies, time-lapse microscopy and spindle severing experiments. The combined data indicate that a variety of cell cycle and polarity kinases phosphorylate LIN-5, with specific phosphorylations promoting pulling force generation while others inhibit LIN-5 function. The combined phosphorylations of the LIN-5 N-terminus and C-terminus are critical in the spatiotemporal control of cortical pulling forces, and thereby for correct chromosome segregation and spindle positioning (Fig 8).
CRISPR/Cas9-mediated genomic engineering has added an important tool to a powerful genetic system, and more efficient procedures are continuously developed [31,42–46]. The use of CRISPR/Cas9 allowed us to precisely alter one or two codons of specific serine/threonine residues within the normal genetic background. Using knockin alleles eliminates unwanted effects of transgene overexpression or silencing. In particular transgene silencing has long hampered lin-5 studies and was also observed in our complementation studies. Transgene expression levels that are close to a threshold level may explain why the lin-5[S659A,S662A] mutation showed a strong loss-of-function phenotype, while the effect of the same mutations introduced in the endogenous locus was less severe. In addition to phosphosite mutations, we also created tagged endogenous alleles of lin-5, gpr-1 and dhc-1 for fluorescent fusion protein expression. This allowed the development of a novel method for analysis of early lethal mutations. This method makes use of a functional eGFP-tagged allele, which acts as a visible balancer and allows the specific removal of wild type function by egfp RNAi.
In a previous study, we revealed in vivo kinase activity through differential labeling of C. elegans cultures with stable nitrogen isotopes, followed by kinase knockdown and quantitative analysis of phosphopeptides by mass spectrometry [15]. This strategy worked well for PKC-3, but various limitations can prevent detection of kinase-substrate relations in vivo. The phosphorylation of LIN-5 by PAR-1 was missed in our previous analysis, because of overlap between the relevant LIN-5 phosphopeptides and unrelated peptides. Identification of mitotic substrates of CDK-1 is difficult in vivo, because CDK-1 knockdown results in complete sterility and arrest of fertilized oocytes before completion of meiosis [47]. Casein kinase I, in turn, is represented by 87 family members in C. elegans [48], making it less likely that knockdown experiments will reveal a quantitative difference in substrate phosphorylation. The in vitro kinase assays in the current study revealed candidate kinases that were otherwise difficult to detect. The PAR-1 in vitro kinase assays pointed to a specific LIN-5 phosphorylation that was subsequently confirmed by our in vivo data. The in vitro phosphorylation of peptides with single phosphorylated residues was instrumental in detecting a probable two-step mechanism for S662 phosphorylation by CKI, following a priming phosphorylation by GSK3. Thus, while detecting direct phosphorylation in vivo remains the ultimate goal, in vitro assays continue to provide meaningful insight.
The combined in vitro and in vivo kinase analyses strongly suggest that PAR-1 phosphorylates LIN-5 at serine 397. Replacing this serine with non-phosphorylatable alanine or phosphomimetic glutamic acid apparently did not affect viability, development, cell division, chromosome segregation or spindle pulling forces. In fact, many phosphorylations that occur in vivo may be bystander rather than regulatory events, and determining which phosphorylations are critical in vivo has received great attention in the current study. The first selection came from alanine substitution mutagenesis combined with complementation of a lin-5 null mutation. This revealed that 4 of the 25 phosphorylated residues are critical for LIN-5 function. As we could only score larval divisions in this assay, we cannot exclude that additional phosphorylations may be critical during embryogenesis. Remarkably, 2 of the 4 critical residues form part of a probable GPR-1/2 binding domain, while the other 2 appear to mediate contact with dynein at the cortex.
We defined the GPR-1/2 binding domain through screening for LIN-5 residues that are essential for GPR-1 interaction in yeast two-hybrid assays. The strong clustering of missense mutations in this screen combined with results from deletion analyses suggests a short linear GPR-interaction epitope. This is in full agreement with results from crystal structure studies of the related NuMA-LGN complex. The TPR repeats in the N-terminal half of LGN form helix-turn-helix repeats that together organize into a superhelical bundle [49,50]. The inner surface of this bundle forms a binding channel for an extended NuMA peptide of 28 amino acids [49]. Many electrostatic and hydrogen interactions between side chains of the NuMA peptide and TPR motifs together provide a high affinity binding site.
The TPR-repeat interaction site in LIN-5 resembles that of NuMA in size, position, and overall amino-acid composition. The exact residues are not well-conserved, however, probably because the many amino acids that contribute weak interactions provide a limited biological constraint for the conservation of individual amino acids. Notably, the core of the binding site contains EPEQLDDW in human NuMA and SPDSLPDF in LIN-5, sharing three identical and two similar residues as well as negative charge. The NuMA peptide contains four acidic residues (D, E), while two D residues and two phosphorylated serines are negatively charged in the LIN-5 peptide.
Phosphorylation offers the opportunity to regulate LIN-5–GPR-1/2 binding. In fact, dual GSK3 and CK1 phosphorylation of the LRP6 Wnt-co-receptor regulates the interaction of LRP6 with axin [24,51]. We did not obtain evidence to support developmentally regulated LIN-5–GPR-1 binding. Alanine substitution of S659 and S662 significantly reduced spindle pulling forces, but division of the zygote, EMS blastomere and seam cells continued to be asymmetric. The latter types of divisions depend on the Wnt-β —catenin asymmetry pathway, which in EMS positions the spindle redundantly with mes-1/src-1 signaling [34]. Even the combined lin-5[S659A,S662A] mutation and mes-1 knockdown did not interfere with A-P positioning of the spindle in EMS. Moreover, we could functionally replace serine 659 and 662 with glutamic and aspartic acid, suggesting that charge, rather than phosphoregulation, is critical for GPR-1/2 interaction.
A contribution of CDK-1 phosphorylation in LIN-5 regulation was expected. CDK1/cyclin B kinases are the master regulators of mitosis that phosphorylate hundreds of substrate proteins [22,52,53]. The LIN-5 N- and C-terminus and corresponding domains in NuMA contain multiple CDK1 consensus sites. CDK1/cyclin B has been shown to regulate Xenopus and human NuMA through phosphorylation of the C-terminus [18,54]. Specifically, phosphorylation at T2055 interferes with the cortical localization of NuMA, thereby inhibiting dynein recruitment until CDK1/cyclin B is inactivated at the metaphase/anaphase transition [18]. Our results indicate that this temporal regulation may also involve critical phosphorylation of the dynein-interacting N-terminus of NuMA by CDK1/cyclin B.
In C. elegans, dynein recruitment to the meiotic spindle and cell cortex, as well as mitotic pulling forces, depend on activation of the anaphase promoting complex/cyclosome (APC/C), and inactivation of CDK-1/cyclin B [37,39,41]. Thus, phosphorylation of specific mitotic substrates by CDK-1/cyclin B is likely to inhibit dynein recruitment and pulling force generation. A recent study identified the p150 dynactin subunit as a likely candidate for inhibition by CDK-1/cyclin B phosphorylation [40]. Our results point to the LIN-5 N-terminus as another critical target for CDK-1 regulation. Supporting this conclusion, T168, T181 are part of CDK consensus sites, are phosphorylated in vivo, and are efficiently phosphorylated by CDK1/cyclin B in vitro. Substitution of LIN-5 T168 and T181 with phosphomimetic glutamic acid or aspartic acid residues resulted in strong loss of LIN-5 function, supporting that CDK-1 phosphorylation normally inhibits LIN-5. More surprising, an indistinguishable phenotype was observed following T168 and T181 replacement with non-phosphorylatable alanine. This could indicate that the threonine residues are critical for LIN-5 folding, or that phosphorylation of these threonines in the N-terminus also contributes to dynein recruitment. In stark contrast to alanine substitution, replacement of the same residues with phosphorylatable serine had no detectable effect on pulling forces, meiotic and mitotic cell divisions, viability and fertility. While other explanations are possible, these data are consistent with a required sequential CDK-1 phosphorylation and dephosphorylation of LIN-5 T168 and T181. Therefore, we propose a two-step model, in which CDK-1/cyclin B induces the assembly of a LIN-5 pre-force generating complex in prometaphase. Subsequent removal of the phosphates, which follows CDK inactivation by the APC/C at anaphase onset, promotes interaction of this complex with dynein.
Many of the lessons learned from studies in worms and flies have subsequently been found to apply broadly to the animal kingdom. The initial discovery of LIN-5 requirement in spindle positioning in C. elegans [19] has contributed to identifying similar functions for NuMA in mammalian systems [55]. It will be intriguing to find out to what extent the phosphoregulation of pulling forces translates from C. elegans to mammalian systems, and specifically whether reversible CDK1 phosphorylation of the NuMA N-terminus controls dynein interaction and spindle positioning.
Strains were cultured on nematode growth medium plates, seeded with Escherichia coli OP50 as previously described [56]. Animals were maintained at 20°C, unless stated otherwise. All strains used in this study are found in S1 Table. Genome modifications in strains SV1568, SV1569, SV1586, SV1588, SV1589, SV1590, SV1600, SV1619, SV1621, SV1622, SV1695 and SV1901 were introduced by making use of CRISPR/Cas9 genome editing as described below.
For functional analysis of wild type and phosphomutant LIN-5, 5 ng/μl Plin-5::gfp::lin-5 DNA, together with 5 ng/μl Psur-5::dsRed and 25 ng/μl Lambda DNA (Fermentas), was injected into the gonad of SV918 young adults. Psur-5::dsRed positive F1 progeny were selected making use of a fluorescence stereo microscope (Leica, MZ16F). After this, lin-5(e1348) homozygous animals were selected based on absence of pharyngeal Pmyo-2::gfp, expressed from the mIn1 balancer chromosome. Rescue analysis of lin-5 null animals was performed by Differential Interference Contrast and fluorescence microscopy, using a Zeiss Axioplan microscope. Intestinal nuclei were counted only in animals expressing Psur-5::dsRed in all intestinal nuclei. Vulval development was assayed for all animals L4 and older.
For quantification of cell numbers in CRISPR/Cas9 knockin mutants, asynchronous populations of animals were fixed, DNA stained with propidium iodide and intestinal and ventral cord nuclei were counted using a Zeiss Axioplan fluorescence microscope. Cells were counted at late larval stages. For the ventral cord, all nuclei of the P2-to-P10 daughter cells and juvenile motor neurons in the region between these cells were counted.
For in vitro CDK1 kinase assays, immunoprecipitations were performed from mitotic lysates of HeLa cells using mouse monoclonal anti cyclin B1 (GNS1) or beads alone as negative control. Immunoprecipitations were incubated for 30 min at 30°C with either Histone H1, bacterially produced GST or GST–LIN-5 in kinase buffer containing 50 mM HEPES at pH 7.5, 5 mM MgCl2, 2.5 mM MnCl2, 1 mM dithiothreitol, 50 μM ATP and 2.5 μCi [γ-32P] ATP. Reactions were terminated by the addition of SDS (5x sample buffer). For mass spectrometry analysis, no [γ-32P] ATP was added to the kinase assays and incubation time was prolonged to 2 hours at 30°C.
For in vitro GSK3 and CK1 kinase assays, peptides (RRRIRCGSPDSLPDFLADN) containing either unphosphorylated, phosphorylated S659 or phosphorylated S662 were used. Kinases were incubated for 30 min at 25°C with synthetic peptide in kinase buffer containing 200 μM ATP, 50 mM HEPES at pH 7.5, 10 mM MgCl2, 1 mM EGTA, 2 mM dithiothreitol, supplemented with 20 μCi [γ-32P] ATP for radioactive kinase assays. Reactions were terminated by the addition of SDS (4x sample buffer).
All other in vitro kinase assays were performed as previously described [15].
In short, kinases were incubated for 30 min at 25°C with bacterially produced GST or GST–LIN-5 in kinase buffer containing 200 μM ATP, 50 mM HEPES at pH 7.5, 10 mM MgCl2, 1 mM EGTA, 2 mM dithiothreitol, supplemented with 20 μCi [γ-32P] ATP for radioactive kinase assays. Reactions were terminated by the addition of SDS (4x sample buffer). For mass spectrometry analysis, no [γ-32P] ATP was added to the kinase assays and incubation time was prolonged to 2 hours at 25°C. Kinases used in this study were: recombinant C. elegans PAR-1 (a kind gift from Erik Griffin and Geraldine Seydoux), and mammalian Aurora B (a kind gift from Susanne Lens), CK1 (New England Biolabs), CK2 (New England Biolabs), and GSK3 (New England Biolabs).
Gel bands were cut and processed for protein in-gel digestion as described elsewhere [15]. Briefly, proteins were reduced with dithiothreitol and then alkylated with iodoacetamide. Trypsin was added at a concentration of 10 ng/μl and the mixture was digested overnight at 37°C. Subsequently, peptides were collected from the supernatants and a second extraction using 10% formic acid was performed. Phosphopeptides from LIN-5 were enriched using TiO2 chromatography [57]. Basically, home-made GELoader tips (Eppendorf, Hamburg, Germany) were packed with TiO2 beads (5 μm, INERTSIL). Peptides were loaded in 10% formic acid and subsequently washed with 20 μl of 80% acetonitrile, 0.1% trifluoroacetic acid (Fluka, Sigma-Aldrich). Phosphopeptides were then eluted twice with 20 μl of 1.25% ammonia solution (Merck, Germany), pH 10.5, and 3 μl of 100% formic acid was finally added to acidify the samples.
Nanoflow LC-MS/MS was carried out by coupling an Agilent 1100 HPLC system (Agilent Technologies, Waldbronn, Germany) to an LTQ-Orbitrap XL mass spectrometer (Thermo Electron, Bremen, Germany). Peptide samples were delivered to a trap column (AquaTM C18, 5 μm (Phenomenex, Torrance, CA); 20 mm x 100-μm inner diameter, packed in house) at 5 μl/min in 100% solvent A (0.1 M acetic acid in water). Next, peptides eluted from the trap column onto an analytical column (ReproSil-Pur C18-AQ, 3μm (Dr. Maisch GmbH, Ammerbuch, Germany); 40 cm x 50-μm inner diameter, packed in house) at ~100 nl/min in a 90 min or 3 h gradient from 0 to 40% solvent B (0.1 M acetic acid in 8:2 (v/v) acetonitrile/water). The eluent was sprayed via distal coated emitter tips butt-connected to the analytical column. The mass-spectrometer was operated in data-dependent mode, automatically switching between MS and MS/MS. Full-scan MS spectra (from m/z 300 to 1500) were acquired in the Orbitrap with a resolution of 60,000 at m/z 400 after accumulation to target value of 500,000 in the linear ion trap. The five most intense ions at a threshold above 5000 were selected for collision-induced fragmentation in the linear ion trap at a normalized collision energy of 35% after accumulation to a target value of 10,000. Peak lists were created from raw files with MaxQuant43. Peptide identification was carried out with Mascot (Matrix Science) against a Caenorhabditis elegans protein database (http://www.wormbase.org) supplemented with all the frequently observed contaminants in MS (23,502 protein sequences in total). The following parameters were used: 10 ppm precursor mass tolerance, 0.6 Da fragment ion tolerance, up to 3 missed cleavages, carbamidomethyl cysteine as fixed modification, oxidized methionine, phosphorylated serine, threonine and tyrosine as variable modifications. Alternatively, MaxQuant and its search engine Andromeda was also employed for peptide identification and quantification. Data are available via ProteomeXchange with identifier PXD004906
Full length gpr-1 was PCR amplified from the ORFeome library (kind gift from Marc Vidal) using KOD polymerase (Novagen) and cloned into bait vector pPC97. Fragments of lin-5 were PCR amplified from the ORFeome library with KOD polymerase (Novagen) and cloned into prey vector pPC86-AN [58]. DB::GPR-1- and AD::LIN-5-encoding plasmids were transformed sequentially into yeast strain Y8930 [58]. Positive interactions were identified on the basis of the activation of the HIS3 and ADE2 reporter genes, indicated by growth on synthetic complete –leucine −tryptophan −histidine + 2 mM 3-Amino-1,2,4-triazole (Sc −leu−trp−his + 3-AT) and synthetic complete –leucine −tryptophan −adenine (Sc −leu−trp−ade) plates.
To generate mutant clones, PCR was performed on pVP054 (pPC86-AN containing nucleotides 1821–2466 encoding amino acid 609–821 of lin-5) with increased MgCl2 concentration of 7 mM. PCR products were cloned into pPC86-AN and transformed to DH5α competent cells. Bacterial colonies were collected and the DNA was isolated using a Nucleobond Xtra DNA purification kit (Macherey-Nagel). Bacterial clones were transformed into MB004 (Y8930 in which ADE2 is replaced by URA3 by homologous recombination) containing DB::GPR-1 and plated on synthetic complete –leucine −tryptophan + 2 g/l 5-fluoroacetic acid (Sc −leu−trp + FOA). Colonies were picked and spotted to synthetic complete –leucine −tryptophan (Sc −leu−trp) plates for validation, PCR and sequencing. Clones with single amino acid changes were re-tested by PCR amplification of the fragment and re-cloning into pPC86-AN. Interaction deficient alleles were identified on the basis of no activation of the HIS3 or URA3 reporter gene, indicated by absence of growth on Sc −leu−trp−his + 3-AT plates and synthetic complete –leucine −tryptophan –uracil (Sc −leu−trp−ura) plates.
For generation of full length mutant clones, plasmids containing fragment clones with selected point mutations were digested and cloned into pVP055 (pPC86-AN containing nucleotides 1–2466 encoding amino acid 1–821 of lin-5). Phosphomutants were generated by either site-directed mutagenesis of pVP055 or Gibson assembly into pVP055 of short regions of lin-5 with point mutations carried in the overlapping region. DB::GPR-1 and AD::LIN-5-encoding plasmids were transformed sequentially into yeast strain MB004. Interaction deficient alleles were identified on the basis of no activation of the HIS3 or URA3 reporter gene, indicated by absence of growth on Sc −leu−trp−his + 3-AT plates and Sc −leu−trp−ura plates.
CRISPR repair constructs were inserted into the pBSK vector using Gibson Assembly (New England Biolabs). Homologous arms of at least 1500 bp upstream and downstream of the CRISPR/Cas9 cleavage site were amplified from either cosmid C03G3 (for lin-5 constructs) or C. elegans genomic DNA using KOD Polymerase (Novagen). Linkers containing the altered cleavage sites and point mutations were synthesized (Integrated DNA technologies). For fkbp::egfp::gpr-1, codon-optimized fkbp was synthesized (Integrated DNA technologies) and codon-optimized egfp was amplified from pMA-egfp (a kind gift from Anthony Hyman). For egfp::lin-5, codon-optimized egfp was amplified from pMA-egfp. For mcherry::dhc-1, codon-optimized mcherry was amplified from TH0563-PAZ-mCherry (a kind gift from Anthony Hyman). Mismatches were introduced in the sgRNA target site to prevent cleavage of knockin alleles. All plasmids and primers used for cloning are available upon request.
Young adults were injected with a solution containing the following injection mix: 30–50 ng/μl Peft-3::Cas9 (Addgene 46168; [59], 30–100 ng/μl u6::sgRNA with appropriate target for dhc-1, gpr-1 and lin-5, 30–50 ng/μl repair template and 2.5 ng/μl pmyo-2::tdTomato. Progeny of animals that express tdTomato were picked to new plates 3–4 days post injection. PCRs with primers diagnostic for homologous recombination at the endogenous locus were performed on F2-F3 populations, where one primer targeted the altered basepairs in the sgRNA site, point mutation or fluorescent tag and the other targeted a region just outside the homology arm. All primers used for genome editing are available upon request.
For yeast protein lysates, cultures were grown overnight at 30°C. Yeast cells corresponding to 4 OD of culture were harvested, treated with Sodium hydroxide and resuspended in 100 2X sample buffer containing β-mercaptoethanol [60].
For C. elegans protein lysates, strains SV1569, SV1663 and SV1664 were grown at 20°C one generation on NGM plates seeded with HB101, followed by a second generation in S-Medium with HB101 bacteria. Gravid adults were harvested and embryos were isolated by hypochlorite treatment. Embryo pellets were snap frozen in liquid nitrogen, grinded using mortar and pestle and resuspended in 5 ml fresh lysis buffer (containing 20 mM Tris-HCl pH 7.8, 250 mM NaCl, 15% glycerol, 0.5% IGEPAL, 0.5 mM EDTA, 50 mM Sodium fluoride, 1 mM β-mercaptoethanol and protease inhibitors (Roche complete, Mini, EDTA-free)). The suspension was passed through a French press 3 times, and the lysate was cleared at 13,000 rpm for 15 min at 4°C.
Protein samples were separated on gradient acrylamide gels and subjected to western blotting on polyvinylidene difluoride membrane (Immobilon-P, Millipore). Membranes were blocked with 5% skim milk in PBST for 1 hour at room temperature, or overnight at 4°C for stripped blots. For protein detection, primary antibodies used in this study were: mouse anti–LIN-5 (1:1000) [19] and rabbit anti-Tubulin (1:1000, Abcam) for stripped blots. Secondary antibodies used were: donkey anti-mouse HRP (1:5000, Abcam) and goat anti-rabbit HRP (1:5000, Jackson Immunoresearch). Proteins were dectected with Signalfire Plus chemiluminescent detection (Cell Signaling Technologies) and a Chemidoc MP Imager (Bio-Rad).
DIC time-lapse imaging was performed on strains N2, SV1568, SV1588, SV1590 and SV1600. Animals were grown overnight at 25°C. RNAi feeding of N2 against gsk-3 was performed with the bacterial clone from the Orfeome-based RNAi Library [61,62]. L4 animals were grown for approximately 32 hours at 15°C before shifting overnight to 25°C for imaging, except SV1901 which was kept at 20°C. Embryos were dissected from adults in a solution of 0.8x egg salt (containing 94 mM NaCl, 32 mM KCl, 2.7 mM CaCl2, 2.7 mM MgCl2, 4 mM HEPES, pH 7.5; [63]) on coverslips and mounted on slides with 3% agarose prepared with egg salt. Embryos were imaged with 5s time intervals with a 100x/1.4 NA lens on a Zeiss microscope at 20°C. Relative positions of the spindle and furrow were analyzed manually using ImageJ.
Live-cell imaging of EMS rotation was performed on strains SV1783, SV1784 and SV1785. L4 animals were grown overnight at 20°C. Embryos were dissected from young adults as above and imaged with a 100x/1.4 NA lens on a Zeiss microscope at 20°C. Spindle rotation in EMS was followed over time with images taken at several time points.
Live-cell imaging of microtubule depolymerization upon nocodazole treatment was performed on strain AZ244. Young adult animals were injected with dsRNA [64] against perm-1 and grown for 20 hours at 15°C. Embryos were dissected from young adults in a solution of 0.8x egg salt containing 1 μM nocodazole on coverslips and mounted on concave slides. Embryos were imaged with a 60x/1.4 NA lens on a Nikon Eclipse Ti microscope with Perfect Focus System and Yokogawa CSU-X1-A1 spinning disk confocal head at room temperature.
Live-cell imaging of mitotic DHC-1 localization was performed on strains SV1619, SV1635, SV1638 and SV1639. Young adult animals were injected with dsRNA [64] against perm-1 and egfp and grown for 20 hours at 15°C. Embryos were dissected from young adults in a solution of 0.8x egg salt containing 1 μM nocodazole on coverslips and mounted on concave slides. Still images of mitotic embryos in metaphase were taken within minutes after nocodazole addition. Eliminating nonspecific toxic effects, embryos on the same slide at the same time continued nuclear envelope degradation, and perm-1(RNAi) embryos continued embryonic development in the absence of nocodazole.
Embryos were imaged with a 60x/1.4 NA lens on a Nikon Eclipse Ti microscope with Perfect Focus System and Yokogawa CSU-X1-A1 spinning disk confocal head at room temperature. mCherry::DHC-1 localization was analyzed in mitotic one-cell embryos after nuclear envelope breakdown. Mitotic embryos were also identified based on presence of a polar body, enlarged centrosomes and remnant of the mitotic spindle.
Live-cell imaging of meiotic DHC-1 localization was performed on strains SV1702, SV1898, SV1899 and SV1902. For RNAi treated animals, young adult animals were injected with dsRNA [64] against egfp and grown for 24 hours at 15°C. Embryos were dissected from young adults as above and imaged with 10s time intervals with a 100x/1.4 NA lens on a Zeiss microscope at 20°C. For images presented in Fig 7B and S12, S15, S18 and S21 Videos, images were processed by subtracting a Gaussian-blur filtered image (Sigma(Radius): 20) using ImageJ. S10, S11, S13, S14, S16, S17, S19 and S20 Videos represent the unprocessed files.
Spindle severing with a UV laser microbeam was performed on strains SV1585, SV1594, SV1596, SV1618, SV1700 and SV1701 essentially as previously described [13]. RNAi feeding of SV1700 and SV1701 against egfp was performed with a bacterial clone containing full length egfp in the L4440 double T7 plasmid [61]. L4 animals were grown on RNAi for approximately 48 hours. For analysis of spindle pulling forces, animals were kept at 25°C for 24 h before ablations. Spindle ablations were carried out at 25°C (Fig 4D) or 20°C (Fig 4E) on a spinning disk confocal microscope. The spindle midzones were severed at anaphase onset and images of GFP-β-tubulin were taken at 0.5 s intervals. For analysis, the position of the spindle poles was automatically tracked using the MTrack2 plugin in ImageJ. Peak velocities of the anterior and posterior spindle poles were determined within a 12.5 s time frame after ablation. Representative videos for every strains are shown in S2–S9 Videos.
Microscope setup: Nikon Eclipse Ti microscope with Perfect Focus System, Yokogawa CSU-X1-A1 spinning disk confocal head, S Fluor 100x N.A. 0.5–1.3 objective (at 1.3), Photometrics Evolve 512 EMCCD camera, Cobolt Calypso 491 nm (100 mW) and Teem Photonics 355 nm Q-switched pulsed lasers, ILas system (Roper Scientific France/ PICT-IBiSA, Institut Curie) to control the UV laser, ET-GFP (49002) filter, ASI motorized stage MS-2000-XYZ with Piezo Top Plate with Tokai Hit INUBG2E-ZILCS Stage Top Incubator (controlled at 25°C), controlled by MetaMorph 7.7 software.
For immunostaining, embryos were dissected from adults in 8 μl of water on poly-L-lysine–coated slides. Embryos were freeze-cracked and fixed for 5 min in methanol at −20°C and then for 20 min in acetone at −20°C. After fixation, embryos were rehydrated in phosphate-buffered saline (PBS) containing 0.05% Tween-20 (PBST) and blocked with blocking solution (PBST containing 1% bovine serum albumin and 1% goat serum [Sigma-Aldrich]) for 1 h. Embryos were stained with primary and secondary antibodies for 1 h and washed after each incubation with PBST four times, 15 min each time. Finally, the embryos were embedded in ProLong Gold Antifade containing 4′,6-diamidino-2-phenylindole (DAPI). Primary antibodies used in this study were: mouse anti–LIN-5 (1:10; [19] and rabbit anti-GFP (1:500, Life Technologies). Secondary antibodies were used at a concentration of 1:500. Secondary antibodies used were: goat anti-rabbit Alexa Fluor 488, and goat anti-mouse Alexa Fluor 568 (Invitrogen). Images were taken with a 63x/1.4 NA lens on a Zeiss confocal microscope.
Embryos were dissected and stained with antibodies as described above. Images were taken with a 63x/1.4 NA lens on a Zeiss confocal microscope using identical microscope setting for all images taken for every secondary antibody. Mean intensity of LIN-5 was measured using ImageJ by selecting fixed size regions that depended on the developmental stage. For every embryo 2 centrosomes and cytoplasmic regions were quantified.
|
10.1371/journal.ppat.1002217 | c-di-AMP Is a New Second Messenger in Staphylococcus aureus with a Role in Controlling Cell Size and Envelope Stress | The cell wall is a vital and multi-functional part of bacterial cells. For Staphylococcus aureus, an important human bacterial pathogen, surface proteins and cell wall polymers are essential for adhesion, colonization and during the infection process. One such cell wall polymer, lipoteichoic acid (LTA), is crucial for normal bacterial growth and cell division. Upon depletion of this polymer bacteria increase in size and a misplacement of division septa and eventual cell lysis is observed. In this work, we describe the isolation and characterization of LTA-deficient S. aureus suppressor strains that regained the ability to grow almost normally in the absence of this cell wall polymer. Using a whole genome sequencing approach, compensatory mutations were identified and revealed that mutations within one gene, gdpP (GGDEF domain protein containing phosphodiesterase), allow both laboratory and clinical isolates of S. aureus to grow without LTA. It was determined that GdpP has phosphodiesterase activity in vitro and uses the cyclic dinucleotide c-di-AMP as a substrate. Furthermore, we show for the first time that c-di-AMP is produced in S. aureus presumably by the S. aureus DacA protein, which has diadenylate cyclase activity. We also demonstrate that GdpP functions in vivo as a c-di-AMP-specific phosphodiesterase, as intracellular c-di-AMP levels increase drastically in gdpP deletion strains and in an LTA-deficient suppressor strain. An increased amount of cross-linked peptidoglycan was observed in the gdpP mutant strain, a cell wall alteration that could help bacteria compensate for the lack of LTA. Lastly, microscopic analysis of wild-type and gdpP mutant strains revealed a 13–22% reduction in the cell size of bacteria with increased c-di-AMP levels. Taken together, these data suggest a function for this novel secondary messenger in controlling cell size of S. aureus and in helping bacteria to cope with extreme membrane and cell wall stress.
| Staphylococcus aureus is an important human pathogen that colonizes the nares and skin of both sick and healthy individuals and causes a variety of infections ranging from superficial skin to invasive infections. The ability of this bacterium to cause disease depends on many factors and is, in part, due to multi-functional cell surface structures. One such structure is lipoteichoic acid (LTA), which is crucial for bacterial growth. In this study we show that LTA is also important for growth of a clinically relevant community-acquired methicillin resistant S. aureus (CA-MRSA) strain and not only for laboratory strains as previously described. We set out to investigate if S. aureus can find a way to survive without LTA and identified strains that can grow and divide almost normally in its absence. Using a whole genome sequencing approach, we found that alterations in one gene, gdpP, allow these strains to grow in the absence of LTA. We show that this mutation causes an increase in the recently identified signaling molecule, c-di-AMP, within the cell. Therefore, with this study we provide information on one of the first functions of this novel secondary messenger, which is in helping bacteria to cope with extreme cell wall stress.
| Staphylococcus aureus is a very prevalent human pathogen that permanently colonizes the nares and skin of approximately 20% of the population, while another 60% are colonized transiently [1]. Infections caused by this pathogen are becoming increasingly more difficult to treat due to its resistance to antibiotic therapy. Where once methicillin was the antibiotic of choice, now only around 60% of S. aureus isolates remain sensitive to this drug. There has also been a rise in the number of community acquired methicillin resistant S. aureus (CA-MRSA) cases in recent years often resulting in severe skin and soft tissue infections as well as invasive diseases such as sepsis, necrotizing pneumonia or osteomyelitis [2], [3]. The ability of S. aureus to cause such a wide range of diseases depends on many factors and is, in part, due to the diverse functions that are linked to its cell envelope. A myriad of proteins are embedded in this structure that allow bacteria to take up nutrients and adhere to diverse surfaces or niches within the human host. It also protects bacteria from environmental insults and at the same time allows the cells to sense and respond to changes in their surroundings, a function crucial for the survival of this pathogen in the host. In addition, the cell wall helps bacteria to maintain their shape and functions to counteract the high internal turgor pressure. Because the cell envelope has such essential functions, it also forms a weak point of the cell, as the inhibition of enzymes required for its synthesis is often lethal or leads to virulence defects. Therefore, this structure has been, and remains, an attractive target for therapeutic interventions.
A typical cell wall of Gram-positive bacteria consists of proteins, peptidoglycan (PG) and the cell wall polymers wall teichoic acid (WTA), which is covalently linked to PG, and lipoteichoic acid (LTA), a polymer anchored to the outside of the bacterial membrane via a lipid moiety [4], [5], [6]. Synthesis of these cell wall components is highly coordinated and any mistakes can lead to cell lysis and death. From studies on the Gram-positive model organism Bacillus subtilis, it has emerged that PG and WTA synthesis enzymes form multi-protein complexes, which are further linked in this organism with cytoplasmic cell shape determining proteins, thereby coordinating and physically linking extracellular and intracellular synthesis processes [7], [8], [9], [10].
S. aureus LTA is an anionic cell wall polymer consisting of a linear chain of glycerolphosphate repeating units that is anchored via a glycolipid to the membrane [11]. The glycerolphosphate subunits are derived from the head group of the membrane lipid phosphatidylglycerol and polymerized on the outside of the cell by the membrane-linked lipoteichoic synthase enzyme LtaS to form the backbone chain [12], [13], [14], [15]. A large fraction of the glycerolphosphate subunits are substituted with D-alanine residues, a modification known to play a key role in the resistance of S. aureus and several other Gram-positive pathogens to cationic antimicrobial peptides [16], [17], [18], [19]. A function for this polymer in the formation of biofilms has been identified and multiple interactions between LTA and eukaryotic cells have been described [20], [21]. An interaction between LTA and macrophage scavenger receptors is thought to occur and help the host to clear bacterial infections [22], [23]. In agreement with this suggestion, scavenger receptor knockout mice are more susceptible to infection with S. aureus [24]. This polymer has also been shown to act as a ligand for Draper, a phagocytic receptor in Drosophila, which upon binding of LTA results in the phagocytosis of S. aureus by Drosophila hemocytes [25]. On the other hand, LTA is also thought to act as an anti-inflammatory molecule on skin cells by suppressing the TLR-3-mediated responses upon skin injury, a key pathway in the induction of inflammation [26].
Recently defined mutants lacking the entire LTA polymer have been constructed and phenotypic analysis indicated an important role for this polymer for normal bacterial growth and morphology [13], [27], [28], [29], [30]. LTA-deficient S. aureus strains have severely impaired growth and can initially only be propagated in medium containing high salt or sucrose concentrations, which are thought to act as osmoprotectants, or at low temperature [13], [27]. Even under conditions permissive for growth, these cells have severe morphological defects, such as an increased cell size and the tendency to clump. In addition, misplacement of cell division septa is observed, highlighting that the lack of LTA on the outside of the cell negatively affects fundamental processes in the cytoplasm of the cell. However, it is currently not known how these processes are coordinated.
In this study we set out to investigate if S. aureus can find a way to survive without LTA and identified LTA-deficient suppressor strains that can grow and divide almost normally in the absence of this multi-functional cell wall polymer. Using a whole genome sequencing approach, it was determined that these strains have acquired mutations in a gene encoding a protein named GdpP (for GGDEF domain protein containing phosphodiesterase). We show that as a consequence of this mutation, the intracellular levels of the novel cyclic dinucleotide c-di-AMP increase drastically in the suppressor strain and gdpP mutant laboratory or CA-MRSA strains. This provides the first experimental evidence that c-di-AMP is produced in S. aureus and that GdpP functions as a c-di-AMP phosphodiesterase in vivo. With this study we provide information on one of the first functions of this novel secondary messenger, which is in helping bacteria to cope with extreme cell wall stress in addition to controlling the cell size of S. aureus, as revealed by microscopic analysis.
In order to determine whether it is possible for S. aureus to compensate for the cell wall stresses introduced by deleting LTA, the RN4220-derived ltaS deletion strains SEJ1ΔltaSN and SEJ1ΔltaSS were created and the lack of LTA confirmed by western blot (Figure S1A in Text S1, data not shown). These strains were constructed under high osmotic conditions in medium containing 7.5% NaCl (N) or 40% sucrose (S) that are, as previously shown, permissive for growth of S. aureus in the absence of LTA (Figure S1B in Text S1) [27]. However, in contrast to this previous study [27], growth of our ltaS deletion strains were dependent on osmoprotectants at both 30°C and 37°C. Plating efficiencies for the ΔltaSN strain decreased from 3.5×108 CFU/ml on 7.5% NaCl containing TSA plates to 8×101 CFU/ml on TSA plates at 37°C and similar low CFUs were obtained at 30°C. SEJ1ΔltaS bacteria had aberrant cell morphologies even under conditions permissive for growth and displayed an enlarged cell size, a tendency to cluster and a misplacement of division sites (Figure 1C) [13], [27]. In summary, our LTA-negative S. aureus RN4220 strains are viable under osmotically stabilizing conditions but not in TSB medium and display the expected morphological and cell division defects.
When SEJ1ΔltaS strains were plated on TSA plates without the addition of sucrose or salt a small number of colonies were obtained. We hypothesized that these colonies arose from bacteria that had acquired compensatory mutation(s) that allow bacteria to grow in the absence of LTA. To analyze this further, five independently isolated suppressor colonies were passed four times in TSB to improve growth. As expected, LTA was still absent in the five suppressor strains 4S4, 4S5, 4N1, 4N2 and 5S4 (Figure 1A), however growth of these strains now more closely resembled that of the parental strain SEJ1 (Figure 1B). Examination of the five suppressor strains by phase contrast and fluorescence microscopy revealed a near WT cell size and a considerable improvement in the accuracy of division site placement, although misplacement of septa still occurred in some cells (Figure 1C).
To further investigate the cell wall properties of these LTA-negative suppressor strains additional assays were performed. The suppressor strains were two to four-fold more susceptible to lysostaphin, nisin, vancomycin, oxacillin, penicillin G and daptomycin (Table S1 in Text S1). In addition, the suppressor strains lysed faster than the control strain in autolysis assays (Figure S2A in Text S1) but had slightly reduced amounts of cell wall-associated hydrolytic enzymes as determined by zymogram assays (Figure S2B in Text S1). LTA production was restored in two suppressor strains, 4S5 and 5S4, by introducing the complementation vector pCN34-ltaS. The amount of cell-associated hydrolytic enzymes increased in these strains (Figure S2C in Text S1), suggesting that LTA has a role in regulating autolysin levels. This is supported by the observations of Oku et al [27] who noted an even greater reduction in the amount of autolysins for their ltaS mutant strain, which could again be complemented. Taken together, S. aureus suppressor strains that can grow in the absence of LTA can be isolated readily and the morphology and cell division pattern defects are significantly improved in these strains. However, differences in autolysis and susceptibility to cell wall active antibiotics are indicative of remaining changes in the cell wall properties of these strains.
It seems likely that the LTA-negative S. aureus suppressor strains 4S4, 4S5, 4N1, 4N2 and 5S4 have acquired mutations elsewhere on the chromosome that allow for improved cell division and growth. A whole genome sequencing approach was used to identify such sequence alterations (for details see materials and methods section). In total, ten genes within these five strains contained mutations with high confidence scores. Five of these mutations were discarded as they were also present in strain SEJ1ΔltaS pCN34-ltaS, an intermediate strain used for the construction of the ltaS deletion strains that still contains an intact copy of the ltaS gene. It is likely that these mutations were introduced during the temperature shift necessary to create the ltaS deletion. The mutations in the remaining five genes were at different positions and consisted of nonsense mutations, amino acid substitutions or DNA inversions (Table 1). The five genes included SAOUHSC_00015, a conserved hypothetical protein with putative diguanylate cyclase and phosphoesterase activity and the only gene mutated in all five suppressor strains; SAOUHSC_01104, encoding for the succinate dehydrogenase SdhA, a TCA cycle enzyme; SAOUHSC_01358, which encodes for a putative permease; SAOUHSC_02001, a conserved hypothetical protein with weak homology to a fusaric acid transporter, and SAOUHSC_02407, a conserved hypothetical protein with homology to DisA, a DNA integrity scanning protein from B. subtilis. In each suppressor strain two to four of these five genes were mutated and all mutations were confirmed by re-sequencing all five genes in each suppressor strain.
As mentioned above, the suppressor strains were passed four times in broth culture to improve growth before phenotypic and sequence analysis. Therefore, one or more of the mutations listed in Table 1 may have arisen during the passing steps to aid with growth and not as a direct consequence of assisting growth of the LTA-negative strains. To address this, chromosomal DNA was isolated from the original suppressor colonies of strains 4S4, 4S5, 4N1, 4N2 and 5S4 and the five genes in question were sequenced. Only the mutations in gene SAOUHSC_00015 were present in all five strains. Thus, it appears that inactivation of SAOUHSC_00015, already named in some S. aureus strains GdpP (for GGDEF domain protein containing phosphodiesterase), compensates for the lack of LTA. The remaining mutations in the four other genes are possibly accessory and function to improve growth.
If disruption of gdpP compensates for a lack of LTA, one would predict that introducing a WT copy of gdpP into a suppressor strain should be lethal, while the introduction of any of the mutant gdpP alleles present in the suppressor strains should not prevent growth. This was indeed the case, and the expression of WT gdpP from an anhydrotetracycline (Atet) inducible promoter containing plasmid but not the expression of the mutant gdpP alleles obtained from suppressor strains 4S4 (inverted sequence, inframe), 4S5 (point mutation) or 4N2 (stop codon) prevented growth of strain 4S5 (Figure 2B). As controls, introduction of the empty vector pCN34iTET or uninduced pCN34iTET-gdpP into the suppressor strain 4S5 had no effect on bacterial growth (Figure 2A, left panel) and expression of GdpP in WT SEJ1 also had no effect on growth, demonstrating that expression of GdpP is not toxic per se (Figure 2A, right panel). These results provide further evidence that disruption of gdpP is essential for the survival of LTA-negative S. aureus suppressor strains. However, the suppressor strains contain other known mutations (see text and Table 1), therefore we set out to recreate the system in an S. aureus strain without these additional mutations. To this end, the S. aureus strain SEJ1ΔgdpP-iltaS with a silent gdpP deletion and IPTG inducible ltaS expression was created (Figure 3A). Normally, the growth of an inducible ltaS strain is dependent on IPTG [13], however this should no longer be the case upon deletion of gdpP and expression of GdpP from a plasmid should restore IPTG-dependent growth to this strain. S. aureus strains SEJ1-iltaS pCN34 (iltaS pCN34) and SEJ1ΔgdpP-iltaS containing the GdpP expression plasmid (ΔgdpP iltaS pCN34iTET-gdpP) were used to test this experimentally. All strains grew in the presence of IPTG (Figure 3C, black filled symbols), produced LTA (Figure S3A in Text S1) and displayed a normal cell shape and placement of division septa (Figure 3B and S3B in Text S1, panels i, iv & vii). It is interesting to note that SEJ1ΔgdpP-iltaS bacteria, which contain a deletion of the gdpP gene, appear to have a reduced cell size (Figure 3B, iv & vii) and this will be analyzed in more detail later on. In the absence of IPTG, LTA was no longer produced (Figure S3A in Text S1) and as expected growth of the inducible ltaS strain ceased (Figure 3C, iltaS pCN34 – white and grey filled squares). Morphologically the cells appeared enlarged, clumped and showed aberrant cell division (Figure 3B and S3B in Text S1, panels ii & iii). However, as predicted growth of the inducible ltaS strain with the gdpP deletion continued even in the absence of IPTG (Figure 3C, ΔgdpP-iltaS pCN38 – white and grey filled triangles) and the morphology of the cells was much improved showing more regular cell sizes and division (Figure 3B and S3B in Text S1, panels v, vi & viii). Addition of Atet (grey symbols) for the expression of GdpP from the complementation vector restored the IPTG-dependent growth phenotype and strain SEJ1ΔgdpP-iltaS pCN34iTET-gdpP ceased to grow in the absence of IPTG (Figure 3C, ΔgdpP iltaS pCN34iTET-gdpP - grey circles) and cells displayed an aberrant morphology (Figure 3B and S3B in Text S1, panel ix). In summary, it can now be concluded that deleting the gdpP gene allows for the growth of an LTA-deficient strain of S. aureus.
The experiments described above confirm that deletion of gdpP compensates for the lack of LTA in an RN4220-derived S. aureus strain. However, this strain is a chemically mutagenized laboratory strain that contains known mutations and defects in regulatory systems [31], [32], [33], [34]. To determine if LTA is also important for growth of other S. aureus isolates, and whether mutations in gdpP can compensate for the lack of LTA, the ltaS gene was deleted in the erythromycin sensitive community-acquired MRSA (CA-MRSA) strain LAC* [35]. Strains LAC*ΔltaSN::erm and LAC*ΔltaSS::erm with complete ltaS deletions could be obtained on TSA plates containing 7.5% NaCl or 40% sucrose and the absence of LTA was confirmed by western blot (Figure 4A). Both LAC*ΔltaS::erm strains could initially only grow in the presence but not in the absence of osmoprotectants (Figure 4B). Identical to the RN4220 ltaS-deletion strains, LAC*ΔltaS::erm suppressor colonies could be obtained on TSA plates. These strains did not produce LTA but were now able to grow in the absence of osmoprotectants (Figure 4). To establish whether the LAC*ΔltaS::erm suppressor strains had acquired mutations within gdpP, this gene was sequenced from eight independently isolated suppressor strains. Of these eight strains, five had mutations in the gdpP gene (Table S2 in Text S1). These results show that LTA is also important for the growth of a CA-MRSA strain and indicate that, as observed in RN4220, inactivation of gdpP provides a mechanism that allows LAC* to grow in the absence of LTA.
S. aureus GdpP contains two N-terminal transmembrane helices followed by a degenerated PAS sensory domain (Pfam00989), a GGDEF domain (Pfam00990), a DHH domain (Pfam01368) and a DHH-associated DHHA1 domain (Pfam02272) (Figure 5A). GGDEF domains are usually associated with proteins containing c-di-GMP cyclase or phosphodiesterase activity [36], [37] and DHH/DHHA1 domain-containing proteins often function as phosphatases or phosphoesterases [38]. The B. subtilis protein YybT is a close homologue to S. aureus GdpP and recently it was shown that recombinant B. subtilis YybT has strong phosphodiesterase activity contained within the DHH/DHHA1 domains [39]. It was further suggested that the cyclic dinucleotide c-di-AMP is the physiological substrate and is converted to 5′-pApA by YybT [39]. In addition, it was found that the GGDEF domain of YybT has weak ATPase activity but no c-di-GMP or c-di-AMP cyclase activity [39]. To investigate whether S. aureus GdpP, like B. subtilis YybT, is a c-di-AMP phosphodiesterase an N-terminally His-tagged fragment of GdpP spanning amino acids 84–655 and containing the PAS, GGDEF and DHH/DHHA1 domains was expressed and purified from E. coli extracts (Figures 5A and S4 in Text S1). Incubation of c-di-AMP with the recombinant rGdpP84–655 protein resulted in the complete conversion of c-di-AMP to 5′-pApA. This was initially determined by mass spectrometry analysis (data not shown) and subsequently quantified by separating reaction products by HPLC and integrating the nucleotide peak areas using c-di-AMP and 5′-pApA standards as controls (Figure 5B). Recombinant rGdpP84-301 protein containing only the PAS and GGDEF domains was also used in this assay and even when present in 4-fold higher amounts did not show any phosphodiesterase activity (Figure 5B). Therefore, identical to B. subtilis YybT, recombinant S. aureus GdpP has in vitro c-di-AMP phosphodiesterase activity and the DHH/DHHA1 domain is essential for this activity. Purified S. aureus rGdpP84–655 did not show any, or at the most very weak, ATPase activity in vitro (data not shown). However reaction conditions were not further optimized, as the phosphodiesterase activity appears to be the biologically relevant activity (see below).
The gdpP mutations present in three of the five sequenced RN4220 LTA suppressors (4N1, 4N2 and 5S4) lead to stop codons in or before the DHH/DHHA1 domain, which will automatically disrupt the phosphodiesterase activity (Figure 5A). On the other hand, the gdpP mutations in 4S4 and 4S5 lead to the expression of GdpP variants with a 24 amino acid inversion (but still in frame) and a G266D amino acid substitution. To test if these alterations affect phosphodiesterase activity, rGdpP4S4 and rGdpP4S5 variants (comprising amino acids 84–655) were produced (Figure S4 in Text S1) and the phosphodiesterase activity measured. While all of the input c-di-AMP was hydrolyzed by WT rGdpP84–655 in less than ten minutes, recombinant rGdpP4S4 and rGdpP4S5 variants converted less than 40% of the input substrate to 5′-pApA in one hour (Figure 5C). These results provide evidence that a disruption of the phosphodiesterase function of GdpP is responsible for compensating for a lack of LTA in suppressor strains.
In B. subtilis YybT, amino acids D225 and R291 in the GGDEF domain and residues D420 and D499 in the DHH/DHHA1 were identified as key residues for the weak ATPase activity and the phosphodiesterase activity, respectively [39]. Alanine substitutions at the corresponding positions D223, R289, D418 and D497 in S. aureus GdpP (Figure 5A) were made and recombinant proteins expressed and purified (Figure S4 in Text S1) for use in in vitro phosphodiesterase assays or were expressed from the Atet-inducible vector pCN34iTET in S. aureus strain 4S5 in order to investigate which activity needs to be inactivated to allow for the survival of the LTA-negative suppressor strain. As expected, recombinant rGdpPD418A and rGdpPD497A with substitutions of key residues in the DHH/DHHA1 domains lacked phosphodiesterase activity (Figure 6A). Interestingly, the rGdpPR289A variant with a substitution in the GGDEF domain also had reduced in vitro phosphodiesterase activity, while the activity of the rGdpPD223A enzyme was identical to WT rGdpP84–655 (Figure 6A). The survival of the LTA-negative suppressor correlated with the defect in phosphodiesterase activity as expression of GdpPR289A, GdpPD418A or GdpPD497A did not affect the growth of strain 4S5, while expression of GdpPD223A with WT phosphodiesterase activity prevented the growth of this strain (Figure 6B). Therefore, the ability of S. aureus to grow in the absence of LTA seems to be independent of any potential ATPase activity GdpP may have, but the survival depends on the disruption of the phosphodiesterase activity and possibly a concurrent increase in c-di-AMP concentration in the cell.
Our results thus far have shown that GdpP has in vitro c-di-AMP phosphodiesterase activity. For this activity to be of any biological significance, S. aureus needs to be able to synthesize c-di-AMP. Until now, c-di-AMP has only been described as a naturally occurring molecule in the supernatant of Listeria monocytogenes cultures and in two very recent studies in the cytoplasm of B. subtilis and Streptococcus pyogenes [40], [41], [42]. The c-di-AMP in the supernatant of L. monocytogenes was shown to be involved in the activation of an IFN-β-mediated host immune response [40]. In the same study, a L. monocytogenes protein thought to be essential for bacterial growth and containing a so called DisA_N or DAC domain (Pfam02457) was implicated in the production of c-di-AMP and termed DacA for diadenylate cyclase A [40]. The S. aureus membrane protein SAOUHSC_02407 is homologous to this L. monocytogenes protein and the only S. aureus protein that contains a DisA_N domain. Of note, this gene is also one of the five genes found to contain mutations within two of the SEJ1ΔltaS suppressor strains (Table 1).
To investigate if SAOUHSC_02407 is a c-di-AMP cyclase, this gene was cloned and expressed in E. coli, which cannot naturally produce c-di-AMP. E. coli extracts were prepared and analyzed by LC-MS/MS as previously described for the detection of c-di-GMP [43]. While no c-di-AMP was detected in E. coli extracts isolated from a strain containing the empty plasmid pET28b, extremely high levels of more than 2800 ng c-di-AMP/mg E. coli protein were detected in extracts isolated from the strain expressing SAOUHSC_02407 (Figure 7A). This provides experimental evidence that the S. aureus protein SAOUHSC_02407, renamed DacA, is a c-di-AMP cyclase and that S. aureus should be capable of producing c-di-AMP.
To address experimentally if c-di-AMP is produced by S. aureus and to investigate the involvement of GdpP in adjusting nucleotide levels, cytoplasmic extracts were prepared from cultures of strains SEJ1 and the gdpP mutant SEJ1ΔgdpP::kan. C-di-AMP could be readily detected in samples isolated from both strains using an LC-MS/MS method [43]. To accurately quantify c-di-AMP levels, samples were spiked before extraction with a 13C15N isotope-labeled version of c-di-AMP of known concentration and amounts were quantified based on a c-di-AMP standard curve (Figure S5 in Text S1). A c-di-AMP concentration of 3.33±0.44 ng/mg bacterial dry weight was determined for strain SEJ1 and upon deletion of gdpP the c-di-AMP levels increased more than 13-fold to 45.37±3.83 ng/mg bacterial dry weight (Figure 7B). An increase in c-di-AMP concentration of a similar magnitude was also observed when samples were prepared from the LTA-negative suppressor strain 4S5 (Figure 7B). Furthermore, c-di-AMP was also detected in cytoplasmic extracts of the CA-MRSA strain LAC* and again the c-di-AMP levels increased more than 11-fold in the isogenic gdpP deletion strain LAC*ΔgdpP::kan (Figure 7B). These results establish for the first time the presence of the cyclic dinucleotide c-di-AMP in S. aureus, and provide direct evidence that the S. aureus GdpP enzyme is a c-di-AMP specific phosphodiesterase in vivo. In addition, these results demonstrate that S. aureus strains that lack LTA respond by increasing cellular levels of the secondary messenger c-di-AMP. It is also important to note that no quantifiable amounts of c-di-GMP could be detected in S. aureus extracts as judged by LC-MS/MS. This is consistent with the findings by Holland et al. [44], that GdpS (GGDEF domain protein from Staphylococcus), the only staphylococcal protein with a potentially intact GGDEF domain, is unable to synthesize c-di-GMP. Taken together, our findings indicate that c-di-GMP is absent in S. aureus but underscores the importance of c-di-AMP as a secondary messenger in S. aureus.
To examine whether a deletion of gdpP, and the concomitant increase in c-di-AMP, directly affects cell wall properties and could in that way compensate for the lack of LTA, the cell wall characteristics of the mutant strains were examined. Firstly the level of hydrolytic enzymes present on the cell surface of LAC* and the isogenic gdpP mutant was analyzed. The gdpP mutant strain possessed increased amounts of autolysins as judged by zymographic analysis (Figure S6 in Text S1).
Next LAC* and the isogenic LAC*ΔgdpP::kan mutant strain were incubated with increasing concentrations of the cell wall or membrane targeting antimicrobials oxacillin, penicillin G, vancomycin, lysostaphin, daptomycin and nisin. LAC*ΔgdpP::kan displayed an approximately eight-fold decreased susceptibility to oxacillin and lysostaphin and a more than 32-fold decreased susceptibility to penicillin G (Table 2), indicating that changes in the cell wall/peptidoglycan structure have occurred. To investigate this further, peptidoglycan was purified from the WT and the gdpP mutant LAC* strains, digested with mutanolysin and the muropeptides analyzed by HPLC. The overall muropeptide profile of the two strains was very similar (Figure S7A and B in Text S1), however a statistically significant reduction in the amount of monomeric muropeptides and a concurrent increase in higher cross-linked muropeptides (trimers and above) was observed in the gdpP mutant strain (Figure 8A). This increase in the amount of cross-linked peptidoglycan could potentially be responsible for the increased resistance to cell wall targeting antimicrobials observed in this strain and could aid with bacterial survival in the absence of LTA by strengthening the cell wall, however this is speculative and remains to be confirmed.
Following this, the WT LAC* strain and the isogenic LAC*ΔgdpP::kan strain were grown to mid-log phase, stained with BODIPY-vancomycin and observed under the microscope. This analysis revealed a reduction of cell size by more than 13% for the gdpP mutant, which could be restored to the WT cell size upon complementation with gdpP (Figure 8B and S7C in Text S1). Based on this microscopic analysis, average diameters of 1.188±0.025 µm and 0.967±0.015 µm were determined for the LAC* and LAC*ΔgdpP::kan strains, respectively (Table 3). A similar reduction in cell size was observed for the gdpP mutant in the RN4220 strain background, where cell diameters of 1.138±0.081 µm for SEJ1 and 0.879±0.051 µm for the SEJ1ΔgdpP::kan strain were measured. These results indicate a function for c-di-AMP in controlling the cell size of S. aureus with increased cyclic dinucleotide levels leading to a statistically significant decrease in cell size.
Based on the experimentally determined cell diameters and CFU counts for cultures used to prepare the cytoplasmic extracts to detect c-di-AMP (see above Figure 7B) intracellular cyclic dinucleotide concentrations of 2.8±0.6 µM for SEJ1 and 2.1±0.3 µM for the LAC* could be calculated, which increased approximately 15-fold to 42.9±9.0 and 31.5±4.5 µM in the isogenic gdpP mutant strains (Table 3).
In a variety of bacteria an increase in intracellular levels of the related and well-characterized cyclic dinucleotide c-di-GMP, stimulates the biosynthesis of adhesins, promotes biofilm formation and inhibits various forms of motility [45]. No such functions have yet been ascribed to c-di-AMP.
To investigate whether cellular levels of c-di-AMP affect the ability of S. aureus to form a biofilm, cultures of WT SEJ1 and both the silent gdpP mutant SEJ1ΔgdpP, and the marked gdpP mutant SEJ1ΔgdpP::kan, were grown without shaking in 96-well plates containing BHI 4% NaCl. Staining of adherent bacteria with crystal violet revealed that the gdpP deletion strains formed approximately 3-times more biofilm than the WT control strain (Figure S8 in Text S1). However, it should be noted that neither the WT nor the gdpP mutant LAC* strains formed robust biofilms under these conditions. Nevertheless, this indicates that increased cellular levels of c-di-AMP not only affect cell properties that allow bacteria to grow in the absence of LTA but, like c-di-GMP, also influence the production of components involved in biofilm formation at least in some S. aureus background strains.
Clear functions for the secondary messenger molecule c-di-GMP, in controlling gene expression and the switch from planktonic to sedentary lifestyles, have been established in a diverse range of bacterial species [45]. It is now well documented that this cyclic dinucleotide plays an important role in controlling biofilm formation and virulence gene expression in a range of bacteria, including important human pathogens such as Pseudomonas aeruginosa [46]. Recently, it has also been suggested that this signaling molecule, which is widespread in bacterial species but apparently not found in higher eukaryotes, can act as a danger signal in eukaryotic cells prompting studies on the immunomodulatory and immunostimulatory properties of c-di-GMP [47], [48]. On the other hand, until very recently c-di-AMP had not been recognized as a naturally occurring molecule in any living organism. c-di-AMP was noticed for the first time in 2008 during crystallization studies of the B. subtilis DNA binding protein DisA [49], [50]. Additional work confirmed that the N-terminal part of DisA (DisA_N domain) is capable of synthesizing c-di-AMP in vitro from two molecules of ATP [49]. Shortly afterwards, c-di-AMP phosphodiesterase activity was ascribed to the B. subtilis protein YybT using an in vitro assay system [39]. The first evidence for the production of c-di-AMP by living cells came from a study on L. monocytogenes, where this cyclic dinucleotide was detected in the culture supernatant and identified as the molecule that stimulates an IFN-β-mediated host immune response [40]. Very recently c-di-AMP has also been detected in cytoplasmic extracts from S. pyogenes and from B. subtilis [41], [42]. In this study, we have identified c-di-AMP within the cytoplasm of the Gram-positive pathogen S. aureus and also quantified the amounts produced in vivo in both laboratory and clinically relevant strains (Figure 7 and Table 3). Intracellular c-di-AMP concentrations of 2 to 3 µM were detected in the S. aureus cells, which are very similar to the concentration of 1.7 µM reported for B. subtilis during vegetative growth [41]. We also show in this study that GdpP functions in vivo as a c-di-AMP-specific phosphodiesterase and provide experimental evidence that the S. aureus protein SAOUHSC_02407, which was renamed DacA, is capable of producing c-di-AMP (Figure 7 and 9). Similar to what was observed with the c-di-GMP signaling molecule, this study provides a link between this novel cyclic nucleotide and cell wall properties in Gram-positive bacteria, as an increase in c-di-AMP levels allows S. aureus to grow in the absence of LTA and the CA-MRSA gdpP mutant strain shows an increased resistance to the cell wall active antimicrobials oxacillin, penicillin G and lysostaphin (Table 2) and an increase in the amount of cross-linked peptidoglycan (Figure 8A). Our study also revealed that c-di-AMP plays a role in controlling the cell size of S. aureus (Figure 8B). Recently, Oppenheimer-Shaanan et. al have shown that c-di-AMP levels increase 3-fold in B. subtilis at the onset of sporulation and the authors suggested that this increase serves as a positive signal for sporulation to proceed [41]. Since S. aureus does not form spores, this particular function attributed to c-di-AMP cannot apply to S. aureus. However, our observation that an increase in c-di-AMP levels results in a decrease in cell size suggests perhaps a more general function for c-di-AMP in progressing the cell cycle and not just the sporulation process in Gram-positive bacteria. In this respect, it is interesting to note that, in contrast to c-di-GMP, c-di-AMP might be an essential constituent of the cell. Attempts to disrupt the c-di-AMP cyclase DacA in L. monocytogenes were unsuccessful [40]. Furthermore, screens for essential genes in Mycoplasma pulmonis, Mycoplasma genitalium, Streptococcus pneumoniae as well as S. aureus indicated that dacA is essential for cell viability [51], [52], [53], [54].
The c-di-AMP cyclase DacA and the c-di-AMP phosphodiesterase GdpP are both predicted to be anchored to the bacterial membrane and it seems likely that changes in the environment, cell wall structures or in the membrane itself serve as cues to adjust the intracellular cyclic dinucleotide levels. Bioinformatic analysis of GdpP revealed that this protein contains a highly degenerated PAS domain. PAS domains usually function as sensory domains for detecting light, redox potential, oxygen, small ligands, and the overall energy level of a cell, usually by way of an associated cofactor [55]. A recent study has shown that the PAS domain of the B. subtilis YybT protein is capable of binding the cofactor heme and that this binding suppresses the phosphodiesterase activity in vitro [56]. This indicates that the PAS domain of YybT is indeed capable of sensing environmental changes and the most likely output will be a change in c-di-AMP levels. In this regard it is interesting to note that we observed a light brown discoloration of the Ni-NTA columns during the purification process of the GdpP protein, suggesting that the S. aureus protein is also capable of binding heme.
GGDEF domains are typically associated with c-di-GMP cyclases or phosphodiesterases and function to synthesize c-di-GMP. Degenerated domains can also regulate the activity of associated c-di-GMP phosphodiesterase domains [57]. The GGDEF domain of GdpP has the highly divergent amino acid sequence SSDQF, with substitutions of three of the five highly conserved active site residues that are essential for GTP catalysis [58]. The GGDEF domain of B. subtilis YybT can bind ATP and slowly convert it to ADP [39]. However, this domain cannot synthesize c-di-AMP and the ATP binding had no effect on the in vitro phosphodiesterase activity contained within the DHH/DHHA1 domains [39]. The biological relevance of the ATP binding and hydrolysis activity of the degenerated GGDEF domain remains to be determined. Indeed, we provide experimental evidence in this study that argues against a role of any potential ATPase activity for the function of the S. aureus GdpP protein (Figure 6). At the same time our data suggest that this domain can influence, independently of any ATPase activity, the phosphodiesterase activity of the associated DHH/DHHA1 domains, as indicated by a decrease in in vitro phosphodiesterase activity of GdpP variants containing single point mutations in the GGDEF domain (Figure 6). Additional experiments are needed to determine the mechanism by which the GGDEF domain influences the activity of the downstream phosphodiesterase.
As mentioned above, c-di-GMP is typically synthesized by GGDEF domain containing proteins and a search for this domain highlighted two S. aureus proteins that could potentially function as c-di-AMP cyclases. GdpP itself, which contains amino acid alternations in crucial resides in this domain and extrapolating from the rigorous in vitro analysis on the B. subtilis homologue YybT, does not have cyclase activity. The second protein is GdpS [44]. O'Gara and coworkers investigated the possibility that the Staphylococcus epidermidis GdpS protein is a c-di-GMP cyclase but with no success and consistent with this observation we show in this study that c-di-GMP does not appear to be present in S. aureus [44]. It is possible however, that GdpS may be involved in c-di-AMP synthesis instead, a theory that is currently under investigation. In this regard it is interesting to note that, in contrast to the gdpP deletion strain analyzed in this study that causes an increase in biofilm formation (Figure S8 in Text S1), an S. epidermidis gdpS mutant strain has a defect in biofilm formation caused by a decrease in transcription of the polysaccharide-producing ica locus [44].
While the exact number of S. aureus proteins involved in c-di-AMP production and hydrolysis remains to be determined, it is certain that GdpP and DacA are involved in this process (Figure 7). Having shown that a deletion of gdpP results in increased intracellular levels of c-di-AMP the question remains as to how this suppresses the bacterial need for LTA. Rao et al., have demonstrated that the alarmone ppGpp, which is produced during the stringent response to cope with stress, inhibits the phosphodiesterase activity of YybT in vitro [39]. Furthermore, disruption of the GdpP homologues in Lactococcus lactis and B. subtilis renders these bacteria more resistant to acid stress [39], [59]. The deletion in B. subtilis also increases the sporulation efficiency of cells that had been exposed to a DNA damaging agent [39], [41]. All these observations indicate that an increase in intracellular c-di-AMP levels allows bacteria to cope better under stress conditions. Undoubtedly depleting LTA from the cell wall will place bacteria under stress. While bacteria may not naturally be able to respond to such a stress, one mechanism that allows bacteria to survive is by irreversibly disrupting the function of GdpP and in this manner increasing intracellular c-di-AMP levels (Figure 7B). By analogy with c-di-GMP, we assume that c-di-AMP then acts as a secondary messenger to up- or down regulate the activity or expression of a certain set of target proteins (Figure 9) [45]. Alternatively, c-di-AMP might also bind to RNA molecules to affect protein expression, as in the case of c-di-GMP-dependent riboswitches [60]. As an increase in c-di-AMP concentration in LTA depleted cells has served to improve cell growth and division (Figure 3) it is tempting to speculate that c-di-AMP is involved in regulating some components of the cell division machinery. A function for c-di-AMP in controlling cell division in S. aureus is also consistent with our observations that gdpP mutant strains are smaller in size (Figure 8B and Table 3). Therefore, an increase in intracellular c-di-AMP appears to allow S. aureus cells to initiate cell division before they have reached their normal size. Furthermore, the increased resistance of a gdpP mutant LAC* strain to lysostaphin, oxacillin and penicillin G and the decrease in monomeric peptidoglycan subunits provides experimental evidence that proteins involved in peptidoglycan synthesis are regulated by c-di-AMP. It was also noted that three LAC* suppressor strains contained WT gdpP sequences (Table S2 in Text S1). It is possible that in these strains the molecular targets of c-di-AMP are mutated instead of altering the concentration of the second messenger molecule. While beyond the scope of this study, it will be interesting to identify c-di-AMP target proteins in future studies, in particular those involved in helping the bacteria cope with stress. One could suppose that by interfering with c-di-AMP synthesis cells may be less able to respond and cope with the stresses encountered during infection. By identifying a way to combine this inhibition with currently available drug treatments, it could be possible to control even the most multi-drug resistant S. aureus infections.
Inactivation of GdpP was the first step that allowed all five sequenced suppressor strains to grow in the absence of LTA, however several of the mutations acquired in subsequent steps are interesting to note. Two suppressor strains acquired mutations that lead to amino acid substitutions in the c-di-AMP cyclase DacA (Table 1). These substitutions might function to improve cyclase activity and, analogous to a gdpP deletion, this may result in an overall increase in c-di-AMP levels. It is also plausible that these mutations function to reduce the drastically increased levels of c-di-AMP observed in suppressor strains, levels which could prove toxic if not regulated properly. A third suppressor mutation resulted in an amino acid substitution in the succinate dehydrogenase SdhA (SAOUHSC_01104). SdhA is a TCA cycle enzyme and it is hard to predict how alterations in a central metabolism enzyme could help LTA-deficient bacteria to survive. The remaining two mutated genes, encode for a putative permease (SAOUHSC_01358) and a conserved hypothetical protein with sequence homology to a fusaric acid transporter (SAOUHSC_02001) (Table 1). Premature stop codons have arisen within these two proteins presumably inactivating their function. Woodward et al have shown that over-expression of multidrug resistance transporters results in increased secretion of c-di-AMP into the supernatant of L. monocytogenes cultures [40]. It is therefore plausible that disrupting the functions of these two putative S. aureus transporters may again serve to increase intracellular c-di-AMP levels.
Eukaryotic host cells often recognize essential bacterial cell components as a means of detecting an infection. One well-studied example of this is the recognition of peptidoglycan by the intracellular host proteins Nod1 and Nod2, which upon detection results in the activation of the NF-κΒ pathway and an immune response [61]. Recently, Woodward and coworkers discovered that c-di-AMP, which is released by L. monocytogenes inside host cells, is also detected in the cytosol where it triggers a host immune response [40]. S. aureus is another pathogen that is capable of invading epithelial and endothelial cells through fibronectin binding protein-mediated adherence to the host cell integrin α5β1 and subsequent endocytosis [62]. Once inside the cell it is equally likely that S. aureus secretes this potentially essential nucleotide into the cytosol, a possibility that may then be exploited by the host immune system. The eukaryotic proteins involved in the detection of c-di-AMP and the mechanism leading to immune activation remains to be determined as our current understanding is only rudimentary [63].
Taken together we have shown that it is possible to create viable LTA-negative strains of S. aureus that compensate for the loss of this important polymer by increasing intracellular levels of the secondary messenger c-di-AMP. The unambivalent identification of c-di-AMP in the cytoplasm of S. aureus and the ability to regulate its level opens up the exciting possibility of identifying target proteins or other compounds through which this cyclic dinucleotide exerts its function and regulates cellular processes. This is especially intriguing for Gram-positive pathogens such as S. aureus, S. pneumonia, S. pyogenes and L. monocytogenes considering that c-di-AMP may be essential for cell viability. Further studies are required to fully elucidate the role of this messenger in the cell and to ultimately discover what targets are altered in response to a deletion of LTA. This will hopefully help us to more fully understand the biological significance of this polymer and to identify novel essential cellular survival mechanisms that could be exploited as therapeutic drug targets.
Strains used in this study are listed in Table S3 in Text S1 and primers used for cloning in Table S4 in Text S1. E. coli and B. subtilis strains were grown in LB and S. aureus strains were grown in TSB medium at 37°C with aeration, if not otherwise stated. When required, media were supplemented with antibiotics and inducers as indicated in Table S3 in Text S1. Details on plasmid and strain constructions are provided in the supplementary Materials and Methods section in Text S1.
WT and LTA suppressor strains were grown overnight in TSB medium. Unsuppressed ΔltaS strains were grown in TSB containing either 7.5% NaCl or 40% sucrose. Overnight cultures were washed three times in the appropriate medium and diluted to a starting OD600 of 0.05. Cultures were incubated at 37°C with aeration and OD600 values determined at 2 h intervals. Cultures containing ltaS under IPTG inducible control were grown overnight in the presence of IPTG and the appropriate antibiotics. Bacteria were washed three times in TSB and diluted to an OD600 of 0.05 in 5 ml TSB with or without 1 mM IPTG and 100 ng/ml Atet as appropriate and OD600 values determined. Where stated, the cultures were diluted after 4 h 1∶100 into fresh medium with the appropriate antibiotics and inducers to maintain cultures in the exponential growth phase and growth continued for a further 6 h. The 4 h time point is then represented as T = 0. Growth curves were performed in triplicate and representative graphs are shown. CFUs per ml culture were determined by emulsifying a colony in 1 ml TSB, normalizing the OD600 to 0.05, performing serial dilutions and plating 100 µl on the appropriate plates. Plates were then incubated at 30°C or 37°C as indicated and colonies were enumerated after overnight growth. Counts were performed twice with representative figures stated in the text.
For whole genome sequence determination, chromosomal DNA was isolated from the S. aureus reference strain SEJ1 and the 5 suppressor strains 4S4, 4S5, 4N1, 4N2 and 5S4. Sequences were determined using a SOLiD 3 System (Applied Biosciences) and DNA fragment libraries. Fifty bp fragment libraries were generated by mechanical shearing, treated and coupled to beads using emulsion PCR and deposited on glass slides following protocols supplied by Applied Biosciences. The generated reads resulted in genomic coverage of between 140x and 213x and have been submitted to the ENA Sequence Read Archive (SRA) under accession ERP000528 (http://www.ebi.ac.uk/ena/data/view/ERP000528). Reads from SEJ1 were aligned to the known genome sequence of S. aureus strain NCTC8325 with the BioScope mapping pipeline (Applied Biosystems) using an anchor length of 25 bp allowing 2 mismatches. Next the reads from the suppressor strains were individually aligned to NCTC8325 using the same alignment procedure. Using these analysis tools, large DNA deletions in place of the prophage sequences and the spa gene as well as 84 point mutations with high confidence scores were detected in all strains (SEJ1 and the 5 suppressor strains) when compared to the NCTC8325 genome sequence. Of these 84 mutations, 76 were identical to those listed in the recently published paper on the RN4220 genome sequence [34]. Of the 8 that differed, 7 were located in a highly repetitive genome region and are likely to represent misalignments of reads. Of the extra 45 snps identified by Nair et. al., [34], a number are caused by insertion or deletion events, mutations we could not detect because of limitations of the analysis software and due to the use of a fragment library rather than using mate-pairs. Ten additional mutations, identified using the BioScope diBayes pipeline (Applied Biosystems) and ‘high’ call stringency setting, were present in the suppressors strains but absent in strain SEJ1 and these mutations are described in detail in the results section. Suppressor strains specific differences were verified by re-sequencing the genes in question at the MRC Clinical Science Centre Sequencing Facility at Imperial College London.
These assays were performed using standard procedures and details can be found in the supplementary Materials and Methods section in Text S1.
His-tagged S. aureus GdpP protein variants and the B. subtilis DisA protein were purified from 1 to 2 liter induced E. coli BL21(DE3) cultures containing the respective pET28b expression vectors (Table S3 in Text S1). Protein induction and purification was performed as described in Rao et al. with minor modifications and details can be found in Text S1 [39].
The enzymatic activities of the different purified GdpP protein variants was determined by incubating 20 µM c-di-AMP (BioLog) with 1 µM purified protein in buffer containing 50 mM Tris pH 8.5, 20 mM KCl and 0.1 mM MnCl2. Enzyme reactions were stopped at the indicated time by the addition of an equal volume of 0.1 M EDTA pH 8 and incubation for 3 to 5 min at 95°C. Enzymatic activity was determined by separating 15 µl of the reaction mixtures by HPLC (Agilent LC1200) using a Luna 150×2, 3 µm particle size RP C-18 column and a 0.1 M triethylamine acetic acid pH 6.1 (Buffer A) and 80% acetonitrile containing 20% buffer A (Buffer B) solvent system. The column temperature was set to 35°C and the flow rate to 0.25 ml/min and a constant buffer B concentration of 5% was used for the runs. Nucleotides were detected at A254 and authentic c-di-AMP and 5′-pApA (BioLog) were used as standards to determine nucleotide specific retention times.% c-di-AMP hydrolysis was calculated based on integrated nucleotide peak areas. Four independent experiments were performed (using proteins from two separate purifications) and the average and standard deviation of all four values is plotted.
Overnight cultures of S. aureus cells were diluted to a starting OD600 of 0.05 and grown for 4 h at 37°C with aeration. Cultures were adjusted to an approximate OD600 of 2, CFU counts determined by plating appropriate dilutions on TSA plates and bacteria from a 10 ml culture aliquot were also collected by centrifugation, washed and freeze dried to determine the dry weight, for normalization purposes. A 5 ml aliquot from the same culture was removed and bacteria collected by centrifugation at 9,000×g for 5 min. The pellet was suspended in 1 ml ice-cold extraction buffer (acetonitrile/MeOH/H2O - 40∶40∶20; LC-MS gradient grade, VWR) containing 0.58 µM internal 13C15N isotope labeled c-di-AMP standard. Samples were snap frozen with liquid N2 for 15 sec before being heated to 95°C for 10 min. Samples were mixed with 0.5 ml of 0.1 mm glass beads and lysed in a Fast-Prep machine 2 x for 45 sec at setting 6 (FP120, MP Biomedicals, LLC). Glass beads were separated by centrifugation at 17,000×g for 5 min at 4°C. The supernatant was removed and stored at 4°C and the remaining glass beads/cell debris mixture was washed with 1 ml extraction buffer without internal standard, incubated on ice for 15 min and again lysed. Samples were once again spun and the supernatant combined with the previous one. Glass beads were washed with 1 ml extraction buffer, incubated on ice for 15 min and centrifuged. All supernatants were combined and samples dried at 40°C under a stream of N2 and stored at −80°C. E. coli overnight cultures were diluted 1∶100 into fresh LB medium and grown at 37°C to an OD600 of 0.5 at which point 1 mM IPTG was added and cultures were grown for a further 3 h. For normalization purposes, culture aliquots corresponding to an OD600 of 2 were withdrawn, washed once in PBS pH 7.4, suspended in 800 µl 0.1 M NaOH and heated to 95°C for 15 min. The samples were centrifuged for 5 min at 17,000×g and the protein content of the supernatant was determined using a BCA assay kit (Pierce). Aliquots corresponding to an OD600 of 20 were withdrawn from the same cultures and centrifuged for 5 min at 9,000×g. The pellet was suspended in 300 µl ice-cold extraction buffer (acetonitrile/MeOH/H2O; 40∶40∶20) containing cXMP (BioLog) as an internal standard at a final concentration of 200 ng/ml. Samples were frozen with liquid nitrogen for 15 sec and afterwards heated to 95°C for 10 min. Samples were centrifuged at 17,000×g for 5 min at 4°C and the supernatant removed. The pellet was suspended in 200 µl extraction buffer without cXMP, incubated on ice for 15 min and centrifuged. The supernatant was combined with the previous one and the pellet was once again suspended in 200 µl extraction buffer followed by a 15 min incubation on ice and a centrifugation step. All supernatants were combined and dried at 40°C under a flow of N2. The c-di-AMP concentration was determined based on a c-di-AMP standard curve and values are presented as ng c-di-AMP/mg E. coli protein.
For the synthesis of 13C15N isotope labeled c-di-AMP, 1 µM rDisA was incubated with 500 µM adenosine-13C15N-5′-triphosphate in 10 mM Tris-HCl, pH 7.5, 10 mM MgCl2 and 0.1% BSA for 18 h at 30°C with gentle mixing (300 rpm). Based on an identical experiment using unlabeled ATP as the substrate, it was deduced that substrate turnover is complete under these conditions. The reaction was stopped by heating samples to 95°C for 15 min and the suspension was clarified by centrifugation (20,800×g; 10 min; 4°C). The concentration of 13C15N-c-di-AMP in the supernatant was determined by measuring the absorption at 259 nm (ε259 nm = 30,000 M−1cm−1). Further purification steps were not necessary for the use of 13C15N-c-di-AMP as an internal standard and were therefore omitted.
c-di-AMP was detected and quantified by LC-MS/MS using a similar protocol as published previously for the detection of c-di-GMP [43]. Details are given in the supplementary Materials and Methods section in Text S1.
For muropeptide analysis, 1 liter TSB medium was inoculated with overnight cultures of strains LAC* (AH1263) or LAC*ΔgdpP::kan (ANG1961) to an OD600 of 0.06. The cultures were grown at 37°C to mid-log phase (OD600 of approximately 1.5), cooled on ice and bacteria subsequently collected by centrifugation. Peptidoglycan was purified using a well-established method, which we have used previously [64], [65]. Next, 5 mg purified peptidoglycan was digested in a final volume of 1.24 ml in 12.5 mM phosphate buffer pH 5.9 with 50 µg mutanolysin from Streptomyces globisporus (Sigma) for 20 h at 37°C. The suspension was subsequently boiled for 5 min and the insoluble material removed by centrifugation for 5 min at 17,000×g and the soluble fraction stored at 4 or −20°C. Immediately before HPLC analysis, a portion of the soluble muropeptide fraction was mixed with an equal volume of 0.5 M Na borate buffer pH 9 and reduced for 30 min at RT by the addition of sodium tetraborohydride. Subsequently the pH was adjusted to 2–3 with 20% phosphoric acid and 100 µl was analyzed by HPLC using a 3 µm-particle-size 120A pore size octyldecyl silane Hypersil 250×4.6 mm C18 column equipped with a 10×4 mm guard column made of the same material (Thermos Electron Corporation). The column temperature was set to 52°C and a sodium phosphate/methanol buffer system and gradient conditions were used as previously described with the exception that the sodium azide was omitted from buffer A [64], [66]. HPLC traces were recorded at 205 nm and the muropeptide profile from three independently grown cultures was determined and a representative graph is shown. For quantification purposes, the area of muropeptide peaks was integrated and quantified using the Agilent Technology ChemStation software and shown as percentage of the total (min 20 and 145) peak area and the average values and standard deviations of the three experiments is given. The two-tailed two sample equal variance Student's t-test was used to determine statistically significant differences between muropeptide peak areas and statistically significant differences with p-values below 0.05 are indicated with an asterisk (*).
S. aureus NCTC8325: ltaS - SAOUHSC_00728; gdpP - SAOUHSC_00015; dacA - SAOUHSC_02407; gdpS - SAOUHSC_00760; gene containing suppressor mutation 2 - SAOUHSC_01104; mutation 3 - SAOUHSC_01358; mutation 4 - SAOUHSC_02001; atl – SAOUHSC_00994; B. subtilis str.168: yybT - BSU40510; disA - BSU00880; L. monocytogenes str. EGD-e: dacA – lmo2120; Streptococcus pyogenes SF370: dacA - Spy1036
|
10.1371/journal.ppat.1003369 | Identification of Fibroblast Growth Factor Receptor 3 (FGFR3) as a Protein Receptor for Botulinum Neurotoxin Serotype A (BoNT/A) | Botulinum neurotoxin serotype A (BoNT/A) causes transient muscle paralysis by entering motor nerve terminals (MNTs) where it cleaves the SNARE protein Synaptosomal-associated protein 25 (SNAP25206) to yield SNAP25197. Cleavage of SNAP25 results in blockage of synaptic vesicle fusion and inhibition of the release of acetylcholine. The specific uptake of BoNT/A into pre-synaptic nerve terminals is a tightly controlled multistep process, involving a combination of high and low affinity receptors. Interestingly, the C-terminal binding domain region of BoNT/A, HC/A, is homologous to fibroblast growth factors (FGFs), making it a possible ligand for Fibroblast Growth Factor Receptors (FGFRs). Here we present data supporting the identification of Fibroblast Growth Factor Receptor 3 (FGFR3) as a high affinity receptor for BoNT/A in neuronal cells. HC/A binds with high affinity to the two extra-cellular loops of FGFR3 and acts similar to an agonist ligand for FGFR3, resulting in phosphorylation of the receptor. Native ligands for FGFR3; FGF1, FGF2, and FGF9 compete for binding to FGFR3 and block BoNT/A cellular uptake. These findings show that FGFR3 plays a pivotal role in the specific uptake of BoNT/A across the cell membrane being part of a larger receptor complex involving ganglioside- and protein-protein interactions.
| Botulinum neurotoxin serotype A (BoNT/A) is one of seven neurotoxins (BoNT/A-G), produced by the bacteria Clostridium botulinum that are both poisons and versatile therapeutics. These toxins enter motor neurons where they prevent the release of acetylcholine at the neuromuscular junction. The specific uptake of BoNT/A across the neuronal cell membrane is dependent on specific receptor interactions. Binding to high density ganglioside GT1b mediates the initial binding step and via a low affinity interaction concentrates BoNT/A on the cell surface. Once anchored in the membrane, lateral movements within the plasma membrane facilitate intermolecular interactions of BoNT/A with additional lower density but higher affinity protein receptors. Here we present data supporting the identification of Fibroblast Growth Factor Receptor 3 (FGFR3) as a high affinity receptor for BoNT/A. We show that BoNT/A binds to FGFR3 with high affinity and functions as an agonist ligand for FGFR3. The identification of this novel receptor for BoNT/A represents an important advance in the understanding of the mechanism of action of BoNT/A, especially on the initial steps of neuronal uptake, and can be the basis for the development of new specific countermeasures and new BoNT/A-based therapeutics.
▸ Recombinant HC/A binds to the two extra-cellular loops of FGFR3b with a KD∼15 nM ▸ Recombinant HC/A acts as an agonist ligand for FGFR3 ▸The level of BoNT/A uptake is dependent on FGFR3 expression ▸ FGFR3 is expressed in motor nerve terminals
| Botulinum neurotoxin serotype A (BoNT/A) is produced by Clostridium botulinum and is a member of the Clostridial neurotoxin family that includes BoNT/A-G and Tetanus neurotoxin (TeNT). BoNT/A causes transient muscle paralysis by entering motor nerve terminals (MNTs) where it cleaves nine amino acids from the C-terminus of the soluble N-ethylmaleimide-sensitive factor attachment receptor (SNARE) protein SNAP25 (SNAP25206) to yield SNAP25197 [1]. Intact SNAP25 is required for neurotransmitter release and cleavage of SNAP25 disrupts exocytosis, which blocks neurotransmitter release [2]–[5]. BoNT/A has become a useful pharmacological and biological tool. Because of its high potency and specificity for pre-synaptic nerve terminals, BoNT/A at picomolar concentrations, is used to treat a wide range of neuromuscular disorders [6]–[8], pain disorders including migraine [9], and excessive sweating [10]. The key to the exceptional specificity of BoNT/A is believed to be the mechanism of uptake across the presynaptic membrane of neurons that involves a combination of low and high affinity interactions known as the double receptor model [11]–[13].
The low affinity receptor for BoNT/A is the ganglioside GT1b with a binding pocket within the C-terminal portion of the receptor binding domain [12], [14], [15]. According to the APR receptor model [13], an array of presynaptic receptors (APRs), clustered in microdomains at the presynaptic membrane, are responsible for specific uptake of neurotoxins, including BoNT/A. It is the binding to high density ganglioside GT1b that mediates the initial binding step and via a low affinity interaction concentrates BoNT/A on the cell surface. GT1b has been shown to bind BoNT/A with a KD∼200 nM in vitro [16]. Once anchored in the membrane, lateral movements within the plasma membrane facilitate intermolecular interaction of BoNT/A with additional lower density but higher affinity protein receptors, including the three isoforms of Synaptic Vesicle (SV) glycoprotein 2, SV2A (ENSG00000159164), B (ENSG00000185518) and C (ENSG00000122012) that are exposed on the outer plasma membrane after fusion of synaptic vesicles to the presynaptic membrane [17]–[22]. BoNT/A specifically recognizes the fourth luminal domain (LD4) of SV2 [17], [18]. The specific sequence in the BoNT/A binding domain that interacts with SV2 has not been identified [23]. Glycosylated SV2A, B, and C have also been identified as receptors for BoNT/F [22], [24] and glycosylated SV2A and B have been identified as receptors for BoNT/E [20]. BoNT/D was reported to enter neurons via two ganglioside binding sites, one site at a position previously identified in BoNT/A, B, E, F, and G, and the other site resembling the second ganglioside-binding pocket of TeNT [25]. Recently, BoNT/D has also been shown to use SV2 (all three isoforms) to enter hippocampal neurons, but BoNT/D bound SV2 via a mechanism distinct from BoNT/A and E [26]. Surprisingly, SV2A and SV2B have also been reported to mediate binding and entry of TeNT into central neurons [27].
Analysis of the first crystallographic structure of BoNT/A revealed a ganglioside binding site with structural homology to that within TeNT [28], [29](Figure 1A). A potential protein binding site was also identified with structural homology to basic fibroblast growth factor (FGFb or FGF2; ENSG00000138685), agglutinin, and the toxin abrin. The report of a potential FGF2 protein binding site within BoNT/A served as the basis for the identification of FGFR3 (Fibroblast Growth Factor Receptor 3, ENSG00000068078) as a receptor for BoNT/A. FGFR3 [30], [31] is one of four receptor-tyrosine kinases (FGFR1–4) that act as receptors for FGFs. FGFRs are composed of an extra-cellular ligand-binding domain consisting of three immunoglobulin-like loops (L1–L3), a transmembrane domain, and a split cytoplasmic tyrosine kinase domain. FGFRs are activated by dimerization induced by ligand binding that enables the cytoplasmic kinase domains to transphosphorylate one another at specific tyrosine residues [32]–[35]. FGFR1–3, but not FGFR4, exist in three different splice variants that differ in the C-terminal half of L3 and determine their individual ligand affinity and specificity. The splice variants are referred to as “a”, “b”, and “c” [36]–[39]. The “b” and “c” variants are expressed on the cell surface, while the “a” splice variant, which lacks the transmembrane domain [40], becomes a secreted extra-cellular FGF-binding protein [41]. Among the 22 known FGFs, FGF1 (ENSG00000113578), members of the subfamilies FGF8 (FGF8, 17, 18) and FGF9 (FGF9; ENSG00000102678, 16, 20) have been shown to function as ligands for both FGFR3b and c but with different levels of affinity. FGF2, and the subfamilies FGF4 (FGF4, 5, 6) and FGF19 (FGF19, 21, 23) have been shown to function as ligands for FGFR3c [39], [42]. Moreover, FGFs bind their receptors in the presence of one or more low affinity co-factors including heparin sulfate (HS), gangliosides, neuropilin-1, Klotho, and anosmin that function to modulate receptor activity [43]–[48].
The studies presented in this manuscript identify FGFR3 expressed in motor neurons at MNTs as a functional protein receptor for BoNT/A. The C-terminal binding domain of BoNT/A, HC/A, binds to the second and third extra-cellular ligand binding domain of FGFR3 and results in the phosphorylation of FGFR3. It is demonstrated that cellular uptake of BoNT/A is dependent on the level of FGFR3 expression. Native ligands for FGFR3; FGF1, FGF2, and FGF9 compete for binding to FGFR3 and block BoNT/A uptake in a cell-based assay. Moreover, peptides derived from the FGFR3 subtype b and c extra-cellular domain block BoNT/A uptake in neuronal cells. Both FGFR3 subtype b and c bind to rHC/A, but FGFR3b has the highest affinity with a KD∼15 nM in vitro. These data suggest that FGFR3 is a potential high affinity component of a receptor complex for BoNT/A on the presynaptic membrane.
Analysis of the BoNT/A crystal structure (Figure 1A) revealed that the HC/A subdomain has structural homology to basic fibroblast growth factor (FGF) [28]. To investigate the interaction of BoNT/A and FGFR, pull-down assays were performed with Neuro-2a and PC-12 cell lines that have been shown to take up BoNT/A with high efficacy after differentiation by serum starvation and trophic factors, and Nerve Growth Factor (NGF) for PC-12 cells [49], [50]. Figure S1A shows a representative experiment of how differentiation increases BoNT/A uptake in PC-12 cells. BoNT/A uptake was determined by treatment of cells with BoNT/A, followed by incubation and Western blot analysis of the SNAP25197 cleavage product. After optimization of differentiation and treatment conditions, differentiated Neuro-2a and PC-12 cells were treated with BoNT/A and EC50 values of 60±5 pM and 47.1±13 pM, respectively were determined (Figure S1B).
A complex containing BoNT/A and its receptor was isolated using three alternative pull-down methods, the results from two of these methods are shown here. First, Sulfo-SBED BoNT/A was used in a biotin transfer experiment to pull down the receptor from intact Neuro-2a cells. Antibodies against FGFR3 detected a band of the correct molecular weight for FGFR3 as part of a 250 kDa complex with BoNT/A. Figure S1C, D, and E demonstrate that a 250 kDa protein complex was isolated and that specific bands for BoNT/A and FGFR3 within this complex could be detected using antibodies specific to BoNT/A or FGFR3. Second, a complex containing both BoNT/A and FGFR3 was isolated from Neuro-2a cells treated with biotin labeled BoNT/A and the cross linking reagent Bis(Sulfosuccinimidyl) suberate (data not shown). Finally, the recombinant binding domain of BoNT/A, His-rHC/A, was used to pull down FGFR3 from differentiated PC-12 cell lysates without the use of cross-linking reagents, demonstrating a strong interaction (Figure 1B).
Having identified FGFR3 as a binding partner for BoNT/A, we investigated the role of FGFR3 as a functional receptor for BoNT/A. Competition experiments utilizing native ligands for FGFR3; FGF1, FGF2, and FGF9 [39], [42], demonstrated that these ligands competed for binding to FGFR3 and blocked BoNT/A uptake in a cell-based assay with differentiated Neuro-2a cells (Figure 1C–D). rHC/A, was used as a positive control and produced a strong blockade of BoNT/A uptake. As a negative control, FGF10 (ENSG00000070193), which is not a ligand for FGFR3, but closely related to the other FGF ligands tested, was used. Pre-incubation with FGF10 did not affect BoNT/A uptake. The experimental data from at least four independent experiments were compiled and fitted to a non-linear exponential decay model; Y = 100*e- CC*log(concentration). The Competition Constant (CC) for each fitted curve was calculated and demonstrated similar competition of the three FGF ligands. The data strongly suggest that BoNT/A utilizes FGFR3 to gain entry into neuronal cells since native FGFR3 ligands blocked its uptake. The hypothesis that BoNT/A acts as an agonist for FGFR3 was further supported by demonstrating that treatment with rHC/A resulted in phosphorylation of FGFR3, achieving similar levels of activation as cells treated with identical concentrations of FGF2 (Figure 1E–F).
The ligand binding site for FGFR3 has been identified as the second and third extra-cellular loops of FGFR3 (Figure 2A) [39], [51], [52]. To further verify the functional role of FGFR3 as a receptor for BoNT/A, we demonstrated that pre-incubation of BoNT/A with a peptide spanning the second and third extra-cellular loops of FGFR3b (FGFR3b Loop 2,3) inhibited BoNT/A uptake presumably via binding to the receptor binding domain of BoNT/A (Figure 2C). Inhibition, although to a lesser extent, was also observed using the peptide spanning the luminal domain (LD4) of SV2C, SV2C529–579 (Figure 2C). SV2C529–579 (Figure 2B) has previously been reported as the minimal peptide region for binding to BoNT/A [17]. As a positive control for inhibition, BoNT/A was pre-incubated with a neutralizing monoclonal antibody directed to the binding domain of BoNT/A, Anti-HC/A. As a negative control, Synaptotagmin II (aa1–20, Syt II1–20), the receptor for BoNT/B [53]–[56] was used. The experimental data from at least three independent experiments were compiled and fitted to a non-linear exponential decay model; Y = 100*e- IC*log(concentration). The Inhibition Constant (IC) for each fitted curve was calculated and demonstrated that FGFR3b Loop 2,3 inhibited BoNT/A uptake and was a more effective uptake inhibitor than SV2C529–579 (Figure 2D).
Initial experiments designed to explore the combined effect of FGFR3b Loop 2,3 and SV2C529–579 showed that the peptides bound with good affinity in vitro (data not shown). To address the question as to whether FGFR3 and SV2 interact in neurons, we performed a series of Co-IPs experiments. We tested if an antibody to FGFR3 could pull-down SV2 isoforms from a differentiated Neuro-2a cell lysate, and vice versa, antibodies to SV2 isoforms could pull-down FGFR3. An interaction between FGFR3 and SV2 was detected using the Anti-SV2B (sc-28956) antibody, which recognizes SV2B and, to a lesser extent, SV2A and SV2C. No bands were detected when using antibodies for SV2A or SV2C (data not shown). The result suggests that FGFR3 and SV2B interact in differentiated Neuro-2a cells (Figure 2E–F).
To characterize the binding of FGFR3 and SV2C to rHC/A, the binding affinity of the two receptor surrogate peptides, FGFR3b Loop 2,3 and SV2C529–579 to rHC/A was tested in a Surface Plasmon Resonance (SPR) binding assay. FGFR3b Loop 2,3 bound to rHC/A with an average KD = 15.0±3 nM, n = 4, ka = 1.77E+04 1/Ms, kd = 2.40E-04 1/s (Figure 3B–C). This is similar to what has been reported earlier upon binding of FGFR2b Loop 2,3 to FGF2, KD = 12.8±0.3 nM [57]. SV2C529–579 bound to rHC/A with an average KD = 105±6 nM, n = 3, ka = 2.34E+03 1/Ms, kd = 2.47E-04 1/s (Figure 3A,C). The difference in affinity between FGFR3b Loop 2,3 and SV2C529–579 is due to a 10 times faster association, ka is estimated to be 10 times higher for FGFR3b Loop 2,3 versus SV2C529–579. It can be seen on the curves as a more shallow slope and longer time to equilibrium for SV2C versus FGFR3b Loop 2,3 (Figure 3A versus 3B).
In order to compare the binding affinity of FGFR3b Loop 2,3 to rHC/A to the binding affinity of native ligands for FGFR3, the binding affinity of FGFR3b Loop 2,3 to FGF2 and FGF9 was also measured in the SPR binding assay. FGFR3b Loop 2,3 bound to FGF2 with an average KD = 12.3±4 nM, n = 3, ka = 1.65E+04 1/Ms, Kd = 1.59E-04 1/s. FGFR3b Loop 2,3 bound to FGF9 with an average KD = 31.2±1 nM, n = 3, ka = 2.92E+03 1/Ms, kd = 9.25E-05 1/s (Figure 3C–E).
Having identified rHC/A as an agonist ligand for FGFR3 (Figure 1E–F) and shown that rHC/A binds to FGFR3b Loop 2,3 in vitro with similar affinity as native ligands for FGFR3, we evaluated if FGFR3 would facilitate uptake of rHC/A and native ligands in a similar fashion. We utilized HEK 293 cells as a model system, because they express FGFR3 but no measurable levels of any of the SV2 isoforms (Figure S1G). Consequently, uptake of rHC/A via SV2 should be absent in these cells. HEK 293 cells do not express SNAP25 and therefore SNAP25 cleavage could not be used as a measure for BoNT/A uptake. Instead, uptake was measured as an increase in intracellular fluorescence after addition of fluorescently labeled rHC/A or FGF2, a native ligand for FGFR3. The results showed slightly less uptake of rHC/A compared to FGF2, but similar kinetics (Figure S1H). The slightly higher uptake of FGF2 compared to rHC/A could be due to FGF2 having more receptor targets, since it is a general ligand for FGFRs. These data suggest that FGFR3 can mediate BoNT/A uptake independently of SV2.
To explore the binding sites of FGFR3b Loop 2,3 and SV2C529–579 on the binding domain of BoNT/A, we performed a series of dual binding experiments using the BIAcore. We tested if the peptides and anti-HC/A, the neutralizing monoclonal antibody previously used in the cell based inhibition assay (Figure 2C & D), could bind to rHC/A simultaneously. rHC/A was captured by anti-HC/A monoclonal and FGFR3b Loop 2,3 or SV2C529–579 were flowed across. The results show that binding of anti-HC/A monoclonal blocks binding of FGFR3b Loop 2,3, but not SV2C529–579 to rHC/A in vitro (Figure 3F–G), demonstrating that FGFR3 and SV2 bind to different sites on the BoNT/A binding domain. Interestingly, these data also suggest that inhibition of BoNT/A uptake by the neutralizing monoclonal anti-HC/A antibody in the cell based assay is due to blockage of FGFR3 binding.
These results demonstrate that FGFR3b Loop 2,3 and SV2C529–579 can both inhibit the activity of BoNT/A in a cell-based assay, but FGFR3b Loop 2,3 is a stronger inhibitor than SV2C529–579 . They show that in an in vitro binding assay, the binding affinity for FGFR3b Loop 2,3 upon binding to rHC/A is higher, due to an estimated 10 times faster association, than the binding affinity for SV2C529–579 upon binding to rHC/A. The binding affinity for FGFR3b Loop 2,3 to rHC/A, is similar or identical, to the binding affinity for FGFR3b Loop 2,3 upon binding to FGF2 and FGF9, two native ligands for FGFR3. Also, uptake of rHC/A in HEK 293 cells, that express FGFR3, but not SV2, is comparable to uptake of FGF2, supporting a case for uptake of BoNT/A via FGFR3 independent of the presence of SV2. Finally, dual in vitro binding studies using a neutralizing antibody to HC/A, show that the FGFR3 and SV2C peptides bind to rHC/A at different sites, FGFR3 at a site close to or overlapping the binding site for the anti- HC/A, and SV2C in a site distal from the anti- HC/A binding site. Different binding sites for FGFR3 and SV2 would allow a multi-receptor complex to form.
Differentiation of neuronal cells increases BoNT/A uptake (Figure S1A and B). It has been suggested that the increased sensitivity in differentiated PC-12 cells is due to increased expression of the SNAP25b subtype that is most sensitive to BoNT/A [50]. Since the increased sensitivity could also be a result of increased expression of a receptor for BoNT/A, we studied the expression of FGFR3 as well as SV2A, B, and C before and after differentiation in both Neuro-2a and PC-12 cells. FGFR3 expression levels were similar in both cell lines and the amount of FGFR3 was unchanged after differentiation. Neuro-2a cells expressed mostly SV2C, while PC-12 cells expressed all three SV2 isoforms. Surprisingly, differentiation resulted in decreased expression of SV2 isoforms in both cell lines (Figure S1F).
Assuming FGFR3 is a functional receptor for BoNT/A, one would expect overexpression of FGFR3 to result in increased binding of BoNT/A on the cell membrane. If receptor binding is a rate-limiting step, this should also result in increased sensitivity to BoNT/A. Experiments to test the sensitivity to BoNT/A were performed under non-depolarizing conditions, where the exposure of SV2 on the cell surface is presumed to be limited. Overexpression of FGFR3 in PC-12 and Neuro-2a cells increased the sensitivity to BoNT/A and produced higher efficacy (increased maximal signal), in a Western blot SNAP25197 cell-based assay, while overexpression of SV2C did not (Figure 4A–B).
FGFR3 overexpression also increased binding of transfected cell membranes to rHC/A in a SPR binding assay, while overexpression of SV2C did not (Figure S2A and C), suggesting that if there is more FGFR3 on the cell surface more BoNT/A will bind, while more SV2C does not increase BoNT/A binding. Human neuroblastoma SH-SY5Y cells were also evaluated in the SPR binding assay because they have low sensitivity to BoNT/A [58] and express very little endogenous SV2C. Even in this situation, there was no effect as a result of over expressing SV2C (Figure S2B).
We also demonstrated, utilizing shRNA, that reduced expression of FGFR3 resulted in reduced sensitivity to BoNT/A. A 65% reduction of FGFR3 protein expression resulted in a 5.7-fold decrease in potency and a ∼5-fold increase in EC50 when compared to control cells (Figure 2C). No change in the protein expression levels of either SV2A, B, or C was detected in those samples. A separate experiment with siRNAs for FGFR3 and SV2C demonstrated that a 4.2-fold reduction in SV2C mRNA resulting in a 2-fold reduction in protein levels did not cause a reduction in BoNT/A uptake (Figure S2E–F). While a 3-fold reduction of FGFR3 mRNA resulting in a 2-fold reduction in protein levels reduced sensitivity to BoNT/A causing a 3-fold shift in relative potency when compared to control cells (Figure S2D and F), confirming that, under non-depolarizing conditions, binding to FGFR3 is a rate-limiting step in BoNT/A uptake.
The interaction between FGFR3 and rHC/A was also observed in a photobleaching experiment using the FRET partners AF-488 and TMR. We detected an increase in the fluorescent signal from the AF-488 labeled rHC/A (donor) after photobleaching the TMR labeled FGFR3 (acceptor). The data shows that FGFR3 and rHC/A are proximal enough within PC-12 cells to FRET, suggesting that FGFR3 not only binds BoNT/A on the cell surface, but it is also trafficking with BoNT/A within the cells. There was little change in the fluorescence observed when the experiment was performed with TMR labeled SV2C as the acceptor (Figure 4E–G).
BoNT/A causes transient muscle paralysis through presynaptic blockade of acetylcholine release at the neuromuscular junction. If FGFR3 functions as a receptor for BoNT/A in vivo, then it would be reasonable to presume that FGFR3 should be expressed at the MNTs. The expression pattern of the FGFR3 receptor was examined on cross-sections of rat skeletal muscle to look for potential co-expression with SV2C, SNAP25, and nicotinic acetylcholine receptors (nAChRs). Overall, immuno-reactive (IR) staining for SV2C and SNAP25 were co-expressed exclusively at neuromuscular junctions (NMJs) throughout the muscle (Figure 5A, A–D). These NMJs were specifically defined by using fluorescently labeled α-bungarotoxin (α-Bgt) nAChRs. In contrast, FGFR3-IR was not only detected at NMJs, but also in extra-synaptic structures, such as myoblasts and blood vessels (Figure 5A, E). At the NMJs however, the FGFR3 staining pattern corresponded to that of SNAP25 and nAChRs (Figure 5A, E–H). To verify expression of SV2C and FGFR3 within BoNT/A sensitive NMJs, we treated rat Tibialis Anterior (TA) muscles with BoNT/A and analyzed the staining patterns for SV2C and FGFR3 together with IR-staining for cleaved SNAP25 (SNAP25197). Focusing on individual synapses, we observed overlapping patterns for SV2C-IR and SNAP25197-IR that were adjacent to the pattern of post-synaptic nAChR expression (Figure 5A, I-L). Similarly, the patterns for FGFR3-IR and SNAP25197-IR at the NMJ were overlapping and appeared adjacent to the pattern of nAChR expression (Figure 5A, M–P). Saline-treated rat muscles showed no immuno-staining for SNAP25197 (Figure 5B). These qualitative results demonstrate that FGFR3 receptors are present on MNTs and are co-expressed with SV2C and SNAP25.
We have shown that FGFR3b Loop 2,3 binds to BoNT/A with low nanomolar affinity. To further identify the binding site for BoNT/A and to test whether the two subtypes of FGFR3, subtype b and c, bound with similar affinities, we constructed eight deletion mutants of FGFR3, containing either FGFR3 Loop 1,2,3 (long or short version), Loop 2,3, or Loop 3 of both subtypes (Figure 6A). The difference between FGFR3b and FGFR3c lies in the most C-terminal part of Loop 3 (Figure S3A). All the deletion mutant FGFR3 peptides were able to inhibit BoNT/A uptake in a cell-based inhibition assay, presumably via binding to the receptor binding domain of BoNT/A and preventing binding to cells (Figure 6B–C). However, in a SPR binding assay a significantly lower affinity was observed for the peptides spanning only Loop 3 compared to the peptides spanning Loop 2,3 or Loop 1,2,3. The association (on-rate) of the peptides spanning only Loop 3 was ∼10 times lower than the on-rate of the longer peptides, while the dissociation (off-rate) was similar (Figure 6E and S3B–C). The similar off-rate can explain why the peptides are able to inhibit equally well BoNT/A binding to the receptor in the cell-based inhibition assay. In the cell-based assay sufficient time (20 min) is available for even slow associating peptides to bind and the ability to inhibit relies more on a slow dissociation. In the SPR binding assay, on the other hand, binding is observed in real time (5–10 min). Based on the lower on-rate of the Loop 3 peptides observed in the SPR binding assay, Loop 2,3, which is the binding region for native FGF ligands, was identified as the minimal optimal binding region for BoNT/A. FGFR3b bound with slightly higher affinity than FGFR3c (Figure 6D).
In this study we identified FGFR3 as a high affinity protein receptor for BoNT/A. Pull-down experiments with neuronal cells resulted in the identification of a protein complex containing BoNT/A and FGFR3. Native ligands for FGFR3; FGF1, FGF2, and FGF9 compete with rHC/A for binding to the receptor and binding of rHC/A results in phosphorylation of FGFR3, demonstrating that BoNT/A acts as an agonist ligand for FGFR3. Since ligand binding and activation of FGFRs are known to result in receptor-mediated endocytosis of both receptor and ligand [59], we propose that binding of BoNT/A to FGFR3 also results in endocytosis and that FGFR3 may mediate BoNT/A uptake in both stimulation independent and stimulation dependent manners. This hypothesis is supported by the fact that depolarization of nerve cells increases uptake (stimulation dependent), while at the same time BoNT/A uptake can take place in resting neurons (stimulation independent) [3], [4], [60]–[63]. This is also supported by the observation that the uptake, but not the initial binding step, is altered by nerve stimulation [64].
Motor neurons at MNTs take up BoNT/A with high affinity, resulting in inhibition of exocytosis and muscle paralysis. Thus, MNTs should presumably express a BoNT/A receptor(s), and our results clearly demonstrated that both FGFR3 and SV2C are present at MNTs. These data support the hypothesis that FGFR3 functions as a high affinity receptor for BoNT/A uptake, and that most likely, SV2 is only available as a receptor for BoNT/A after depolarization and vesicular exocytosis.
Using a SPR binding assay, we demonstrated that a peptide spanning the second and third extra-cellular loops of FGFR3, FGFR3b Loop 2,3, binds to rHC/A with a KD∼15 nM and that a peptide spanning the luminal domain of SV2C, SV2C529–579, binds to rHC/A with a KD∼100 nM in vitro. The observed ∼15 nM affinity for binding of rHC/A to FGFR3b Loop 2,3 was similar or identical to the affinity for binding of two native ligands for FGFR3, FGF2 and FGF9 to FGFR3b Loop 2,3 in the same assay. Also, comparable uptake of rHC/A and FGF2 was observed in HEK 293 cells, a cell line that express FGFR3 and not any of the SV2 isoforms, suggesting that FGFR3 can mediate uptake of BoNT/A independently of SV2.
A recent publication [63] clearly supports our findings. The authors observed limited co-localization of SV2C and HC/A or BoNT/A after treatment of spinal cord motor neurons under resting conditions and this co-localization did not significantly increase under depolarizing conditions. Moreover, inhibition of exocytosis by pre-treatment with BoNT/D did not prevent the internalization of HC/A. The authors concluded that BoNT/A may exploit an alternative pathway(s), largely independent of stimulated synaptic endo-exocytosis, to enter neuronal cells in both resting and depolarizing conditions.
Pre-incubation of BoNT/A with the FGFR3 and SV2C peptides before treatment of cells, blocked uptake in neuronal cells, presumably by interacting with the binding domain of BoNT/A and preventing binding to the receptor on cells. In accordance with the observed lower affinity for SV2C529–579 compared to Loop 2,3 of FGFR3, the FGFR3 peptide produced a stronger blockade than the SV2C peptide. These data suggest that FGFR3 may function as a high affinity receptor for BoNT/A and that SV2C may function as a medium affinity receptor.
SPR experiments demonstrated that the FGFR3b and SV2C peptides bind to different sites on HC/A, FGFR3 in a site overlapping the epitope of a neutralizing monoclonal antibody to HC/A (6B1, provided by Dr. L. Smith, USAMRIID) and SV2C in a site distal from both. So far the binding site for SV2 has not been identified, but it has been suggested that SV2 binds to the C-terminal half of the binding domain, HCC, similar to how Synaptotagmin II binds to BoNT/B [23]. Different binding sites for FGFR3 and SV2 on the binding domain of BoNT/A would allow formation of a multi-receptor complex. Interestingly, Co-IP experiments show that FGFR3 and SV2 can interact in live Neuro-2a cells, suggesting a step-wise binding and/or formation of a multi-receptor BoNT/A complex.
We concur with others in the field that the specificity of BoNT/A for neuronal cells and specially motor neurons, which is higher than binding to a single receptor can explain, is due to the fact that uptake of BoNT/A is a multi-step process involving at least two crucial steps [13], [23]. The first crucial step is binding to gangliosides like GT1b (KD∼200 nM) that are abundantly present in the outer leaflet of the plasma membrane of neuronal cells. This initial step increases the local concentration of BoNT/A and allows it to diffuse in the plane of the membrane to bind its protein receptor(s) [15], [16], similar to what has been observed for heparin sulfate and FGF2 [65]. BoNT/A that is diffusing within microdomains of the plasma membrane will be presented to FGFR3 and/or SV2, bind to the receptor, and undergo endocytosis representing a second crucial step.
There are several lines of evidence suggesting that the initial binding to gangliosides is critical to specifically accumulate BoNT/A on the membrane of neuronal cells. For example, It has been shown [66] that, in the absence of GT1b, Neuro-2a cells are insensitive to BoNT/A and that knockout mice defective in the production of polysialogangliosides show reduced sensitivity to BoNT/A and BoNT/B [20], [26], [67]. Moreover, a mutant version of HC/A, W1266L & Y1267S that does not bind to GT1b, does not extend paralysis time caused by BoNT/A in murine phrenic nerve-hemidiaphragm preparations demonstrating an impaired ability to bind to neuronal cells [68]. Here we propose that only after BoNT/A is anchored at the neuronal membrane the second crucial step, binding to FGFR3 and/or SV2, can occur. This explains how BoNT/A can specifically enter motor neurons by recognizing FGFR3, a receptor also expressed by non-neuronal cells that lack gangliosides in their membranes. As evidence for a second crucial step, we demonstrate that, in Neuro-2a cells, if either FGFR3 or SV2C binding is blocked, BoNT/A uptake is impaired.
BoNT/A uptake is affected by the cellular levels of FGFR3 expression. We demonstrated that overexpression of FGFR3 increased binding of membrane extracts to rHC/A as well as BoNT/A uptake in three different neuronal cell lines, while down-regulation of FGFR3 reduced uptake of BoNT/A. In contrast, no changes in BoNT/A uptake were observed when increasing or decreasing the expression of SV2C, suggesting that FGFR3, but not SV2C represents a rate-limiting step in BoNT/A uptake under resting conditions. This is consistent with our finding that FGFR3, but not SV2C, co-localized with BoNT/A in un-stimulated PC-12 cells.
By testing eight deletion mutant peptides of the FGFR3b and c extra-cellular domain in the SPR binding assay, we identified the extra-cellular Loop 2,3 of FGFR3 as the minimal optimal binding site for rHC/A. Native ligands for FGFR3 also bind to Loop 2,3 [39], [51], [52] and these data demonstrate that the binding site for rHC/A overlaps the binding site for native ligands of FGFR3. The affinity measurements also demonstrated that FGFR3b bound with slightly higher affinity than FGFR3c, the KD for the subtype b peptides was ∼15 nM, while the KD for the subtype c peptides was ∼25 nM. The FGFR3c subtype is the subtype expressed in the nervous system, while expression of the FGFR3b subtype is restricted to epithelial structures [69], [70]. It is therefore more likely that BoNT/A utilizes the FGFR3c subtype in vivo to gain access into neuronal cells.
In conclusion, this paper presents evidence for FGFR3 as a high affinity receptor for BoNT/A, potentially being part of a larger receptor complex involving sugar- and protein-protein interactions. FGFR3 is present in the target motor neurons. Overexpression of FGFR3 in several neuronal cells increases efficacy and sensitivity to BoNT/A while decreased FGFR3 expression renders the cells less sensitive. BoNT/A binds to FGFR3 at the same extra-cellular region and with the same affinity as native ligands for FGFR3 and functions as an agonist ligand inducing FGFR3 phosphorylation. Moreover, BoNT/A uptake can be blocked by native FGFR3 ligands or by peptide fragments containing the extra-cellular region of FGFR3. Together, these results expand our knowledge of BoNT/A uptake in neuronal cells and present a potential new pathway mediating BoNT/A entry and trafficking into neurons under both resting and depolarizing conditions.
Unless otherwise stated tissue culture reagents were from Invitrogen (Carlsbad, CA)
PC-12- Rat pheochromocytoma cell line (CRL-1721; ATCC) was cultured in collagen IV plates (354528; BD). Growth media: RPMI media with 2 mM GlutaMAX, 5% Fetal Bovine Serum (heat-inactivated), 10% Equine Serum, 10 mM HEPES, 1 mM Sodium Pyruvate, 100 U/ml Penicillin, and 100 µg/ml Streptomycin. Differentiation media: RPMI media with 2 mM GlutaMAX, 1× B27 supplement, 1× N2 supplement, 10 mM HEPES, 1 mM Sodium Pyruvate, 50 ng/ml NGF, 100 U/ml Penicillin, and 100 µg/ml Streptomycin. Neuro-2a- Murine neuroblastoma cell line (CCL-131; ATCC) was cultured in Costar Tissue Culture Flasks (CLS3150; Corning). Growth media: EMEM with 2 mM GlutaMAX, 0.1 mM Non-Essential Amino-Acids, 10 mM HEPES, 1 mM Sodium Pyruvate, 100 U/ml Penicillin, 100 µg/ml Streptomycin, and 10% Fetal Bovine Serum. Differentiation media: EMEM with 2 mM GlutaMAX, 0.1 mM Non-Essential Amino-Acids, 10 mM HEPES, 1× N2 supplement, and 1× B27 supplement. SH-SY5Y- Human neuroblastoma cell line (94030304; ECACC) was cultured in Costar Tissue Culture Flasks (CLS3150; Corning). Growth media: EMEM with 2 mM GlutaMAX/F12, 0.1 mM Non-Essential Amino-Acids, 10 mM HEPES, 1 mM Sodium Pyruvate, 100 U/ml Penicillin, 100 µg/ml Streptomycin, and 10% Fetal Bovine Serum. Differentiation media: EMEM with 2 mM GlutaMAX, 0.1 mM Non-Essential Amino-Acids, 10 mM HEPES, 1× N2 supplement, and 1× B27 supplement. HEK 293- Human Embryonic Kidney 293 cells (CRL-1573; ATCC) were cultured in Costar Tissue Culture Flasks (CLS3150; Corning). Growth media: EMEM with 2 mM GlutaMAX, 0.1 mM Non-Essential Amino-Acids, 10 mM HEPES, 1 mM Sodium Pyruvate, 100 U/ml Penicillin, 100 µg/ml Streptomycin, and 10% Fetal Bovine Serum.
For differentiation, PC-12, Neuro-2a, and SH-SY5Y cells were plated in 96-well plates at 5×104 cells/well in 100 µl differentiation media for three days.
FGFR3b/c peptides and rHC/A were expressed from pET-29 b (+) in E.Coli, Acella Electrocompetent BL21(DE3) (42649; Edge Biosystems). Expression was induced by 1 mM IPTG (V3955; Promega) at either 37°C for 16 hours (FGFR3b/c peptides) or at 16°C for 16 hours (rHC/A). For purification of rHC/A, the supernatant was collected after centrifugation and the protein was purified using the MagneHis Protein Purification System (V8500; Promega). FGFR3b/c peptides were purified from inclusion bodies. After expression, cells were first lysed for 1 hour in five times the cell wet weight in lysis buffer containing 50 mM Tris-HCl pH 8.0, 10 mM EDTA, 100 mM NaCl, 10 mM DTT, 5% (v/v) glycerol, protease inhibitor (P1860; Sigma), 150 mU/ml rLysozyme (71110; EMD Chemicals), and 50 mU/ml benzonase nuclease (70746; EMD Chemicals) and then sonicated for 5 minutes. Pellets were collected by centrifugation and washed three times, first time with wash buffer (50 mM Tris-HCl pH 8.0, 100 mM NaCl, 10 mM DTT, 5% glycerol and 2% Triton X-100) plus 10 mM EDTA, second time with wash buffer only, and third time with wash buffer plus 2 M urea. The inclusion bodies were dissolved in 50 mM Tris-HCl pH 8.0, 500 mM NaCl, 10 mM DTT, 8 M urea, 10 mM imidazole and the peptides were isolated using the Magne-His Protein Purification Resin (V8560; Promega). Wash buffer: 50 mM Tris-HCl pH 8.0, 500 mM NaCl, 10 mM DTT, 8 M urea, and 20 mM imidazole. Elution buffer: 50 mM Tris-HCl pH 8.0, 500 mM NaCl, 10 mM DTT, 8 M urea, and 500 mM imidazole. After elution, the buffer was exchanged to 50 mM Tris-HCl pH 8.0, 1 mM EDTA, 3.8 mM GSH, 1.2 mM GSSH and 1 M arginine using a FastDialyzer fitted with 5 kDa MWCO cellulose acetate membranes (Harvard Apparatus).
The assay was performed according to the protocol from Pierce (21277; Pierce). 150 µg of rHC/A (26 mg/ml stock conc.) was used as “Bait” protein and 500 µl of differentiated PC-12 cell lysate (from 1.5×106 cells) was used as “Prey” protein. As negative controls, samples without either “Bait” or “Prey” protein were run in parallel. The eluted samples were analyzed by SDS-PAGE and Western blot analysis.
10 µg BoNT/A was reacted with 2 mM Sulfo-SBED (33073; Thermo Scientific) (solubilized at 125 mM in DMSO) in 0.1 ml PBS for 2 hours. The reaction was stopped by addition of 0.1 µl 0.4 M Tris. As a control 10 µg of BSA was also reacted with 2 mM Sulfo-SBED. Sulfo-SBED BoNT/A and BSA were added to 1×108 Neuro-2a cells and mixed by rotisserie at 4°C for 4 hours. The reagent was photoactivated for 15 minutes with a UV light source. The cells were washed 4 times with cold TBS and then lysed by incubation for 2 hours in T-X-100 lysis buffer (50 mM Tris, 150 mM NaCl, 1% Triton X-100, 10 mM EDTA, pH 7.2). The biotinylated proteins were precipitated using Monomeric Avidin (Thermo Scientific), washed 4 times in the TX100 lysis buffer and then analyzed by SDS-PAGE and Western Blot Analysis.
Samples were dissolved in 2× SDS-PAGE loading buffer (LC2676; Invitrogen), heated to 95°C for 10 min, resolved in 12% 26-well Criterion gels (345-019; Bio-Rad) or 12% Bis-Tris Novex NuPage gels (NP0341BOX; Invitrogen), transferred to 0.45 µm nitrocellulose membranes (62-0233; Bio-Rad), blocked for 1 hour in TBS buffer (170-6435; BioRad) plus 0.1% Tween 20 (161-0781; BioRad) (TBS-T) and 2% blocking agent (RPN418V; GE Healthcare), and incubated overnight with primary antibody, either; anti-SNAP25197 (Allergan) rabbit polyclonal antibody diluted to 1 µg/ml, anti-HC/A (Allergan) rabbit polyclonal antibody diluted to 1 µg/ml, anti-FGFR3 (1∶500, sc-123; Santa Cruz Biotechnology), anti-FGFR3 (1∶1000, Ab133644; Abnova), anti-SV2A, (1∶200, sc-28955; Santa Cruz Biotechnology), anti-SV2B (1∶200, sc-28956; Santa Cruz Biotechnology), anti-SV2C (1∶500, sc-28957; Santa Cruz Biotechnology) or anti-Syntaxin (1∶200, sc-12736; Santa Cruz Biotechnology) in TBS-T plus 2% blocking agent. Secondary antibody was anti-rabbit IgG H+L HRP conjugate (81-6120; Invitrogen), anti-rabbit IgG veriBlot for IP secondary antibody (HRP) (ab131366, Abcam) (used for IP only), and anti-mouse IgG H+L HRP conjugate (62-6520; Invitrogen) diluted 1∶5000 in TBS-T plus 2% blocking agent. Membranes were developed using ECL Plus Western Blotting System Detection Reagents (RPN2132; GE Healthcare). The Chemiluminescence was captured using a Typhoon 9140 (GE Healthcare) set to the following parameters: 455 nm excitation laser and detector set to all wavelengths below 520 nm emissions. The intensity of the gel bands were calculated using Image Quant software TL V2005 (GE Healthcare). The data was analyzed using PLA and SigmaPlot v 10.0 (Systat Software Inc.). Intensity values were plotted against concentration of BoNT/A in log scale and fitted to a 4-parameter logistics function (Y = Y0+a/[1+(X/X0)b]) without constraints. Based on the fitted curves the EC50 values, corresponding to “X0”, were determined.
Cells were transfected with pcDNA3.1 (+) (V790-20; Invitrogen), FGFR3c (EX-Y0098-M50; Genecopoeia), SV2C (EX-S2660-M050; Genecopoeia), RNAi Hi GC (12935-400; Invitrogen), FGFR3 siRNA-88 (FGFR3RSS331488; Invitrogen), FGFR3 siRNA-89 (FGFR3RSS331489; Invitrogen), SV2C-1 siRNA (AM16708; Ambion). Membrane extractions were performed with a Native Membrane Protein Extraction kit (444810; Calbiochem). Total protein concentration was measured using Bradford Reagent (500-0205; Bio-Rad).
Differentiated Neuro-2a cells transfected with FGFR3 were treated with 0.5 nM or 50 nM FGF2 (233-FB; R&D Systems) or rHC/A (26 mg/ml stock concentration) for 10 minutes. Membrane extracts were prepared and 100 µg of total protein was incubated with 40 µl of a 50% slurry of anti-phosphotyrosine conjugated beads (16-101; Millipore) for 24 hours at 4°C. Samples were washed 4 times with MEB buffer (50 mM Tris pH 7.5, 150 mM NaCl) containing phosphatase inhibitor cocktail 1 and 2 (P2850 and P5726; Sigma) and complete protease inhibitor cocktail (11 873 580 001; Roche) and analyzed by SDS-PAGE and Western blot analysis, using antibody against FGFR3.
Differentiated cells (see Cell Lines and Growth Conditions) were treated with 0–30 nM BoNT/A (0.41 mg/ml stock concentration) in differentiation media for 6 or 24 hours followed by overnight or two day incubation in toxin-free media. Cell lysates were analyzed by SDS-PAGE and Western blot analysis, using antibody against cleaved SNAP25, anti-SNAP25197 (Allergan).
For competition, before treatment with 1 nM BoNT/A (150 kDa, Metabiologics), Neuro-2a cells were pre-treated for 30 min with increasing concentrations of FGF1, FGF2, FGF9, FGF10 (negative control) (132-FA; 233-FB; 273-F9, and 345-FG; R&D Systems), or rHC/A (Figure 1A, positive control). For inhibition, before treatment onto cells, 1 nM BoNT/A (150 kDa, Metabiologics) was incubated for 20 min with increasing concentrations of either; FGFR3b/c deletion mutant peptides (Figure 6A), SV2C529–579 (JPT Peptide Technologies; aa529–279, H-NTYFKNCTFIDTVFDNTDFEPYKFIDSEFKNCSFFHNKTGCQITFDDDYSA-NH2, Figure 2B), monoclonal anti-HC/A 6B1 (Provided by Dr. L. Smith, USAMRIID; positive control), or Synaptotagmin II1–20 (JPT, aa1–20, H-MRNIFKRNQEPIVAPATTTA-NH2; negative control). The competitor/inhibitor was added at concentrations of 2, 5, 20, 50, and 200 molar excess of BoNT/A. Cells were incubated with BoNT/A plus competitor/inhibitor for 2 hours. The toxin containing media was then removed and replaced with fresh media followed by overnight incubation. Cells lysates were analyzed by SDS-PAGE and Western blot analysis, using antibody against cleaved SNAP25, Anti-SNAP25197. Blots were quantified and the amount of SNAP25197 produced at each concentration, as a measure of BoNT/A uptake, was used to calculate the percent competition/inhibition. The amount of BoNT/A uptake for each competitor/inhibitor was compared to the amount of BoNT/A uptake after pre-treatment/pre-incubation with the negative control. Each experiment was conducted at least three independent times and each dose was tested in triplicates in each individual experiment, the percent average for each of the three or more independent experiments were used to generate inhibition curves. The curves were fitted to a non-linear exponential model; Y = 100×e −b*log(concentration), where “b” was defined as either the competition constant (CC) or the inhibition constant (IC).
Differentiated Neuro-2a cells were washed with PBS and then lysed by incubation at 4°C for 30 minutes in lysis buffer containing 20 mM Tris, 150 mM NaCl, 1% Triton X-100, 1 mM EDTA, and 1 mM EGTA pH 7.2 plus complete protease inhibitor cocktail (11 873 580 001; Roche). The supernatant was collected by centrifugation and the total protein concentration was measured using Bradford Reagent (500-0205; Bio-Rad). The Co-IP reaction was performed by mixing 1 mg of cell lysate with 10 µg antibody in a total volume of 1 ml. The reaction was incubated at 4°C overnight. As negative controls, a sample without antibody (lysate only) and a sample without lysate (antibody only) were prepared in parallel. Then 100 µl Protein A/G Magnetic Beads (88802; Thermo Scientific) was added and the reaction was incubated at 4°C for 1 hour. Three times the beads were sedimented using a dynamag-2 magnet (Invitrogen; 12321D) and washed with PBS. Finally, the beads were re-suspended in 2× SDS-PAGE running buffer and Western Blot was performed. Cell lysate was run in parallel.
Purified rHC/A (0.4 mg) was dialyzed against 4 L of 50 mM HEPES pH 7.0–7.2 150 mM NaCl in a 0.5 ml dialysis unit (Harvard Apparatus) for 16 hours at 4°C using a 25 kDa cut-off cellulose acetate membrane. FGF2 (0.2 mg) (233-FB/CF; R&D Systems) was re-suspended in 0.5 ml of 50 mM HEPES pH 7.2, 150 mM NaCl. rHC/A was labeled with either Alexa Fluor 488 C5-maleimide (A10254; Invitrogen) (cell imaging and photobleaching) or Alexa Fluor 633 C5-maleimide (A20342; Invitrogen) (rHC/A uptake) and FGF2 was labeled with Alexa Fluor 633 C5-maleimide (10∶1 molar ratio free label to protein) overnight at 4°C in the dark. To remove free label, the proteins were either dialyzed again, using the conditions listed above, or ran on a PD-10 desalting column (17-0851-01; GE Healthcare). The column was equilibrated with 50 mM HEPES pH 7.2 150 mM NaCl. The concentrations of rHC/A and FGF2 were determined by measuring the UV absorbance at 280 nm using a Beckman Coulter DU 800.
Cells were plated at 20,000 cells per well in a Cell Carrier microplate (Perkin Elmer) in culture media and allowed to attach overnight. Cells were treated with 1 µg/ml Cell Tracker Green CMFDA (C2925; Invitrogen) and Hoechst 33342 (H10295; Invitrogen) for 30 minutes prior to adding fluorescently labeled rHC/A or FGF2. After removing the staining media, 0.1 ml of 0–25 nM of Alexa Fluor 633-rHC/A or -FGF2 was added. Fluorescence was measured using the Operetta High Content Imaging System (Perkin Elmer), set to the following parameters: 20× WD objective, 9 fields, non-confocal, 15% excitation, Blue (Ex 380–410/Em 430–460), Green (Ex 460–490 nm/Em 500–550 nm) and Far Red (Ex 630–645 nm/Em 660–900 nm). Uptake, measured as increasing amounts of Far Red signal in the cells, was monitored for 15 hours, with 30 minutes time points. The results were analyzed using Harmony 3.1 software.
PC-12 cells transfected with Halo tagged FGFR3 or SV2C were plated on Collagen IV coated glass bottom dishes (P35GCOL-0-10-C; MatTek) and differentiated for 3 days. The cells were incubated with 5 µM Halotag TMR ligand (G8251; Promega) for 15 minutes and washed 3 times for 10 minutes with fresh media. Cells were then treated with 1 µM Alexa Fluor 488 labeled rHC/A. After 2 hours incubation, the cells were fixed using 5% paraformaldehyde. Cells were imaged using a LSM710 confocal microscope and analyzed using ZEN 2009 software (Carl Zeiss INT, Germany). The Alexa Fluor 488 label and TMR star labels were imaged using the following respective settings: excitation 488 nm/emission 500–510 nm, excitation 561 nm/emission 595–620 nm. The TMR was photobleached by exciting the fluor with the 561 nm laser 100 times for 0.1 s with 5% laser power. The amount of fluorescence intensity of the donor fluor (AF-488) was measured before and after photobleaching of the acceptor (TMR).
Sprague-Dawley rats (200–250 g; Charles River) were injected with 10 units of BOTOX (Allergan) into the tibialis anterior (TA) muscle of the right hind limb. Animals receiving injections of 0.9% saline into their TA muscle served as controls. Rats were sacrificed 3 days following injections and their TA muscles were harvested. Muscles were embedded in OCT compound, frozen in liquid nitrogen and stored at −80°C. Prior to staining, muscles were cross-sectioned (10 µm) using a cryostat (Leica), mounted onto microscope slides and stored at −20°C until use. Frozen, slide-mounted muscle sections were thawed to room temperature and immediately fixed with 2% paraformaldehyde for 10 min. Sections were blocked with 5% normal serum in PBS, pH 7.4 for 60 minutes and then incubated with primary antibodies for 3 hours at room temperature: rabbit anti-SV2C (1∶400, sc-28957; Santa Cruz Biotechnology), rabbit anti-FGFR3 (1∶200, sc-123; Santa Cruz Biotechnology), mouse anti-SNAP25 (1∶200, SMI-81, Covance), and mouse anti-SNAP25197 (1∶200, Allergan). Muscle nicotinic acetylcholine receptors (nAChR) were labeled with α-bungarotoxin (α-Bgt) Alexa-Fluor 647 conjugate (1∶500, Invitrogen). Sections were then washed and incubated with secondary antibodies for 30 minutes at room temperature. Following a final wash, slides were coverslipped and analyzed. Images were acquired using a Zeiss LSM-710 confocal microscope (Carl Zeiss INT).
Experiments were performed on a BIAcore 3000 instrument (GE Healthcare). Ligands, rHC/A, anti-HC/A 6B1 (Provided by Dr. L. Smith, USAMRIID), FGF2, or FGF9 (233-FB; and 273-F9; R&D Systems), were immobilized on a CM5 chip (BR-1003-99, GE Healthcare) using an amine coupling kit (BR-1000-50, GE Healthcare). Analytes, either SV2C529–579 (JPT Peptide Technologies, dissolved in 100% DMSO), FGFR3b/c deletion mutant peptides (Figure 6A), rHC/A, or membrane extracts were injected over the ligand surfaces at concentrations ranging from 0–5000 nM, or for the membrane extractions at 5 µg/ml. The flow rate was set to 20 or 30 µl/min. Running buffer: HBS-EP buffer (BR-1006-91, GE Healthcare). The surfaces were re-generated by two 1-min injections at 30 µl/min of 10 mM Glycine, pH 1.5 (rHC/A) or 1-min injections at 30 µl/min of either; 10 mM Glycine, pH 1.5 and 0.125% SDS (FGFR3), 10 mM Glycine, pH 1.5 and 20 mM CHAPS (Membrane extracts), or 10 mM NaOH (SV2C). The sensorgram curves were evaluated using the BIAevaluation 3.0 software. The curves were fitted to a 1∶1 Langmuir binding model (A + B ↔ AB, where A is the analyte and B is the ligand immobilized on the sensor surface). Based on the fitted curves the association constant, ka, the dissociation constant, kd, and the equilibrium constant, KD (KD = kd/ka) were determined. The FGFR3b/c peptide curves were also visually compared using the “normalization” wizard in the BIAevaluation 4.1 software.
To assess the significance of the differences from the BoNT/A cell-based competition/inhibition assays and the SPR Binding Analysis assays t-tests were performed using online Graphpad software; www.graphpad.com/quickcalcs/ttest1.cfm?Format=SEM (GraphPad Software Inc).
FGFR3 (ENSG00000068078), SV2A (ENSG00000159164), SV2B (ENSG00000185518), SV2C (ENSG00000122012), FGF1 (ENSG00000113578), FGF2 (ENSG00000138685), FGF9 (ENSG00000102678), and FGF10 (ENSG00000070193).
|
10.1371/journal.pntd.0003624 | Costs Of Using “Tiny Targets” to Control Glossina fuscipes fuscipes, a Vector of Gambiense Sleeping Sickness in Arua District of Uganda | To evaluate the relative effectiveness of tsetse control methods, their costs need to be analysed alongside their impact on tsetse populations. Very little has been published on the costs of methods specifically targeting human African trypanosomiasis
In northern Uganda, a 250 km2 field trial was undertaken using small (0.5 X 0.25 m) insecticide-treated targets (“tiny targets”). Detailed cost recording accompanied every phase of the work. Costs were calculated for this operation as if managed by the Ugandan vector control services: removing purely research components of the work and applying local salaries. This calculation assumed that all resources are fully used, with no spare capacity. The full cost of the operation was assessed at USD 85.4 per km2, of which USD 55.7 or 65.2% were field costs, made up of three component activities (target deployment: 34.5%, trap monitoring: 10.6% and target maintenance: 20.1%). The remaining USD 29.7 or 34.8% of the costs were for preliminary studies and administration (tsetse surveys: 6.0%, sensitisation of local populations: 18.6% and office support: 10.2%). Targets accounted for only 12.9% of the total cost, other important cost components were labour (24.1%) and transport (34.6%).
Comparison with the updated cost of historical HAT vector control projects and recent estimates indicates that this work represents a major reduction in cost levels. This is attributed not just to the low unit cost of tiny targets but also to the organisation of delivery, using local labour with bicycles or motorcycles. Sensitivity analyses were undertaken, investigating key prices and assumptions. It is believed that these costs are generalizable to other HAT foci, although in more remote areas, with denser vegetation and fewer people, costs would increase, as would be the case for other tsetse control techniques.
| Sleeping sickness remains a serious threat in Sub-Saharan Africa. The disease is normally controlled by medical screening of the human population and treatment of individuals found to be infected. The disease is transmitted by tsetse flies but vector control is rarely used for control. A major reason given is that is too expensive in resource poor settings. We have developed a novel technology based on insecticide treated screens (= tiny targets) to control flies more cost-effectively. A 250 km2 field trial of tiny targets has been performed in Northern Uganda and we made use of this to undertake a full costing analysis of tiny target technology. The cost of the operation was costed at USD 85.4 per km2. This represents a major reduction in the cost of tsetse control. The reductions are largely due to the low costs of tiny targets and to the ease with which they can be deployed.
| Tsetse control technologies and their mode of delivery are evolving all the time. A major purpose of this evolution is to develop approaches that can reduce the incidence of human and animal African trypanosomiasis (HAT and AAT) more cheaply and/or more effectively. Measuring cost-effectiveness accurately, and in such a way that different operations and approaches are fully costed and can be validly compared, is essential to underpin decision-making in this field [1]. This paper analyses the costs of an actual field operation using the new technology of tiny targets undertaken in Arua District, Uganda in 2012/2013 whose ultimate aim was to reduce transmission of HAT by controlling Glossina fuscipes fuscipes [2].
In contrast to analyses of tsetse control operations primarily undertaken to control AAT, costs of such operations undertaken in HAT foci have only been intermittently reported on in the entomological literature [3]. With the development of lower cost devices a major concern, these reports have focussed on their unit cost and related this cost to the km2 and the human population ‘protected’. The term ‘protected’ was introduced to indicate the area and the people within that area who benefited from tsetse control, as against the much more restricted area of tsetse habitat where traps or targets were actually deployed. Thus, excluding manpower, the newly developed Vavoua trap was reported as costing about half as much as the standard biconical and pyramidal traps [4]. Four projects using traps and screens in HAT foci published cost-effectiveness estimates for Côte d’Ivoire [5], Congo [6], Equatorial Guinea [7] and Uganda [8]. Coincidently these all relate to the 5-year period 1986–1990. In Uganda, the project area’s population was around 320,000; the other three all worked in HAT foci containing about 25,000 people. All cite trap costs, which can be compared to levels today by converting from local currencies to United States dollars (USD) at the historical rates applicable at the time, and then updated to current (2014) prices by applying the USD inflation rate (http://inflationdata.com/inflation/Inflation_Rate/HistoricalInflation.aspx historical data) which for this period yields a factor ranging from 2.16 to 1.88, thus roughly doubling all prices. Thus, at 2014 prices in Côte d’Ivoire screens cost USD 6.6 and Vavoua traps USD 13.6; in Equatorial Guinea [7] pyramidal traps costing an estimated USD 9 [9] were used. Meanwhile, in the Congo [6], pyramidal traps costs USD19 and villagers were supplied with a repair kit for the traps costing USD 6, giving an average annual cost per trap of USD10, all at 2014 prices. In Uganda [8] pyramidal traps costing USD 6 were being used on a large scale, and the newly developed mono-screen trap [10] cost USD 8.8 at 2014 prices. In 2012/13 pyramidal traps were bought for use in Uganda at a cost of USD 10—indicating that their relative price has remained surprisingly stable over time.
In order to evaluate the relative cost-effectiveness of traps and targets/screens, these monetary costs needed to be assessed alongside measures of effectiveness against tsetse populations. For traps, catches can be compared (e.g. [10]). However, in order to compare traps and screens a more sophisticated metric is required because targets do not retain the flies killed. The tsetse control operation analysed here follows on from a decade’s research into increasing the ‘cost-effectiveness’ of the targets themselves, measured in terms of tsetse caught or killed per m2 of cloth for G. f. fuscipes [2,11] and G. palpalis palpalis [12]. This standard metric has made it possible to compare the effectiveness of the classic targets or biconical / pyramidal traps, in use since the 1980s as described above, with much smaller devices. The amount of fabric required gives a clear and measurable indicator of trap/target cost which can be compared over time and across countries and currencies. For G. p. palpalis, the killing efficiency of a medium-sized horizontal target design 0.5 m2 was shown to be 6 times greater than of the classic 1 m2 target [11]. The adoption of the ‘tiny’ 0.125m2 (0.5 X 0.25 m) target for use in this trial follows directly from these studies [11] showing the killing efficiency of G. f. fuscipes per m2 to be between 5.5 and 15 higher than for 1 m targets and up to 8.6–37.5 greater than for biconical traps. Similar results were recently obtained by [13] for G. p. palpalis, showing 0.25 m2 targets to be promising as cost-effective devices, but using relative catches as a metric. The cost-effectiveness of different traps for G. f. fuscipes has also been studied using catch per linear m of fabric as a metric [14].
The four historic projects cited above went on to calculate trap costs per person protected, which at 2014, prices came out to USD 11, USD 11, USD 0.5 and USD 1 for Côte d’Ivoire, Equatorial Guinea, Congo and Uganda respectively. For Côte d’Ivoire, adding the deployment costs for fuel and vehicle maintenance plus trap replacement and reimpregnation with insecticide increased the cost to USD 13 per person protected. Although labour and staff costs were not costed [5] provided a detailed inventory of all inputs, including people’s time alongside full instructions for estimating the costs of operations. In Uganda, adding cost for staff, labour and transport increased the cost per person protected to USD 2. Costs were also given per km2. These costs reflect very different population densities in the HAT foci from 17 per km2 in Côte d’Ivoire to 100 in Uganda. Devices were also placed at very different densities with 25 per km2 protected in Côte d’Ivoire, 10 to 15 per km2 in Uganda. Costs per km2 in Uganda, at 2014 prices, worked out at USD 85, rising to USD 179 if staff, labour and transport were included; in Côte d’Ivoire the cost per km2 protected was higher at USD 217. These costs refer to the first year of deployment. For all the projects, it was thought that costs would fall in the second year of operation, with trap life sometimes extending beyond one year and deployment sites having been selected and prepared.
However, a full analysis of the cost-effectiveness also has to include all delivery costs. It has long been known that the tsetse control techniques described as “expensive” and” high-tech” and usually deployed on a larger scale, such as aerial spraying and the sterile male technique, used on their own, require less expensive ground level support and thus have apparently lower delivery costs than targets and traps, since flying time is usually included in the ‘core’ cost of the technology. As far back as the late 1970s it could be shown that the differential between the apparently high cost of helicopter spraying and ground spraying was greatly eroded when the full delivery costs for ground spraying were included [15] and the same was true for targets and aerial spraying [16]. More recent comparisons [1] also indicate that while total costs of bait technologies (whether stationary: traps/targets or live: insecticide-treated cattle) can be substantially lower, in relative terms, their delivery costs are substantially higher in relation to their core costs (insecticides and traps/targets) than is the case for than aerial spraying or the sterile male technique (core costs of insecticide, sterile males and flying time). The need to reduce delivery costs for the bait technologies was part of the reason why many projects have tried to involve local communities, not just by informing them about the objectives and benefits of using traps and targets, but also in terms of contributing labour and ensuring traps /targets remained in place and in working order [5]. However, community involvement has had mixed success [17], with the needs of communities often being treated as secondary to the entomological objectives. Although not part of a project budget, inputs by community members impose an economic cost on that community, so that an economic analysis should value these inputs.
Lastly, whereas the costs of traps and targets or insecticide can be reduced, delivery costs do not necessarily decline proportionately. For this reason, simply multiplying the trap or insecticide cost by a constant to estimate the delivery cost is unlikely to be reliable. To date, apart from the meticulous detailed information recorded by and the preliminary estimates made for the use of insecticide-treated cattle in southeastern Uganda and reported in [1] there is no published accurate assessment of the delivery costs of such an operation in a HAT focus. Accordingly, alongside the entomological monitoring of the control operation in Arua reported by [18] an important component of the project’s work was the detailed recording and pricing of all inputs.
The study focussed on a control operation using tiny targets which covered 250 km2. Work began in June 2012 with a sensitisation operation, and continued to the end of June 2013, thus covering a period of 13 months. It was split into six sub-activities, spread over that period as given in Table 1 and illustrated in Fig. 1. To monitor costs for each activity, a data sheet was kept, recording the number of days spent in the field (Table 1), staff deployment, labour hired, vehicles used and kilometres travelled, use of other capital items such as global positioning sets (GPS), laptop computers, specialist items (traps, targets and extension materials) and all other running costs (e.g., fuel and oil, vehicle maintenance, hired transport, stationery, GPS batteries, protective clothing for staff, assorted minor consumables). In this way all field and non-field costs were recorded and both variable and fixed costs were fully accounted for. Office overheads were also monitored. The information thus collected provides a full set of data for an actual field operation.
The operational area on which these costs are based consisted of the area surrounding five blocks, each of 7 x 7 km, which were the subject of a control operation initiated in 2011 (Fig. 2). At the beginning of December 2012, the control area was enlarged from ~250 km2 (5 x ~50 km2) to 500km2. The work done in the ‘new’ areas was carefully logged and separated from that done in the five original ‘old’ blocks (Fig. 2). These areas are contiguous and there was little additional travel between locations. Accordingly, all costs were divided by 250 to produce a cost per km2 controlled.
The costing methodology adopted was the ‘full costing’ approach described in [1]. By clearly itemising cost components, the calculations undertaken here are presented so as to enable effective comparison with those presented for other operations.
The overall objective was to produce a replicable costing at current prices for a field tsetse control operation run by local staff in the Ugandan context. Thus costs were adjusted so as to remove the purely research components. The organisation and supervision of the work was undertaken by an academic research team composed of an anthropologist and three entomologists. Supervisory staff inputs were costed at the salaries and travel allowances paid to a Ugandan senior entomologist, for the time spent in the field and on administrative duties. This included a proportionate allowance for weekends and holidays. Each district in Uganda has an entomologist responsible for vector control. Two categories of preparatory work were costed: sensitising local populations and a preliminary tsetse survey. Two costings for the survey are presented, one as actually incurred, the other for a streamlined operation with no research component. The sensitisation programme was treated as a ‘capital’ investment—which would be valid for at least three years before follow up activities were needed.
Within each component activity, the ‘full costs’ for this operation were calculated as follows. Depreciation was estimated for all capital items which outlasted the 13-month project duration. The relevant items and the assumptions used are listed in Table 2. For non-durable items (fuel and oil for vehicles, labour, travel allowances and/or per diems for staff, protective clothing for staff, backpacks, slashers, pangas, stationery, GPS batteries, extension materials and refreshments for villagers during village meetings) the actual recorded expenditure was used. Although the life of traps can exceed one year, they were classified along with targets as ‘specialised equipment’ and their working life was conservatively estimated at 150 days deployment.
However, for a project undertaken within an existing government structure, some capital and recurrent costs would be spread over several activities, of which tsetse control in a specific area would only be one. In the case of this project, as explained above, in addition to the parallel control work undertaken in the ‘old’ project’s five 50 km2 squares, a substantial proportion of the time was allocated to research. Thus only 25% of the office overheads were allocated to the project being costed. Also, for this reason only a share of basic salaries, reflecting the time spent on the project, was included for entomologists. Similarly, for some items (e.g., motorbikes, GPS sets, laptops, traps) an appropriate share of their annual use (based on kilometres travelled or days used, as recorded in the data sheets) was attributed to the project. However, for the 4x4 pickup truck, the cost was based on the total recorded distance travelled which was virtually all for this 250 km2 control project.
Prices were all converted to US dollars (USD) mainly from 2012/13 Uganda Shillings (UGX) (and occasionally other currencies). The average rate for the project period was 2615 UGX = 1 USD, ranging from 2416 to 2700 (http://www.oanda.com/currency/historical-rates/). The rate of 2615 UGX = 1 USD was applied throughout. Other currencies used were the British pound (GBP) and the Euro (EUR), whose conversion rates to 1 USD for the period were 0.6382 GBP and 0.7755 EUR. Total costs are rounded to the nearest USD without removing the effects of rounding, thus in some cases the totals will appear not to add up, but all individual figures are accurately rounded. Costs per km2 are rounded to the nearest USD 0.1. All costs are at market prices applicable at the time and place where they were incurred and include the cost of shipping to Uganda as relevant.
Before undertaking tsetse control in the new enlarged area, a preliminary entomological survey was undertaken. The objective of this was to identify suitable sites for locating traps. Traps were deployed and visited daily for 3 days, 1 to deploy and 2 to monitor. The total number of field days was 20, deploying 8 traps (Table 3). The work was undertaken by a senior entomologist and driver, using the 4x4 pickup truck. Total travel was just over 2000 km, accounting for 40% of the vehicle’s mileage during the study period. Some use was also made of the office motorbike. Pangas and slashers were used to cut down the vegetation around the traps sites. The costs are summarised in Table 3. Total costs came to USD 5901, which works out at USD 23.60 per km2. The costs were dominated by cost of depreciation (44%) and maintenance (25%) of the pickup truck, reflecting its low annual travel of some 5000 km and relatively high maintenance costs, which were only to some extent offset by its low depreciation, both reflecting the fact that it was 15 years old. Staff salaries and allowances accounted for a further 23% of costs, and fuel 6%.
During the course of the project, the reliance on the pick-up truck was gradually reduced, and teams of trap and target attendants were trained. They accessed the project area either by motorbike or bicycle, sometimes using public transport, transporting traps and targets in backpacks. To investigate the impact of this technological and logistical improvement, the preliminary monitoring costs were recalculated (Table 4) based on the timings achieved in activity D (monitoring the actual target deployment) and allowing for more intensive monitoring (12 traps and 25 field days, monitoring a total of 100 trap sites, or 4 per 10 x 10 km square). This intensity is equivalent to the monitoring undertaken in study zone in 2010 when work first began in Arua area. The work would be done by a team of two trap attendants using a motorbike. This approach would allow for considerable cost savings, reducing the total cost to USD 1281, or USD 5.12 per km2.
Before undertaking the sensitisation campaign preparatory research activities were [19] undertaken. This highlighted people’s wariness, and in some cases fear, of the targets and traps, and underlined the need for an effective public awareness initiative. The campaign proceeded in several steps. Three different sensitization materials were developed in English and translated into Lugbara: letters for communities, information leaflets and flip-charts used for training and for house-to-house sensitisation.
Preliminary meetings with sub-county authorities were carried out and they were briefed about the project activities and sensitization campaigns planned in their area. All the villages in the new control area were identified and mapped, using geographical positioning systems (GPS). This work was initiated by the research team, led by an anthropologist, and then taken over by the target attendants. In total 130 villages were involved in the campaign. The costs (Table 5) are based on what was actually experienced, although the speed of work varied, with the initial research team of 4 managing to walk 12 km a day and identify 18 villages a day (4.5 per person per day). When this was taken over by the target attendants, whose task included this work and then later placing and renewing targets, the rate fell to 1.2 villages per person per day. Eight training meetings were organised for the village health teams (VHTs) at the sub-county headquarters at which refreshments were provided and travel costs paid. VHTs, consisting of two volunteers chosen by each village, are part of the health delivery system in Uganda. VHTs were paid two daily lunch allowances, which corresponded to the usual national VHT rates. Out of the 260 VHT’s, 257 came for training and then returned to their villages to carry out two days of house-to-house sensitization using picture-based flip charts and samples of targets. They were requested to place information sheets on the usual village notice boards and trees. VHTs were moving on foot or using their own bicycles, so no extra transport cost was included in this exercise. On the third day the research team recollected sensitisation materials and the sensitization forms on which the number of people who received the message in house-to-house campaign was recorded. On this occasion VHT members received their lunch allowances and transport fees. This was followed by training meeting for VHTs from another sub-county, where flip charts and targets were reused. In total 8,713 households were visited and 56,983 people received the message.
The campaign took place over six months and its costs are summarised in Table 5. The total cost came to USD 11,984. Transport accounted for 43% of costs, labour and staff a further 47%, with only 5% required for the extension materials. It is considered that at least three years would elapse before further sensitisation activities would be needed. Accordingly a third of the costs were allocated to the one-year tsetse control operation, coming to USD 15.86 per km2.
In addition, during the course of the project, trap and target attendants undertook sensitization regularly on an informal basis as part of their daily routines: talking to people washing by the rivers, or working in fields. This had no cost implication, as the full cost of their work is included in the activities described below.
The target deployment activity began in early December 2012, and has been described in detail [18]. Targets were deployed at about 6 per km2 when averaged over the whole area, but at a higher density in the riverine habitat as shown in Fig. 2). Because they are so close to each other, the targets in the ‘new’ project area costed here show up as a green dotted line, whereas those in the ‘old’ project area form a red dotted [18]. A total of 1,536 were deployed, and 1,551 provided for in the cost estimate, thus allowing for a slight excess. Targets were manufactured by Vestergaard-Frandsen (Lausanne Switzerland) and shipped from Vietnam. The cost per target was USD 1. Effective target life was assumed to be more than six months and less than a year (see activity E) so targets were treated as recurrent rather than capital cost items. The cost of shipping and insurance varied greatly. In storage tiny targets have a long shelf life—estimated to be about two years. A rapid air consignment cost USD 0.40 per target, lower cost air transport was quoted at USD 0.17 for 10,000 targets and USD 0.126 for 50,000. Larger quantities could be sent by sea, at an estimated cost of USD 0.045 for 100,000 and USD 0.012 for 500,000. If larger scale target deployment were coordinated by the Ugandan government, the sea route would be preferable. In these cost calculations, a cost of USD 0.10 was used, on the assumption that targets would mostly be transported by sea. In order to deploy the cloth targets, wooden supports had to be prepared and glued, a task under taken by the ‘target fixer’.
The deployment was done by teams with bicycles. Targets and supports were transported in backpacks. For the more distant sites, the target teams used public transport (‘boda-boda’ taxi motorbikes and small pick-up trucks) to reach the deployment zone and were paid transport allowances to fund this. The total cost for the activity came to USD 7,370, or USD 29.5 per km2 (Table 6). The single largest cost item was the targets, accounting for 27.5% of costs, followed by labour at 25%.
Once a substantial proportion of the targets were deployed (end of December, 2012) the monitoring activity began (Fig. 1). This continued until the end of the evaluation period. Traps were deployed in the new control zone at 12 sites every twice a month for 3 days, and monitored, as illustrated in Fig. 2. Some of the new trap sites were outside the project area, in order to monitor the impact of the control operation on the boundaries of the treated area. This work was done by the trap attendants, who were supplied with motorbikes. The payment modalities gradually evolved—from allowances to reclaiming actual costs. In the cost calculations (Table 7) all the costs were based on the fuel required for the actual mileage and the associated maintenance costs plus depreciation. The total cost for this activity came to USD 2,250 which worked out at USD 9.00 per km2. Vehicle running and depreciation accounted for 46% of this cost, and labour a further 29%.
The target maintenance operation began at the end of March 2013. All target sites were visited and targets were repaired or replaced as required, the vegetation around them cleared, etc. By the end of the evaluation period 950 targets had been replaced. The modalities were the same as for the initial deployment, with trap attendants using bicycles or hiring local transport to access the intervention areas. However, the time required was much less, since the sites for the targets had already been determined and only some vegetation clearance was required. The costs are summarised in Table 8. The total costs came to USD 4,290 or USD 17.16 per km2. The main cost item was vehicle running (34%) followed by targets (29%) and labour (24%).
Lastly, it is important in field-based projects such as this not to neglect the costs of administering and organising the work. The costs of each field activity include non-field days (itemised in Table 1), mainly for supervisory and research staff. Over and above this it was necessary to run an office, allowing internet access, other communications and processing of data for research purposes as well as routine administration and organisation. The cost components are itemised in Table 9. The total cost of running the office for the evaluation period came to USD 8,724. The office served the two 250 km2 control operations. About half the time was taken up with research. Accordingly, 25% of the total cost was attributed to the ‘new area’ control operation costed in this paper. This worked out at USD 8.72 per km2.
Combining the costs from the six component activities produced the results given in Table 10. The overall total for the control work in the ‘new area’ was USD 21,337, coming out at USD 85.4 per km2 or USD 13.8 per target deployed. Of this over half (55%) was for deploying and maintaining targets. The cost of the targets themselves came to 13% of total project costs. By expenditure category, the single largest cost component was transport (35%) followed by labour (24%).
In order to test the robustness of the cost estimates, a range of sensitivity analyses was undertaken (Table 11). These looked first at the impact on overall costs of cost increases of a third in crucial components: targets, traps, labour, senior staff and fuel and public transport costs. The most sensitive items were labour (8.0% increase) and fuel and public transport costs (6.1%), reflecting their relative share in total costs. Secondly, the implications of varying some of the key assumptions made in the cost estimation were examined. This showed that while varying the share of office overheads allocated to the tsetse control operation had only a limited impact, if the sensitisation programme had to be repeated every two rather than every three years, costs would increase by 9.3%. Using the preliminary survey as an example showed that if just this single activity were undertaken using senior staff and a vehicle rather than local staff with bicycles and motorbikes, the overall cost would increase by over 20%.
It is important to set these results in context. As stated in [18], a fully inclusive cost of USD 85.4 per km2 involves a marked reduction on earlier estimates. If office overheads, sensitisation and preliminary studies (which account for 35% of costs), are excluded the figure comes out as a ‘field cost’ of USD 55.7 per km2, which is the figure that should be compared to the costs published in much of the literature. Referring back to Fig. 2 and the methods section, it should be noted that cost are given for the whole area ‘protected’ rather than per mile of riverine habitat treated. In historic terms this cost is well below the USD 179 (at 2014 prices) noted by [8] also for Uganda with 10–15 traps per km2. In terms of contemporary estimates it is also far lower than the USD 556 estimated by [1] for 10 traps per km2. The human population density in the control zone was estimated at 500 per km2, based on [20] together with gridded data obtained from http://www.afripop.org/. Thus the cost per person ‘protected’ is thus very low, at USD 0.17.
The results of the sensitivity analysis (Table 11) help to explain why the operation was so-cost effective, and to underpin a discussion of the factors which might limit this cost-effectiveness. Looking first at the cost of targets, one of the reasons this type of operation is much less expensive than those undertaken in the late 1980s, as reported on above, is that targets now come made of fabric which has been pre-impregnated with insecticide, so repeat impregnations are not required. Also, when comparing to operations targeting other Glossina species, especially those of the morsitans group, it should be noted odours are not required as bait here. However, while the target cost was fixed at USD 1 by the supplier, the costs of shipping from Vietnam, where the targets were made, fluctuated a lot. A simple 33% increase in the cost of targets would take the overall cost of the operation up by 4.3%. As explained above, based on the quotes received and in the hope that shipments could be by seas, the cost used in this study was USD 0.10 per target, a figure which seems a reasonable compromise. If the transport cost were doubled to USD 0.20 per target, the cost of the project would rise to USD 86.5 per km2. Shipping by sea is only feasible for large quantities, so, in order to keep costs low an in-country project would need to be implemented in several sites of the type studied and to buy in bulk. The only other specialist cost items—traps and extension materials—account for 0.1% and 1.0% of total costs, so cost increases in these items would have little overall impact on the project.
Looking at the other items for which price increases were costed in Table 11, the most significant is labour. Labour accounts for 24.1% of all costs, so increasing it by a third would result in an 8.0% increase in project costs. The high reliance on labour, whether government employees, locally sourced or provided by the community is a characteristic of the bait technologies. In this project in particular, the low overall cost was achieved by first training target/trap attendants who did much of the work and second by providing them with motorcycles, bicycles or hired local transport, thus reducing dependence on large project-owned vehicles. This was only possible because the small targets can be easily transported in a backpack, so that one person can carry up to 30 tiny targets. The impact of this is nicely demonstrated by comparing the two costings for preliminary surveys (Tables 3 and 4). The cost reduction from USD 23.6 to USD 5.1 per km2 is achieved by limiting senior staff involvement to supervision and by replacing the use of a 4x4 project vehicle with a good quality office motorbike and cheaper motorbikes used by the target attendants. The more expensive option would involve 21.6% higher costs (Table 11). Another unknown in relation to the preliminary survey is how extensive an area needs to be covered. In this project, some knowledge of the basic area already existed. If a new project was targeting a completely unknown area, or a region where previous tsetse surveys were out of date, the preliminary survey may need to be more extensive. A 50% increase in the area surveyed would increase overall costs by 6.4% (Table 11) and larger increases would have a linear impact on cost increases, if done at the same intensity. However, this survey was intended to help site targets, and it could be argued that a survey aimed simply at identifying future intervention sites might cover a larger area, but at a lower trapping intensity.
Ensuring that local populations want, support and understand the tsetse control measures is key to success [5,8,17]. Pre-intervention attitudes and knowledge in local populations highlighted that effective sensitisation would be a vital part of this project [21]. One unknown in the costings was, for a longer control operation, after how long and to what extent it would be necessary to visit communities again and reinforce the information provided at the start of the operation. As explained above, the village representatives continued to remind their communities of the purpose of the control activities, as did the trap and target attendants in the course of their work. For the latter, easy contact with local people was facilitated by their mode of travel using bicycles or motorcycles. Accordingly a compromise figure of three years was decided on—so that the sensitisation activity was effectively ‘depreciated’ over a 3 year period. If it were deemed necessary to repeat the operation in full after two years, the cost of the operation would increase by 9%. On the other hand, if nothing more had to be done for four years, the cost of the operation would be reduced by 4.6% (Table 11), which is a more likely scenario.
Attributing a share of the office overheads, for what was in many respects more a research than a control operation was also somewhat subjective. Sensitivity analysis indicated that varying the 25% proportion initially allocated from 20% to 33% had only a small impact on total costs (Table 11). Again, as was the case for the preliminary survey assumptions, the differences between the full cost of the office overheads, and those attributable to the control operation illustrates how costs are reduced if the research components are removed.
Thus, looking at individual components of the costs which might change in value or whose underlying assumptions might need changing, it is clear that foreseeable changes in the cost or quantities of a single item are unlikely to have a major effect on the costs. Of course, if several changes occur together, then a cumulative effect would be more significant. But overall, the costs can be described as robust. However, it is important to state that these costs apply to a specific tsetse control operation, and thus they only include what was done as part of that operation. Wider studies—such as surveys of trypanosomiasis in human and livestock populations were not undertaken as part of this tsetse control operation, although livestock were sampled as part of the separate research activities. The area was well known as a focus of HAT and controlling the disease in livestock was not a driver for the work. There were no separate training courses. The entomologists were already qualified in tsetse control. The target and trap attendants received their training in GPS use, trap placement and monitoring and target placement and repairs in the field, under supervision from the entomologists. On the sensitisation side, project staff were similarly trained by the anthropologist. The existence of the VHT’s within the Ugandan health service meant that at the village level, a support network for this type of work already existed. Long term monitoring will also provide more detail on target life and replacement rates. There may be some cost reductions in a second deployment as deployment costs in year two could fall because sites have already been identified and trap and target attendants are implementing well practiced routines.
Lastly, it is important to ask—how applicable will these costings be to similar work undertaken in other parts of Africa? Differences are likely to be experienced at three levels. The first one involves different prices, salary structures and a different organisational set up at the level of government services. Secondly, there are also important issues around economies of scale and shared resources to consider. For example, these costs assume the presence of a district entomologist who could allocate a costed share of his time to a particular vector control activity; shipping costs for targets are very dependent on scale, etc. These costings apply to a relatively small area—but should be regarded as ‘lean’, in the sense that almost all resources are fully used and no spare capacity is included. If even smaller areas were targeted, some extra travel from area to area might be required. Also, by spreading the cost of sensitisation over three years, the programme is implicitly assumed to go on for that long. Both of these levels are country and project specific. The third level is ecological, integrating a number of factors. Although the maintenance of T. b. gambiense HAT foci relies on the presence of a human reservoir, population densities can vary greatly: from under 20 per km2 as discussed above for the forest zone of Côte d’Ivoire [5] and the 500 per km2 estimated for the north-western Uganda study zone costed here. In the T. b. rhodesiense focus of south-eastern Uganda [8], where cattle have been shown to be the major reservoir [22] of the disease, the human population density at the time was 100 per km2. In areas with lower human population densities, tsetse habitat is likely to be more dense, access more difficult, terrain could be rugged, overnight stays or camping may be required and local labour be less easy to recruit. All of these factors will drive cost per km2 upwards and in areas of low human population density the cost per person protected will rise steeply. Ultimately, in some isolated or rugged areas it might not be possible to rely on all of the lower cost forms of transport, but the benefit of using cheaper and more portable targets will be maintained.
All the considerations discussed above (sensitivity to changes in price and assumptions, price and organisational differences, ability to harness economies of scale, accessibility and its links to human and livestock population density and tsetse habitat) would apply equally to any ground-based tsetse control technology, and to some extent to all vector control methods. Thus while the actual cost levels achieved in this exercise may not be replicable in every situation, the principles on which the cost savings are based will be: low cost delivery using motorbikes or bicycles and local labour together with a cheap and highly portable target with a high killing efficiency.
|
10.1371/journal.ppat.1004072 | Potent Dengue Virus Neutralization by a Therapeutic Antibody with Low Monovalent Affinity Requires Bivalent Engagement | We recently described our most potently neutralizing monoclonal antibody, E106, which protected against lethal Dengue virus type 1 (DENV-1) infection in mice. To further understand its functional properties, we determined the crystal structure of E106 Fab in complex with domain III (DIII) of DENV-1 envelope (E) protein to 2.45 Å resolution. Analysis of the complex revealed a small antibody-antigen interface with the epitope on DIII composed of nine residues along the lateral ridge and A-strand regions. Despite strong virus neutralizing activity of E106 IgG at picomolar concentrations, E106 Fab exhibited a ∼20,000-fold decrease in virus neutralization and bound isolated DIII, E, or viral particles with only a micromolar monovalent affinity. In comparison, E106 IgG bound DENV-1 virions with nanomolar avidity. The E106 epitope appears readily accessible on virions, as neutralization was largely temperature-independent. Collectively, our data suggest that E106 neutralizes DENV-1 infection through bivalent engagement of adjacent DIII subunits on a single virion. The isolation of anti-flavivirus antibodies that require bivalent binding to inhibit infection efficiently may be a rare event due to the unique icosahedral arrangement of envelope proteins on the virion surface.
| Dengue virus (DENV) is a globally important mosquito-transmitted human pathogen for which there is no approved vaccine or antiviral therapy. In recent years, the number and severity of DENV human infections have increased due to the expanded geographic range of the virus. Neutralizing antibodies are a key component of a protective natural and vaccine-induced immune response against human DENV infections. One recently described monoclonal antibody (E106) protects mice against infection of DENV-1 when administered before or several days after virus infection. Because of these results, we investigated the mechanism of action of E106 using a combination of structural and functional approaches. E106 engaged an epitope on domain III of the viral envelope protein that is a composite of two previously described epitopes. Unexpectedly, and in contrast to the intact IgG, Fab fragments of E106 were ineffective at neutralizing virus; this was explained by their weak micromolar affinity for virus particles. Our results suggest that neutralization by E106, our most potently inhibitory and protective anti-DENV MAb, requires bivalent binding of adjacent DIII subunits on a single virion. Immunization strategies with intact virions that skew the selection of neutralizing antibodies to those with bivalently binding properties could augment the potency of antiviral humoral responses against DENV and other flaviviruses.
| Dengue virus (DENV) infection in humans causes symptoms ranging from a mild febrile illness to a severe and sometimes fatal disease. Over 3.6 billion people globally are at risk for DENV infection, with an estimated 390 million infections annually and no currently approved vaccine or antiviral therapy [1]. DENV belongs to the Flaviviridae family of medically important positive-stranded RNA viruses. Within the DENV serocomplex, there is significant diversity, including four serotypes (DENV-1, -2, -3, and 4) that differ at the amino acid level of the envelope (E) protein by ∼25 to 40 percent and multiple genotypes within a serotype that vary by up to ∼3 percent [2], [3].
A humoral response against DENV infection is believed to contribute to lifelong immunity against challenge by the homologous serotype. In comparison, protection against a heterologous DENV serotype infection is more transient (∼6 months to two years) [4], [5], allowing re-infection and disease to occur with a heterologous serotype in hyper-endemic areas of the world. Estimates suggest that greater than 90% of severe cases occur during secondary infection with a heterologous DENV serotype, possibly because sub-neutralizing amounts of cross-reactive antibody facilitate viral entry into myeloid cells expressing Fc-γ receptors, a phenomenon termed antibody-dependent enhancement of infection (ADE) [6]. Antibody-mediated protection against homologous DENV infection correlates with a neutralizing antibody response directed predominantly against the viral E protein [7]. The ectodomain of E is comprised of three domains: domain I (DI), a central nine-stranded β-barrel that connects domain II (DII), which contains the fusion peptide at its distal end, and an immunoglobulin-fold like domain III (DIII) [6]–[9]. Although neutralizing antibodies have been mapped to all three domains of the E protein, many potently inhibitory anti-DENV mouse MAbs map to DIII [10]–[17], specifically to the lateral ridge or A-strand epitopes, and block flavivirus infection at a post-attachment stage, likely by preventing E protein dimer-to-trimer transitions that are required for viral fusion [11], [15], [18].
We recently described a highly therapeutic MAb, DENV-1-E106 (hereafter termed E106), which neutralized infection of strains corresponding to all five DENV-1 genotypes and protected against lethal DENV-1 infection when administered as a single dose even four days after virus inoculation [13]. To understand the basis for the potency (plaque reduction neutralization titer (PRNT50) of 0.6 ng/ml) [13] and specificity of this MAb, we solved the crystal structure of E106 Fab in complex with DENV-1 E DIII. Our analysis revealed a small antibody-antigen interface with contact residues corresponding to two previously characterized DIII epitopes. Remarkably, a ∼20,000-fold disparity in neutralization by intact IgG and Fab correlated with distinct abilities to bind intact virions. Our results are consistent with a model in which our most potently inhibitory and therapeutic DENV-1 MAb requires bivalent binding through dual and simultaneous engagement of two antigen binding sites on a single virion to neutralize infection. E106 is therefore one of only a few unique antibodies described to date where effective neutralization requires a bivalent binding mechanism, and is the first such characterized MAb directed against flaviviruses.
E106 is a sub-complex specific therapeutic MAb that binds to DENV-1 and DENV-4 infected cells and neutralizes infection of all five DENV-1 genotypes efficiently (EC50 ranging from 1 to 50 ng/ml), without neutralizing DENV-2, 3 and 4 serotypes or WNV [13]. The 2.45 Å crystal structure of E106 Fab bound to DIII revealed only nine contact residues, from the A-strand (K307, K310), the end of the B-strand (K325, Y326), and the connecting BC (E327, T329, D330) and DE (K361, E362) loops (Fig. 1A and Table 1); these results are consistent with prior mapping data by yeast surface display, which implicated five of these residues as essential recognition determinants (K310, T329, D330, K361 and E362) [13]. Yeast surface display results also implicated G328, P332, and P364 in E106 binding, and mutation of any of these residues would likely result in an altered presentation of the direct contact residues. Charge reversals at either E384 or K385 in the FG-loop also diminished E106 recognition, and this loop is adjacent to the primary E106 epitope. Overall, the contact residues contributed 24 van der Waal interactions, 14 hydrogen bonds, and 10 electrostatic interactions to the interface (Table S1). The E106 structural footprint represents a unique composite of previously identified DIII-specific neutralizing epitopes on flaviviruses including the lateral ridge (N-terminal region, BC, DE and FG loops) [19], [20] and A-strand epitopes [10], [21]. DENV-1 DIII was engaged by 11 heavy chain residues from CDR1 (I30, G31, Y32 and Y33), CDR2 (N52, E50, and R53), and CDR3 (R95, I196, N97 and W98) (Fig. 1B, top panel) and four light chain residues from CDR1 (D30, D32), CDR2 (E50), and CDR3 (L94) (Table S1).
A comparison of the DIII structure in complex with Fab versus unbound DIII revealed small conformational differences, with a root mean square displacement of 0.9 Å in the α-carbons over 98 residues. Of the DIII residues that directly interacted with the E106 Fab, the greatest differences in α-carbon positions involved residue T329 (1.2 Å); this was significant because a recently identified E106 MAb neutralization escape mutant showed a T to A amino acid change at position 329 (Fig. 1B, bottom panel) and [22]).
The E106 structural epitope on DIII was characterized by a high shape complementarity score (Sc = 0.73, with a perfect fit being 1) [23], which is greater than typical antibody-antigen interactions (Sc = 0.64–0.68) [23] but similar to anti-flavivirus MAb-E protein interactions (Table 2). The combined surface area buried by the DIII-E106 Fab complex was ∼1,243 Å2 (Fig. 1C and Table 2) [24] which is less than most antibody-antigen (1,400–2,300 Å2) [25], [26] and anti-flavivirus MAb-E protein interactions (Table 2). Typical of many antibody-antigen complexes, the majority (∼70%) of the DIII-E106 Fab interface was contributed by the heavy chain (Fig. 1C), with a combined buried surface of 877 Å2 (401 Å2 of the heavy chain and 476 Å2 of domain III). The light chain contributed the remaining buried surface (172 Å2 of the light chain and 194 Å2 of domain III).
All nine DIII contact residues were conserved in the five DENV-1 genotypes (K361 is replaced by the conservative substitution R361 in genotype 5 strains); this likely explains why E106 neutralized infection of all five DENV-1 genotypes efficiently (Fig. 1D and 1E), and [13]). In comparison, only one of the nine contact residues (Y326) was conserved in DENV-2, DENV-3, DENV-4 or WNV, a finding that is consistent with virological data showing that DENV-1-E106 MAb neutralizes infection in a serotype-specific manner [13]. E106 binds to DENV-4 but not to DENV-2 or DENV-3 [13]. The number of conserved contact residues did not correlate with DENV serotype binding specificity (DENV-4 has four whereas DENV-2 and DENV-3 each have five, Fig. 1E), which may instead be accounted for by other factors, including differential maturation [27] or relative virion dynamics [28].
To determine the significance of the small buried interface of our E106-DIII complex, we investigated E106 binding to DIII by surface plasmon resonance (SPR). Increasing concentrations of purified DENV-1 DIII monomer were flowed over immobilized E106 Fab (Fig. 2A). Equilibrium analysis surprisingly revealed a micromolar affinity (4.8±2.1 µM) for this interaction. A similar result was observed when increasing concentrations of DENV-1-DIII monomer was flowed over immobilized intact E106 MAb in the solid phase (3.2±0.8 µM, Fig. 2B); this experiment eliminates the possibility that papain cleavage and the removal of the E106 Fc region altered the monovalent binding parameters. Binding to the ectodomain of DENV-1 E (DI-DII-DIII) also appeared equivalently weak (1.1±0.1 µM). As an independent measurement of affinity, we performed isothermal titration calorimetry under similar experimental conditions as SPR by injecting DENV-1 DIII into a solution of E106 Fab. Using this method with completely unmodified proteins we again measured micromolar affinity for the E106 Fab/DIII interaction (KD = 0.7±2 µM) (Fig. 2C).
The micromolar monovalent affinity of the highly therapeutic E106 antibody was unanticipated in light of its picomolar inhibitory activity (4±2 pM or 0.6±0.3 ng/mL; our most potent neutralizing anti-flavivirus MAb isolated to date); as a comparison, our therapeutic DIII-specific anti-WNV MAb E16 (inhibitory activity 30 to 80 pM or 4–18 ng/mL), which has advanced to human clinical trials [29], has a monovalent affinity of 3.4 nM [30]. We hypothesized that while E106 MAb potently inhibited DENV-1 infection, Fabs should lack this activity. To test this, we compared the ability of Fab and intact IgG from E106 and E103, a lateral ridge DIII-specific DENV-1 neutralizing antibody [13], to inhibit infection. While monovalent E103 Fab showed a 114-fold decrease in neutralization potency compared to the intact bivalent IgG, the Fab of E106 Fab showed a remarkable 18,150-fold decrease in inhibitory activity when compared to intact IgG (Fig. 3A).
To investigate this observation further, we performed a virion-binding assay by ELISA. DENV-1 virions (strain 16007) were captured with humanized DIII A-strand-specific antibody (DENV-1-E50) [31] and then detected with Fab or intact IgG of E103, E106, or WNV E16 (negative control). Notably, the amount of virus detected with E106 IgG was indistinguishable from E103 IgG (Fig. 3B). In comparison, E106 Fab bound virions to a significantly lower level (P<0.0001) at all concentrations tested, than those derived from E103 (Fig. 3B). Thus, disparate neutralization of E106 MAb and Fab correlated with discordant binding to DENV-1 virions and was consistent with the biophysical measurements: monovalent binding by DENV-1-E106 is surprisingly inefficient given the potent inhibitory activity of intact antibody.
Based on these experiments, we hypothesized that efficient neutralization of DENV-1 infection by E106 required bivalent binding. Using bio-layer interferometry (BLI), we measured the affinity and kinetics of MAb and Fab binding to intact DENV-1 virus particles [32] (Fig. 4). E106 MAb (Fig. 4A) bound DENV-1 particles with an apparent affinity (avidity), KDapp of 13±2 nM. In contrast, E106 Fab had an affinity of KD>1 µM, with a rapid dissociation rate (t1/2<2 sec) that was at least 800-fold faster than E106 MAb (t1/2>400 sec) (Fig. 4B). These results contrast with more comparable binding of E103 MAb (KDapp of 0.8±0.1 nM) (Fig. 4C) and E103 Fab (KD of 7±1 nM) to DENV-1 particles (Fig. 4D). Importantly, the binding affinities of E106 Fab engaging isolated DIII measured by SPR and ITC is remarkably similar to that observed by BLI for the binding to DENV-1 virions, suggesting that our structurally defined DIII epitope corresponds to the entire virion surface recognized by a single Fab.
We next investigated the time- and temperature-dependence of E106 neutralization, as this analysis can provide information as to the relative accessibility of epitopes [28], [33]. Changes in the time or temperature of incubation did not appreciably affect E106 neutralization (Fig. 5A and B). By performing pre- and post-attachment neutralization assays, we found that, similar to several other potently neutralizing DIII-specific antibodies against flaviviruses [15], [30], [34], E106 can neutralize infection even after virus attaches to cells (Fig. 5C). Finally, we tested the neutralization of E106 MAb as a function of the maturation state of the virus. DENV virions are a heterogeneous mixture of immature, partially mature and fully mature virions, with immature virions being generally less or non-infectious. In this assay, E106 MAb neutralization proved to be independent of the maturation state of the virus (Fig. 5D). In comparison, neutralization by E60, a DII fusion-loop-specific MAb, was sensitive to virion maturation, as seen previously [27].
In contrast to non-enveloped viruses where a structural understanding of bivalent antibody binding has emerged [35]–[37], there currently is no such data for icosahedral enveloped viruses including flaviviruses. To address how E106 might recognize DENV-1 bivalently we docked our structure onto the cryo-EM derived model of the mature DENV virion (Fig. 6) [38]. While the E106 epitope is predominantly exposed on all 180 E protein monomers (Fig. 6A), unimpeded binding is readily apparent only on monomers in the 3-fold and 2-fold symmetry axes, similar to what we observed for E16 binding to WNV [20], [39]. However, minor reorientation of DIII subunits on the inner 5-fold symmetry axis would allow for up to two E106 Fabs to bind there at the same time as three Fabs could bind to the outer 5-fold (2-fold) related epitopes. We measured the distance separating the docked Fab CH1 domain C-termini with the expectation that distances greater than 50 Å would be unlikely spanned by 16 hinge residues [36], [37]. This analysis indicated the possibility for limited bivalent bridging, with the primary candidates being adjacent outer 5-fold epitopes (49 Å CH1 separation) (Fig. 6B and D) as well as adjacent inner and outer 5-fold epitopes (24 Å CH1 separation) (Fig. 6C and D). These epitopes are 85 Å and 79 Å apart, respectively, within the expected reach of a single IgG molecule (117–134±40 Å) [40]. We also examined the E106 epitope on the recent cryo-EM reconstruction of DENV-2 at 37°C [41], [42]; this ‘bumpy’ virion supports a similar model of bivalent binding to DIII on the 5-fold and 2-fold symmetry axes.
Epitope mapping studies have enhanced our understanding of the mechanisms of virus neutralization and identified sites on the E protein of flaviviruses that are targeted by neutralizing antibodies [7]. These include the lateral ridge of DIII of WNV and JEV [19], [20], the A-strand of DIII of DENV [10], [21], the CC′ loop of DIII of DENV-1 [33], the fusion loop of DII of WNV and DENV [43], a DI epitope of DENV-4 [11], and a complex epitope centered at the hinge of DI and DII on WNV [44] and DENV [45], [46]. Here, we describe a composite epitope, comprised of regions of the lateral ridge and A-strand of DIII that is targeted by the therapeutic MAb E106. DIII residues contacted by E106 were highly conserved among DENV-1 genotypes but variable in other DENV serotypes. Consistent with this, E106 potently neutralized all five DENV-1 genotypes, but not other DENV serotypes nor WNV [13].
The E106 Fab-DIII complex was characterized by a small-buried interface, which correlated with an unexpectedly weak micromolar affinity, as determined both by SPR and isothermal calorimetry. We found no evidence for E106 binding to residues outside of DIII as the binding affinity to recombinant DIII appears to be very similar to the binding of Fab to E ectodomain protein or virions; consistent with this, neutralization escape studies only identified residues in DIII [22]. Monovalent E106 Fab poorly neutralized DENV-1 compared to intact E106 IgG, and this finding correlated with poor binding of Fab to intact virus. Although our structural models suggest that E106 can readily bind the isolated pre-fusion dimer and post-fusion trimer and possibly prevent the ∼70° transition of DIII that is associated with membrane fusion [47], [48], the inability of E106 Fab fragments to neutralize virus efficiently argues against this model. Our data are more consistent with a mechanism of neutralization that requires bivalent binding of E106 IgG to single virions, and cross-linking of E protein monomers in adjacent symmetry groups to prevent requisite E protein rearrangements during infection.
The measurement of micromolar monovalent affinity was unexpected given that E106 is our most potently neutralizing and protective anti-DENV-1 MAb (EC50 of 0.6 ng/ml against DENV-1 strain 16007), which is at least 10-fold more potent than our well-characterized DIII-lateral ridge-specific therapeutic MAb (E16) against WNV [20], [30]. Indeed, E106 had the lowest EC50 value of ∼500 anti-flavivirus MAbs (DENV-1, DENV-2, DENV-3, DENV-4, and WNV) generated to date in our laboratory [13], [15], [30], [49]. Is there a correlation between neutralization potency and E106 bivalent binding to single virions? The icosahedral arrangement of the E protein on the mature DENV virion displays 180 copies of the E protein. Our in silico modeling predicts that in one possible arrangement, up to 48 of these sites may be available for bivalent engagement by 24 intact E106 IgG. Since monovalent binding was insufficient for neutralization, bivalent binding to single virions could neutralize infection by inhibiting an essential stage of the virus lifecycle (attachment, entry, or fusion). Alternatively, bivalent binding across virions could neutralize DENV infection by aggregating virus. Our post-attachment studies suggest that E106 MAb was capable of neutralizing infection even after virus attached to the cell surface. Aggregation also appears less likely because the neutralization curves did not show a characteristic triphasic dose-response curve that was reported in inhibition studies describing antibody-virus aggregation [50]. We favor a model in which bivalent binding of E106 stabilizes and/or cross-links one or more E protein monomers in different symmetry groups, analogous to monovalent binding of WNV CR4354 MAb [44], and thus prevents radial expansion and rearrangement that is requisite for fusion of viral and host endosomal membranes [48], [51].
E106 is one of very few MAbs that have been shown to require bivalent binding for efficient virus neutralization [35], [36], [52], and the first one directed against a flavivirus. While antibodies can be multivalent, with the potential to bind to virus particles with high avidity, the relatively small number of bivalent binding MAbs described to date may be attributed to the following: (i) the limited number of epitopes displayed on a single particle for some viruses; (ii) the position and orientation of epitopes that are beyond the reach of a bivalent MAb, which is limited by torsional flexibility of its hinge; (iii) the radius of curvature of virions, which may restrict the accessibility of epitopes by MAb (iv) post-translational modification (e.g. N-linked glycans) of virions that may hinder bivalent MAb engagement; (v) immunization protocols that rely on isolated recombinant envelope proteins, rather than envelope proteins in the structurally unique form of the intact virion; and (vi) assays that do not screen for bivalent neutralizing MAbs. E106 was generated after priming and boosting with infectious DENV-1 [13]. The repetitive E antigens in the icosahedral orientation of the virion may have facilitated selection of low monovalent affinity yet high avidity antibodies. While some antibodies against HIV have been described as bivalent, they actually are bispecific, with each arm binding distinct epitopes [53]. This is likely due to the paucity of trimeric spikes on the surface of the virus (∼14) and their irregular spacing [54]. Although several MAbs have been proposed to require bivalent binding for efficient virus neutralization [55], [56], compelling evidence is presented only for the non-enveloped positive strand RNA viruses, specifically, human rhinovirus 2 [36] and 14 [35], [52] and the rabbit hemorrhagic disease calicivirus [37]. It may be that the quasi-icosahedral arrangement of the flavivirus envelope creates a landscape that permits limited bivalent MAb engagement.
Bivalent engagement of the virion by antibodies could be an important concept for DENV vaccine development. Immunity against DENV may not be achieved optimally using a subunit vaccine approach, as isolated E protein monomers may not induce antibodies that require bivalent binding for strong binding and neutralization. Analogously, some human MAbs against WNV bind a complex epitope at two independent positions on adjacent E protein monomers in different symmetry groups, which is only present on an intact WNV particle [44]. Human MAbs isolated from DENV-infected individuals are believed to recognize similar quaternary epitopes in E that are present only in the context of the intact DENV virion [45], [46]. Given that E106 MAb was our most potent and therapeutic anti-DENV-1 MAb, strategies that enhance the likelihood of generating and identifying neutralizing antibodies that function through bivalent binding mechanisms may improve the potency of inhibitory humoral responses against DENV and other flaviviruses. Regardless, an understanding of the structural and mechanistic basis for the neutralization activity of E106 MAb provides new insights into the humoral response against flaviviruses.
DENV-1 DIII (residues 293 to 399 of the E protein of strain 16007) was expressed in bacteria and re-folded oxidatively from isolated inclusion bodies as described previously [13]. Fab fragments of E106 were prepared using immobilized papain resin according to the manufacturer's instructions (Pierce). MAb (5 to 10 mg) was digested for 18 hours at 37°C, and passed over a protein A agarose resin to remove Fc fragments and undigested MAb and purified on a S-75 size exclusion chromatography column equilibrated in 20 mM HEPES pH 7.4 and 150 mM NaCl. Antibody–antigen complexes were formed by mixing papain-generated, gel filtered E106 Fab fragments with DIII at a ratio of 1.2∶1 and crystallized by the hanging drop vapor diffusion method at a total protein concentration of 14 mg/mL in a solution of 22% PEG 6,000, and 0.1 M MES pH 5.0 (final pH 5.7). Crystals (in 1 µL crystallization drops) were cryoprotected by the addition of 0.2 µL aliquots of cryobuffer (in 23.5% PEG 6,000, 0.1 M MES pH 5.0, final pH 5.7, and 20% glycerol), then transferred to a fresh drop of cryobuffer and rapidly cooled in liquid nitrogen. X-ray diffraction data were collected at ALS beamline 4.2.2 (Lawrence Berkeley Laboratories) at a wavelength of 0.976 Å at 100 K with a CCD detector, and indexed and scaled in HKL2000 [57]. The crystals diffracted to 2.45 Å resolution and belonged to the space group P212121 with unit cell dimensions of a = 82.7 Å, b = 91.8 Å, c = 92.6 Å, with one E106 Fab-DIII complex per asymmetric unit. Crystallographic phasing was obtained by molecular replacement using the program Phaser [58] and the coordinates of DENV-1 DIII (Protein Data Bank (PDB) 3EGP) and the Fab fragment of CTM01 IgG (PDB 1AD9). Iterative model building and refinement was performed using Coot [59] and Refmac [60] and later Phenix [61]. The final structure was refined to Rcryst = 18.9% and Rfree = 23.9%. The final model includes DENV-1 DIII amino acid residues 297 to 394, E106 heavy chain residues 1 to 214 (Chothia numbering), and light chain residues 1 to 213. The atomic coordinates and structure factors of E106 Fab bound to DENV-1 DIII (CSGID target number IDP00272) have been deposited in the Protein Data Bank (www.rcsb.org) under PDB accession number 4L5F. Structural figures were prepared using CCP4MG [62] and Pymol [63] (surface representation using 1.4 Å solvent probe) and where shown, spheres represent van der Waal radii, vdw * 1.1.
Monovalent antibody affinity analysis was performed using SPR (BIAcore T100, GE Company) and ITC (VP-ITC instrument, Microcal) at 10°C in 50 mM HEPES, pH 7.5 and 100 mM NaCl. For SPR, E106 MAb or Fab was immobilized at low concentrations (∼500 Response Units) to a CM5 chip (GE healthcare) using amine-coupling chemistry. Bacterially-expressed DIII (residues 293–399) of DENV-1 (strain 16007) was injected at a flow rate of 65 µl/min at concentrations ranging from 0.2 µM to 500 µM for three minutes to saturate binding and then allowed to dissociate for seven minutes. The half-life of the monovalent interaction was short and did not require additional regeneration of the chip surface in preparation for the next DIII injection. The observed binding curves were double referenced to a non-reactive antibody (WNV E16) as well as buffer in the absence of DIII. Curves were analyzed by a steady-state fit for a 1∶1 interaction, and a nonlinear least squares fit was used to evaluate the fit of the curve to the observed data. Alternatively, 500 response units (RU) of DIII were immobilized onto a CM5 chip and E106 Fab fragments were injected to saturation. Affinity measurements of E106 for the DENV-1 E ectodomain (DI-DII-DIII) were performed such that insect-derived DENV-1 E glycoprotein (ProSpec-Tany TechnoGene Ltd.) was in the stationary phase and E106 Fab fragments were in the mobile phase, in order to conserve limited protein and avoid avidity affects. Additional regeneration was not necessary because of the short half-life for the interaction. WNV E ectodomain was used as a negative control for E106 binding. Analysis was performed as with the DIII described above. ITC experiments were performed such that 4 to 8 µL of DENV-1-DIII protein (90 to 110 µM) was injected into 1.4 mL of E106 Fab protein (6 to 7 µM) over a total of 36 injections. The titration data were integrated and normalized in Origin (Microcal) to determine the reaction stoichiometry, n, and equilibrium constant Ka ( = K−1d).
Plaque reduction neutralization tests (PRNT) and pre- and post-attachment neutralization assays were performed with DENV-1 strain 16007 on Vero cells as previously described [13], [64]. Binding of intact MAbs or Fabs (E103, E106, and a negative control WNV E16) to DENV-1 virions (strain 16007) was detected by capture ELISA [13], [64]. Briefly, humanized DENV-1 E50 MAb (subcomplex DIII A-strand specific antibody) was coated at 2 µg/ml on MaxiSorp (Nunc) polystyrene 96-well microtiter plates in a sodium carbonate (pH 9.3) buffer. Plates were washed three times in wash buffer (PBS with 0.02% Tween 20) and blocked for one hour at 37°C with blocking buffer (DMEM with 10% FBS). DENV-1 virions (2.5×105 PFU) diluted in DMEM with 10% heat-inactivated FBS were captured on plates for two hours at 37°C. Wells were washed thrice with blocking buffer and DENV-1 MAb or Fab was then added at 100 µg/ml and 4-fold serial dilutions to duplicate wells and incubated for two hours at 37°C. Plates were washed five times and then sequentially incubated with goat anti-mouse (whole molecule) IgG-HRP (Sigma, St Louis, MO) and tetramethylbenzidine substrate (Dako). The reaction was stopped with the addition of 2 N H2SO4 to the medium, and emission (450 nm) was read using an iMark microplate reader (Bio-Rad).
A plasmid expressing the C-prM-E genes of DENV-1 (strain 16007) was co-transfected into HEK-293T cells with a plasmid encoding a WNV replicon expressing GFP. Transfected cells were incubated at 30°C and RVP harvested at 72 and 96 hours post-transfection, filtered through a 0.2 µM filter, and stored aliquoted at −80°C. DENV-1 RVP were incubated with serial dilutions of MAb under conditions of antibody excess at 4°C, 37°C, or 40°C for one or five hours. Subsequently, MAb-RVP mixtures were added to Raji-DCSIGNR cells and incubated at 37°C for 48 hours. Infected cells were assayed for GFP expression using a BD FACSCalibur flow cytometer as described [28].
All bio-layer interferometry studies were performed in PBS buffer supplemented with 1 to 2 mg/ml BSA (PBS-B) at 25°C using an Octet Red biosensor system (ForteBio). DENV-1 reporter virus particles (RVPs) (Western Pacific 74 strain) were produced as previously described [65]. To purify virus particles, RVP production supernatant was harvested, clarified through a 0.22 µm filter (Corning), and PEG precipitated using 7.5% PEG 8000 (Sigma). The virus particles were further purified through two 20% sucrose cushions before resuspension in HBS. Samples were stored at −80°C and gently thawed prior to use. RVPs were loaded onto streptavidin (SA) biosensor tips using a human monoclonal antibody against DENV-1 (1H7.2, the anti-prM antibody, a gift from James Crowe), which was captured using a biotinylated goat anti-human polyclonal antibody (GAH Fc, Southern Biotech). Briefly, GAH Fc was diluted to 5 µg/ml in PBS-B and bound to SA sensor tips for five minutes. Following a brief rinse in PBS-B, 1H7.2 (5 µg/ml in PBS-B) was captured for five minutes. After another brief rinse, DENV-1 RVPs diluted to 10 µg/ml (or 50 µg/ml (E106 Fab)) were loaded for 45 minutes followed by a five-minute stabilization. Antibody association was measured for up to 10 minutes followed by dissociation for 20 minutes (E106) or 45 minutes (E103) in buffer. Non-specific binding was assessed using sensor tips without RVPs as well as using sensor tips loaded with retroviral pseudotypes (Lipoparticles) containing only endogenous cell surface receptors (no viral envelope protein). Data analysis was performed using Octet Data Analysis v6.4 (ForteBio). Binding kinetics were analyzed using a standard 1∶1 binding model.
|
10.1371/journal.pcbi.1003747 | Propagating Waves of Directionality and Coordination Orchestrate Collective Cell Migration | The ability of cells to coordinately migrate in groups is crucial to enable them to travel long distances during embryonic development, wound healing and tumorigenesis, but the fundamental mechanisms underlying intercellular coordination during collective cell migration remain elusive despite considerable research efforts. A novel analytical framework is introduced here to explicitly detect and quantify cell clusters that move coordinately in a monolayer. The analysis combines and associates vast amount of spatiotemporal data across multiple experiments into transparent quantitative measures to report the emergence of new modes of organized behavior during collective migration of tumor and epithelial cells in wound healing assays. First, we discovered the emergence of a wave of coordinated migration propagating backward from the wound front, which reflects formation of clusters of coordinately migrating cells that are generated further away from the wound edge and disintegrate close to the advancing front. This wave emerges in both normal and tumor cells, and is amplified by Met activation with hepatocyte growth factor/scatter factor. Second, Met activation was found to induce coinciding waves of cellular acceleration and stretching, which in turn trigger the emergence of a backward propagating wave of directional migration with about an hour phase lag. Assessments of the relations between the waves revealed that amplified coordinated migration is associated with the emergence of directional migration. Taken together, our data and simplified modeling-based assessments suggest that increased velocity leads to enhanced coordination: higher motility arises due to acceleration and stretching that seems to increase directionality by temporarily diminishing the velocity components orthogonal to the direction defined by the monolayer geometry. Spatial and temporal accumulation of directionality thus defines coordination. The findings offer new insight and suggest a basic cellular mechanism for long-term cell guidance and intercellular communication during collective cell migration.
| The fundamental mechanisms underlying intercellular coordination during collective cell migration remain elusive despite considerable research efforts. We present a novel analytical framework that considers spatiotemporal dynamics across several traits. Our approach was applied to discover new modes of organized collective dynamics of cancer and normal cells. Following disruption of a cell monolayer, a propagating wave of coordinated migration emerges as clusters of coordinately moving cells are formed away from the wound and disintegrate near the advancing front. Activation of Met signal transduction by hepatocyte growth factor/scatter factor, master regulators of cell motility in malignant and normal processes, generates coinciding waves of cellular acceleration and stretching that propagate backward from the wound front and trigger a delayed wave of directional migration. Amplified coordination is intrinsically associated with enhanced directionality suggesting that even a weak directional cue is sufficient to promote a coordinated response that is transmitted to cells within the cell sheet. Our findings provide important novel insights on the basic cellular organization during collective cell migration and establish a mechanism of long-range cell guidance, intercellular coordination and pattern formation during monolayer wound healing.
| Collective cell migration plays an essential role during embryonic development, wound healing, tissue repair and cancer metastasis [1]–[4]. Directional migration and intercellular coordination are two cellular traits that play major roles in collective cell migration. It was previously demonstrated that collective cell migration relies mostly on a directional signal that stems from the moving cluster rather than from external cues [5], directionality might be correlated with metastatic potential [6], and is enhanced by growth factors [7]. Directionality and coordination are affected by substrate stiffness [8], topographic cues [9], cell density [10], and are linked to mechanical intercellular cooperation [11]–[13]. Vitorino et al. defined 3 modules for collective cell migration: motility, directionality and coordination, and classified genes that affect each of these modules [14]. Despite these vast research efforts, the physical mechanisms underlying intercellular coordination are still unknown. We present here a rigorous analytical framework to investigate the dynamic relations between different physical variables of migrating cells over time and space, which suggests new insights regarding the mechanisms that account for directionality and intercellular coordination.
Capabilities of collective behaviors of cancer cells involve some modes of inter-cellular communication, social networking and cooperation between cells, which regulate dissemination, proliferation and colonization within the body [6], [15]–[19]. Revealing common and different cellular and molecular mechanisms that govern intercellular coordination of normal and cancer cells may lead to new therapeutic paradigms to target intracellular signaling processes and intercellular communication in cancer metastasis [20].
In vitro wound healing assays involve the partition of a cell monolayer into two separated segments by scratching. We studied the collective dynamics of such a monolayer, as these segments move towards each other to close the wound. As the wound edge advances, the cell monolayer is moving forward. Here we focus on global organization during collective cell migration of DA3 mammary tumor cells and MDCK normal epithelial cells and the effects of Met signaling activation by its ligand hepatocyte growth factor/scatter factor (HGF/SF), master regulators of cell motility in malignant and normal processes [21]–[24]. Growth factors play a central role in collective cell migration [7], [14], [25]. We choose to study HGF/SF-Met signaling effects on cell motility since: 1) it is a well characterized signaling pathway in different aspects of cell motility; 2) the molecular tools to study this pathway are well established enabling to effectively inhibit Met signaling [26]; 3) The relevance of this signaling pathway to cancer and especially metastasis [27], [28].
We previously established that HGF/SF induces in tumor (DA3) migrating cells a backward propagating wave of increased velocity that is associated with shape modification into larger and more elongated cells. This wave propagates from the wound edge backward into cells located farther away from the advancing front [29] (Supporting Text SI1 in Text S1, also shown for MDCK cells here).
Here we reveal that collective cell migration is more intricate than was previously reported. We applied specially-designed analytical techniques to investigate the spatiotemporal dynamics of acceleration, cellular stretching (strain rate), directionality and coordination, and their associations and temporal order. We found that these quantities can exhibit wave-like phenomena, which move backward in respect to the monolayer's moving direction - away from the wound edge. The wave's profile is similar to a pulse: low acceleration is observed for the cells that are close to the edge, the acceleration increases for the cells that are behind them, and decreases again for cells that are farther away. The location of the maximal acceleration moves backward, i.e., in a direction opposite to the motion of the monolayer itself. We refer to this phenomenon as backward propagating wave, and it is sketched in Fig. 1A.
First, we discovered the emergence of a backward propagating wave of coordination. The wave reflects the formation of clusters of cells moving with coordinated trajectories. These clusters are measured by applying a region-growing segmentation algorithm that gradually merges adjacent trajectories to spatial clusters. Second, we show that HGF/SF treatment generates backward propagating waves of both cellular acceleration (escalating motility) and stretching (strain rate). We refer to these coinciding waves as a wave of acceleration and stretching. Thirdly, we uncover the emergence of a backward propagating wave of directionality, following the HGF/SF-induced wave of acceleration and stretching with an approximated one hour lag. The term ‘directionality’ is defined as the ratio between the cells' velocity towards the front, and the velocity parallel to the wound edge. Finally, we present the association between the waves of different cellular properties and suggest that the wave of directionality may be responsible for the amplified magnitude of the wave of coordination. Our data suggest that increased velocity leads to enhanced intercellular coordination: motility is increased due to acceleration and stretching that in turn increase directionality by temporarily diminishing the velocity components orthogonal to the overall direction defined by the monolayer geometry; accumulation of directionality over space and time thus defines coordination.
The different physical properties of the moving cellular monolayer were measured and averaged spatiotemporally to describe the collective dynamics. To quantify each attribute, the cellular area in each image was discretized to patches or “agents” and their trajectories were tracked. It is important to note that no position-switching between these agents has been observed (also reported in [29]). Acceleration, strain rate and directionality were computed along the migration trajectories of each of the individual “agents”, as detailed below.
Acceleration (increasing motility) was calculated as the change in the velocity along the trajectory (Fig. 1B), ; this local time-derivative was performed by computing the acceleration of the agent along its trajectory using the preceding and following time-frames.
Strain rate is a measure for the local deformation of an object, typically caused by non-uniform force acting on the object which results in a non-uniform stretching. Strain rate was measured as the local spatial derivative of the speed , where x is the local migration direction of each agent. It was calculated as the difference between the velocities of the agents ahead and behind the agent of interest, according to its local direction. Assuming cellular cohesiveness and mass conservation, it is an implicit measure for cellular deformation rate [30], thus cell stretching was taken to be proportional to the strain rate along the trajectory (Fig. 1C).
Directionality is defined as the absolute value of the ratio of the velocity component toward- and parallel- to the wound edge (Fig. 1D). The velocity components are measured with respect to the direction of the wound edge. Higher values of the ratio between the perpendicular and parallel components indicate that the motion is more directed towards the wound edge, while smaller values mean a “noisier” motion. Values lower than 1 mean that the dominant motion is parallel to the wound edge (not closing the wound), while values above 1 indicate a more efficient healing process as the directionality increase.
Intercellular coordination is a measure of the collective migration, rather than a single cell's. It is defined as the fraction of cells migrating in coordinated clusters within the monolayer. Explicit detection of these clusters was performed by applying image segmentation on a dense grid of trajectories, thus identifying and grouping similar trajectories of adjacent agents. It is important to note that directionality and coordination are not necessarily related: a pair of adjacent cells can move coordinately but with poor directionality, or migrate less coordinately with higher directionality (Fig. 1E). Full technical details on the measures that were used are given in Materials and Methods and in Supporting Text SI4 in Text S1.
The results are organized as follows. Next, we present the emerged backward propagating waves of acceleration & stretching, directionality and coordination. Then, to assess the hypothesis that acceleration and strain rate lead to directionality that ultimately defines intercellular coordination, we apply spatiotemporal correlation analysis to reveal the temporal order between these waves.
Further analysis revealed that HGF/SF induced coinciding waves of cellular acceleration and of cellular stretching (strain rate). Average acceleration and stretching were computed as functions of the distance from the wound edge. This was done by averaging over all the agents belonging to each of the parallel layers at different distances from the edge (more details in Supporting Text SI4 in Text S1).
Figure 3 shows spatiotemporal maps (kymographs) of acceleration and strain rate for DA3 cells. Each element (t,d) in the map shows the average acceleration, over time interval Δt = 14.5 minutes (1 frame), (Figs. 3A and 3C) and the average strain rate (Figs. 3B and 3D), at time (t) for all the agents of a layer of width Δd = 12.4 µm (10 pixels), located at a distance (d) from the wound edge. Consecutive time projections (columns) of the spatiotemporal maps for HGF/SF treatment further illustrate the wave-like dynamics of the acceleration and strain rate (Figs. 3E and 3F). Comparison between the collective migration in response to HGF/SF and the control reveals that the wave of acceleration and stretching is generated as a response to HGF/SF treatment, while in the control case the acceleration and strain-rates spread from the edge inwards in a smooth manner without a distinct wave front. Comparison between the time projections of the spatiotemporal maps for acceleration (Fig. 3E) and strain rate (Fig. 3F) reveals the accurate coinciding of the waves of increasing motility (acceleration) and stretching (strain rate). These waves propagate at roughly twice the speed of the advancing front edge, consistently with previously published results [31].
A wave of enhanced directionality emerges following the HGF/SF-induced wave of acceleration and stretching. Here the directionality is measured for each layer at a distance (d) at each time (t) from the wound edge by - the ratio between the average speed towards the wound edge and the average speed in the parallel direction. More specifically, and are the average perpendicular speed and the average of the absolute value of the parallel speed, respectively, of all the agents that belong to a layer at distance (d) and at time (t). Note that since cells do not move backward from the wound edge, . Figures 4A and 4C show kymographs of this directionality measure for DA3 cells. Consecutive time projections (columns) of the spatiotemporal maps further illustrate the wave-like dynamics of the directionality in response to HGF/SF (Figs. 4B and 4D). Comparison between the collective migration in response to HGF/SF and the control reveals that the directionality wave is generated as a response to HGF/SF treatment.
We find the emergence of three waves of collective migration during the wound healing process: a wave of acceleration and stretching, a wave of directionality and a wave of coordination. To address the hypothesis that acceleration and strain rate lead to enhanced directionality and coordination we proceed by suggesting a simplified theoretical model demonstrating a possible mechanism that links cell stretching to directionality and by quantifying the associations between acceleration and strain rate, directionality and coordination waves.
The waves described above and their associations also emerge, under the same conditions with striking similarity, during collective migration of MDCK epithelial cells (Fig. 7), the most common model system for 2D collective cell migration.
Additional control experiments to demonstrate the specificity of the HGF/SF induced wave to Met signaling revealed that DA3 cells treated with either HGF/SF with Met inhibitor (PHA) or with Met inhibitor alone, have similar, though slightly reduced, characteristics compared to control cells (See Supporting Text SI3 in Text S1 for details). These results demonstrate that the HGF/SF-induced wave is specific to Met signaling, serving as a model system to study the effects of growth factor on collective cell migration.
Both wound healing and cancer metastasis are complex processes that require cooperation among many cells to efficiently migrate while keeping the cell sheet intact. We found a new mode of collective migration, a backward propagating wave of coordination that is described as clusters of coordinately moving cells that are formed a substantial distance away from the wound edge and disintegrate closer to the advancing front. We also found that HGF/SF, as a growth factor model, plays an important role in generating an orchestrated wound healing process. A backward propagating wave of acceleration and cellular stretching (strain rate) is generated in response to HGF/SF. This wave leads to the emergence of a wave of enhanced directionality that eventually results in an amplified wave of coordination (Figs. 7A and 7B and 7C).
Backward propagating waves of increased velocities [29], [31], [33] and strain rate [31] were previously reported in epithelial cells. Ng. et al demonstrated that front cells were more coordinated than cells farther from the cell front, general coordination increased with time, and a gradual rise in coordination was observed for distal cells [8]. Our analyses looked at dynamic relationships between several physical variables incorporated across many experiments over time and space to reveal new insights on mechanisms that account for intercellular coordination.
Our data are consistent with the effect of FGF in a dose-dependent manner, suggesting that growth factors mainly affect directionality [14]. These effects are presumably mediated via stimulation of the shape and motility regulators such as alteration of the actin cytoskeleton by the Met/Gab1/Grb2 signaling pathway [34].
Based on these results we propose that directional migration results from cellular stretching forces, corresponding to the strain-rate. To test this hypothesis we devised a theoretical model demonstrating that stretch-rate can be transformed to a chemical directionality signal by its effect on binding/unbinding rates of cell-cell adhesion molecules. This phenomenon may be regulated by cellular inherent elasticity, cell-cell adhesions and self-propulsion. Notably, the shape deformation is consistent with previous studies linking directionality changes to RhoGTPases and cell morphology deformation: increase in RohGTPases leads to lamelliopodia-to-lobopodia transition (stretching) with enhanced directional and persistent motion [35]. Additional experimental effort will be made to validate this hypothesis in the collective migration setting.
The coordination wave occurs in both DA3 and MDCK cell lines under different serum condition, more prominently in the presence of HGF/SF, implying that HGF/SF-Met signaling plays a major, but not exclusive, role in mediating this phenomenon. The association between directionality and coordination accords with the idea that cell-substrate traction is produced by polarized lamellipodia, which tend to polarize neighboring cells in the same direction, eventually forming long-range polarization and intercellular coordination [36]. The data may imply that even a weak directional cue is sufficient to promote a coordinated response that is transmitted to cells within the cell sheet; basal activation of Met or other tyrosine kinase signaling could induce a weak wave of directionality that can explain the observed results. We conclude that the wave of coordination is an intrinsic trait in collective cell migration that is amplified in the presence of HGF/SF, speculatively by progressive mechano-sensing cell-cell communication mechanisms.
Isolated cells, or cells in small clusters respond to HGF/SF by rapid activation of cellular motility mechanisms [37] in an orderly manner; they spread, loose cell-cell adhesions, exhibit increased motility and spatial scattering [38]. Similar relations between cell-cell adhesions (which affect density), motility and intercellular coordination occur in confluent monolayers; Inhibition of cadherin-mediated cell–cell adhesions increased general motility but reduced directionality, persistence and intercellular coordination for confluent monolayers of MCF10A cells [8]. Increased monolayer density is associated with decreased motility and increased intercellular coordination, experimentally [10], [39]–[42], theoretically [43] and known to exist in other systems as well, such as bacterial populations [44].
It is well established that in response to HGF/SF, epithelial monolayers become sparser, maintain higher motility [29], and weaken cell–cell contacts [25]. Nevertheless, higher intercellular coordination was found here as response to HGF/SF (Fig. 8D). This observation may imply an alternative channel of mechanical intercellular communication [25], [45]. For example, a mechanism that is based on emerged collective sensing from individuals modulating their speed in response to local cues through social interaction with their peers [46] may explain our findings. We speculate that the wave of directionality has an important role in this phenomenon, and that increased motility coupled with a general directional cue introduced by the free edge is more prominent than the known density-motility-coordination intrinsic relations, thus leading to increased coordination.
The enhanced coordinated motility induced by HGF/SF-Met resembles its effect on the induction of tubulogenesis, where increased cell-cell association is crucial for tubule formation [21], [47]. Other processes, such as epithelial plasticity, are also known to be common to development and cancer [48]. Thus, the enhanced waves of directionality and coordination found here might imply that even in tumor cells, HGF/SF-Met activate similar pathways to those activated during embryogenesis.
|
10.1371/journal.pntd.0007521 | Absence of accessory genes in a divergent simian T-lymphotropic virus type 1 isolated from a bonnet macaque (Macaca radiata) | Primate T-lymphotropic viruses type 1 (PTLV-1) are complex retroviruses infecting both human (HTLV-1) and simian (STLV-1) hosts. They share common epidemiological, clinical and molecular features. In addition to the canonical gag, pol, env retroviral genes, PTLV-1 purportedly encodes regulatory (i.e. Tax, Rex, and HBZ) and accessory proteins (i.e. P12/8, P13, P30). The latter have been found essential for viral persistence in vivo.
We have isolated a STLV-1 virus from a bonnet macaque (Macaca radiata–Mra18C9), a monkey from India. The complete sequence was obtained and phylogenetic analyses were performed. The Mra18C9 strain is highly divergent from the known PTLV-1 strains. Intriguingly, the Mra18C9 lacks the 3 accessory open reading frames. In order to determine if the absence of accessory proteins is specific to this particular strain, a comprehensive analysis of the complete PTLV-1 genomes available in Genbank was performed and found that the lack of one or many accessory ORF is common among PTLV-1.
This study raises many questions regarding the actual nature, role and importance of accessory proteins in the PTLV-1 biology.
| Primate T-lymphotropic viruses type 1 (PTLV-1) are complex retroviruses infecting both human (HTLV-1) and simian (STLV-1) hosts. It has been shown that the persistence and pathogenesis of these viruses depend on the expression of small, accessory proteins.
A bonnet macaque (a monkey present in India) was found infected with STLV-1. The genome was sequenced and found quite divergent from the other STLV-1 genomes previously described. Intriguingly, this virus does not encode accessory proteins. Analysis of other available sequences found that most strains lack at least one accessory gene. Thus the importance and the role of these proteins in the PTLV-1 biology should be revisited.
| The Primate T-lymphotropic virus type 1 (PTLV-1) constitutes a group of deltaretroviruses infecting humans (HTLV-1) or non-human primates (STLV-1). Phylogenetic studies led to the definition of PTLV-1 viral subtypes [1]. Viruses belonging to the African PTLV-1 subtypes (i.e. subtypes b, d, e, f and g) are found both in humans and non-human primates (NHPs), and the human and simian strains are undistinguishable. Continuous zoonotic spillovers of these strains occur, following severe bites by infected non-human primates, or in the context of bushmeat hunting and handling [2–4]. Two subtypes are exclusively human: PTLV-1a is found in human populations scattered throughout the globe, while PTLV-1c is found in indigenous people of the Australomelanesian continent. In contrast, a group of STLV-1 has been described in macaques and great apes in Asia [5–7]. These viruses are genetically distant from other subtypes (a fact that led some researchers to consider some of them as a separate STLV, STLV-5) [8] and have never been found in humans to date.
HTLV-1 is estimated to infect at least 5–10 million people worldwide [1]. HTLV-1 is the etiological agent of many severe diseases, ranging from an aggressive lymphoproliferation, the adult T-cell leukemia/lymphoma (ATL), to inflammatory syndromes, such as a neurodegenerative disease called HTLV-1 associated myelopathy or tropical spastic paraparesis (HAM/TSP). Pathogenesis does not seem to be restricted to a certain HTLV-1 subtype; for instance, ATL cases have been reported in patients infected with HTLV-1a, -1b or 1c [9, 10]. STLV-1 is also oncogenic. ATL have been reported in many simian species, from Macaques to Gorilla [11, 12].
Deltaretroviruses are complex retroviruses. In addition to the canonical gag, pol, env retroviral genes, the PTLV-1 genome has a series of open reading frames (ORFs) encoding regulatory and accessory proteins [13–15]. Regulatory proteins are essential for viral expression and propagation both in vitro and in vivo. In contrast, accessory proteins are optional for viral expression in vitro, but required for viral persistence in vivo. There are three regulatory proteins in PTLV-1: Tax, Rex, and HBZ. PTLV-1 purportedly encodes four accessory proteins, named P12/P8 (encoded by ORF I), P13 and P30 (encoded by ORF II).
This study reports the first STLV-1 genome from a virus infecting a bonnet macaque (Macaca radiata), a macaque species from India. The virus replicated well in cell culture, and this material was used for sequencing, a full genome analysis of this divergent Asian STLV-1 strain was performed. While the canonical ORFs, as well as the ORFs encoding regulatory proteins were found conserved, the ORFs encoding the accessory proteins were absent or disrupted. This latter observation is intriguing as these proteins are purportedly essential for viral persistence in vivo. In order to determine if the absence of accessory proteins is specific to this particular strain, a comprehensive analysis of the complete genomes of different PTLV-1 subtypes available in GenBank was performed. Except for HTLV-1a strains, strains from other PTLV-1 subtypes all lacked at least one accessory ORF. This raises many questions regarding the actual role and importance of accessory proteins in the PTLV-1 biology.
The bonnet macaque Mra18C9 was housed in an animal rescue center and material was sent to us for routine diagnostic screening.
Serological tests for PTLV-1 gave conflicting results. While the serum tested negative on an in-house ELISA, it reacted positively when tested by a local hospital laboratory. These conflicting results led us to perform additional serological testing using the INNO-LIA HTLVI/II assay (Fujirebio Europe, Ghent, Belgium). The serum reacted strongly with PTLV-1/2 env gp46 and gp21 antigens, but did not react with any type-specific peptide. It was thus considered as an indeterminate sero-reactivity.
As the serological tests were inconclusive, we performed a diagnostic PCR screening on the DNA isolated from PBMCs, by using a generic nested PCR assay able to amplify a fragment of the tax/rex region of both PTLV-I and -II. The PCR tested positive, and BLAST analysis on the 118 bp-long fragment revealed high nucleotide identity percentage (>93%) with STLV-1 identified from Formosan macaques (Macaca cyclopis STLV108, GenBank accession number: KM268809), stump-tailed macaques (M. arctoides MarB43 and marc1, GenBank AY590142 and U76625 respectively), and long-tailed macaques (M. fascicularis MFA-C194, GenBank U59132).
The full-length STLV-1 genome infecting the bonnet macaque Mra18C9 was obtained by using a combination of the sequence-independent VIDISCA-454 technique, which generated 4 segments scattered along the STLV genome, and standard PCRs to bridge the sequence gaps and sequence the LTR.
The flanking long terminal repeats (LTRs) comprise a TATA box, a polyadenylation signal, and three Tax-responsive elements (TxRE) type 1. The TxRE-2 may not be fully functional as an insertion of two nucleotides is present at position 210–211. Interestingly, the LTR was quite divergent from previously published PTLV-LTR sequences, as the nucleotide identity was lower than 80% (Table A in S1 Text).
The canonical retroviral ORFs (gag, pro, pol, env) are conserved, as well as the sequences necessary for the gag-pro and pro-pol ribosomal frame-shifts (Fig 1). Nucleotide and amino acid sequence comparisons confirm that Mra18C9 STLV-1 is a highly divergent PTLV-1 strain. At the nucleotide level, the gag, pro, pol, and env genes show not more than 80% identity with their counterparts from other PTLV, while their encoded proteins differ 10–15% in amino acid identity (Tables A-B in S1 Text).
The ORFs corresponding to the regulatory proteins Tax, Rex and HBZ are also present (Fig 1), and splice-acceptor and -donor sites are preserved. The tax gene sequence shares roughly 82% nucleotide identity, and the encoded Tax protein has a 88–90% amino-acid identity to other PTLV-1 sequences (Tables A-B in S1 Text). Importantly, the PDZ-binding motif of Tax, which is necessary for the interaction of Tax with cellular factors, is mutated in Mra18C9 genome: the ETEV motif is changed into an ETEI motif.
Phylogenetic analyses of the concatenated gag-pol-env-tax genes clearly demonstrated that Mra18C9 falls into the Asian STLV-1 group (Fig 2A). Macaque STLV-1 strains all form long branches, and do not aggregate in a monophyletic group. Instead, they form a paraphyletic group. In order to better estimate the relative position of this strain, phylogenetic analyses were performed using the commonly studied LTR and env sequences (Fig 2B and 2C). The different phylogenetic methods (Neighbor-Joining, and a Bayesian approach) gave very similar results. Results of the Bayesian analyses are shown in Fig 2. The Mra18C9 strain was consistently found in a long branch among the macaque strains; the closest known sequences were the Macaca arctoides strains (when considering either LTR or env sequences) (Fig 2B and 2C) and the Macaca mulata strains MMU-R18 and R22 (when considering the LTR) (Fig 2B).
In conclusion, the Mra18C9 strain has a typical PTLV-1 organization, but this virus is highly divergent from other PTLV, suggesting a long and independent evolution.
Sequence analysis suggests that the accessory proteins encoded by ORF-I may not be functional in the Mra18C9 strain. Indeed, the start codon of the open-reading frame is mutated (ATG > 6821GCG), and multiple stop codons are found. Furthermore, the splice-acceptor site is mutated (ATK 6380CAG/CAAC > Mra18C9 6719TAA/CAAC) and may not be functional either. Thus, Mra18C9 does not encode the P12/P8 proteins.
The splice-acceptor domain necessary for a P30 protein-encoding mRNA is also mutated (ATK 6475TAG/CACT > Mra18C9 6814GGA/CGCT), suggesting that P30 may not be produced by the Mra18C9 strain. This is reinforced by the presence of an early stop codon in the ORF.
Similarly, the splice-acceptor domain, necessary for a P13-encoding mRNA, is also mutated (ATK 6872CAG/CAGG > MRA18C9 7223CAG/TTGG), and therefore the mRNA may not be synthesized. In addition, the ORF is also altered: the start codon is conserved but the stop codon is mutated (TAA > 7549CAA). The resulting protein would then be much longer (137 amino-acid long putative protein instead of 87 aa-long), which can rigorously influence its functionality.
Collectively, our analyses indicate that the STLV-1 Mra18C9 strain completely lacks the accessory genes, which have been reported in HTLV-1a strains.
We wondered if the absence of accessory genes was specific to the Asian macaque STLV-1 strain isolated from M. radiata. For this purpose, an in silico analysis of PTLV-1 complete genomes available in GenBank was performed (Table 1 and Tables C-F in S1 Text).
Accessory proteins were conserved in HTLV-1a strains and in the available STLV-1b strain [12]. In contrast, the macaque STLV-1 strains lacked all of the accessory genes. The loss of accessory genes in MarB43 was previously reported [7]. All the other strains (either HTLV-1b, c; or STLV-1e, f, g) lacked at least 1 accessory protein. For instance, the HTLV-1b strains lack both P12 and P30, and as previously mentioned HTLV-1c strains lack P12 [18].
In conclusion, although accessory genes have been found important for viral infection and persistence in HTLV-1a, many PTLV-1 strains lack one or more accessory genes.
PTLV-1 infects a wide range of non-human primates (NHPs). We report the first strain infecting a bonnet macaque (Macaca radiata). Complete genome analysis revealed that it is a highly divergent strain when compared to the currently known PTLVs. This would explain the conflicting results obtained by ELISA, as well as its low reactivity (indeterminate profile) on the commercial Western blot. Phylogenetic analyses positioned the Mra18C9 strain among other Asian macaque STLV-1 viruses. This confirms the large heterogeneity within the Asian PTLV-1 clade as was described previously [5, 6].
In phylogenetic trees, Asian macaque STLV-1 strains form very long branches when compared to African (including HTLV-1a) strains. This could suggest that these viruses have evolved independently in Asia, with their simian host, for a very long period [19]. The introduction of STLV in Asian NHP has been dated at approximately 200.000 years ago [6, 7]. Under this assumption, the viruses have coevolved for long with their host, and the phylogeny of Asian STLV-1 should mirror the evolution of Asian primates. However, this is not the case for macaque STLV-1. First, they do not form a monophyletic group; instead they form a paraphyletic group. The strains are organized as a ladder, branching deeply next to the PTLV-1 root (Fig 2A). One could argue that this particular topology results partially from genetic saturation. Moreover, among Asian STLV-1 there are only a few monophyletic groups corresponding to simian species (Fig 2B and 2C). Apart from M.fuscata and M.arctoides STLV-1, the other clades are composed of sequences of mixed origins, with STLVs from Pongides and Hylobatides that infect macaques. Even when focusing on sequences isolated from macaques, the distribution of the sequences does not follow the known Macaca phylogeny [20, 21]. Together, this points to interspecies transmission (between macaques or from macaques to orangutans or gibbons) of such viruses, as others previously suggested [5, 6]. Interspecies transmission of STLV-1 has been previously reported, although in the context of captivity [22, 23]; the evidence of such transmission in natura is mostly inferred by phylogenetic analysis [6, 24]. Thus, one could argue that the long branching is not only due to a long independent evolution in Asia, but also to an accelerated mutation rate for these strains, which could be related to frequent interspecies transmission.
The canonical and the regulatory proteins are present in the genome of the Mra18C9 STLV-1. The Mra18C9 STLV-1 strain is a functional, replicative virus (at least in vitro, as it could be amplified on SupT1 cells). However, the strain may have an attenuated phenotype. First, the Tax-responsive element 2 (TxRE2) seems disrupted due to a 2-nucleotide insertion [25]. This may lead to a lower basal transcription and reduced viral expression [26]. Second, the viral transactivator Tax has a mutated PDZ-binding motif (PBM). The Tax PBM is essential for sustained proliferation both in vitro and in vivo [27, 28]. A single mutation of the last amino acid of the PBM was shown to be sufficient to abrogate its function [29–31].
The pX region of the Mra18C9 STLV-1 strain lacks both ORF-I and ORF-II. We first hypothesized that the loss of these ORFs could render the virus less pathogenic. Indeed, accessory proteins have been shown to be important for viral persistence and pathogenesis in HTLV-1a. Mutations of accessory ORFs limit the replicative capacity of HTLV-1a in a rabbit model [32]. Similarly, in the closely related bovine leukemia virus, mutations of the homologue ORFs render the virus attenuated [33]. It was proposed that Australian HTLV-1c strains, because they lack the P12 protein, might be less oncogenic [18]. Nevertheless, ATL cases have been reported in HTLV-1c-infected individuals [9]. Moreover, while the absence of accessory proteins seems to be a general feature of macaque STLV-1, ATL cases have been reported in naturally infected macaques [11]. Thus, even in the absence of accessory proteins, STLV-1 still present an oncogenic potential.
Although STLV-1 is highly prevalent among Asian NHPs, and humans are in frequent contact with macaques and may acquire other retroviral infection, such as simian foamy viruses that are endemic in several species of Asians monkeys [34], no Asian STLV zoonotic transmission has been reported so far. One could hypothesize that the absence of ORFI and ORFII in macaque STLV-1 can limit viral transmission and propagation. Indeed, HTLV-1a P12 and P30 have been found to be essential for viral replication in human and macaque dendritic cells [35], which can play a key role in viral transmission. However, this proposition is nullified by our thorough analysis of the different PTLV-1 subtypes. Indeed, HTLV-1b, which is a very common virus in Central Africa, persists and propagates in humans despite the absence of P12 and P30 (Table 1).
In conclusion, the role and importance of accessory proteins needs to be reconsidered in light of the analysis of the different PTLV-1 strains. Indeed, as most studies have focused on HTLV-1a, the 3 accessory proteins were found conserved and their function in viral persistence and transmission was believed to be essential [17, 32, 35]. This study indicates that some PTLV-1 can persist in the absence of one or many accessory proteins. This raises the question of a dispensable role of these proteins, or the presence of other accessory proteins yet to be identified in the other PTLV-1 subtypes.
The Bonnet macaque was housed in an animal rescue center (animal shelter VZW, Belgium) and material was sent to the BPRC for viral diagnostic screening.
The BPRC is fully licensed by the Netherlands Food and Consumer Product Safety Authority (belonging to the Ministry of Agriculture, Nature and Food Quality) to work with animal products and perform diagnostic services for third parties (Approval no. 1926950).
Sera were assayed for antibodies to PTLV-1 using an in-house developed serological test with purified, lysed HTLV-1 particles as coating antigen (Advanced Biotechnologies Inc., Eldersburg, USA) [36]. Additionally, the serum from the bonnet macaque was also tested by ELISA in a local hospital laboratory (DDL, Delft, The Netherlands), and a western blot (WB) analysis using the INNO-LIA HTLVI/II assay (Fujirebio Europe, Gent, Belgium).
Genomic DNA was isolated from whole blood using the QIAamp DNA blood mini kit (QIAGEN Benelux B.V., Venlo, The Netherlands). A 118 bp tax/rex gene fragment was amplified using the TR101/TR102 and SK43/SK44 nested primer sets, as previously described [37, 38]. Furthermore, A pan-STLV PCR was performed with primers PH1F and PH2R, as described by van Dooren et al. [39]. The amplified fragment of 192 bp fully overlapped the 118 bp fragment from the first PCR.
The virus discovery method VIDISCA-454 was used for the analysis of cell culture supernatant from the PTLV-infected SupT1 cell culture, as previously described [40]. In brief, PBMCs were isolated on a Ficoll and stimulated for 2 days with PHA (1 ug/ml final). Next, they were co-cultured with SupT1 cells until CPE was visible (2–3 weeks). The cell culture supernatant was centrifuged to remove cell debris and treated with TURBO DNase (Ambion, Thermo Fisher Scientific, Breda, The Netherlands). Next, nucleic acids were isolated with a QIAamp Viral RNA Mini Kit (QIAGEN Benelux BV, Venlo, the Netherlands) and reverse-transcribed with SuperScript II (Thermo Fisher Scientific) using non-ribosomal random hexamers. Subsequently, second strand DNA synthesis was performed with 5 U of Klenow fragment (New England Biolabs, Bioke, Leiden, The Netherlands). Double-stranded DNA was purified by phenol/chloroform extraction and ethanol precipitation and digested with Mse I restriction enzyme (New England Biolabs). Adaptors with different Multiplex Identifier sequences (MIDs) were ligated to the digested fragments of the different samples. Before PCR amplification, the fragments were purified with AMPure XP beads (Agencourt AMPure XP PCR, Beckman Coulter, Woerden, The Netherlands).
A 28-cycles PCR with adaptor-annealing primers was performed. The program of the PCR-reaction was: 5 min 95°C, and cycles of 1 min 95°C, 1 min 55°C, and 2 min 72°C, followed by 10 min 72°C and 10 min 4°C. After purification with AMPure XP beads, the purified DNA was quantified with the Quant-it dsDNA HS Qubit kit (Invitrogen, Carlsbad, CA, USA) and diluted to 107 copies/μl. Samples were pooled and Kapa PCR (Kapa Biosystems, Wilmington, MA, USA) was performed to determine the quantity of amplifiable DNA in each pool. Subsequently, the Bioanalyser (hsDNA chip, Agencourt) was used to determine the average nucleotide length of the libraries. The pools were diluted until 106 copies/μl, titrated with beads (DNA:beads ratio of 0.5:1, 1:1, 2:1 and 4:1) and used in an emulsion PCR according to the supplier’s protocol (LIB-A SV emPCR kit). Sequencing was done on a 2 region GS FLX Titanium PicoTiterPlate (70x75) with the GS FLX Titanium XLR 70 Sequencing kit (Roche, Woerden, The Netherlands). Sequence reads were analyzed using the blastn and blastp algorithms (National Center for Biotechnology Information).
PCR primers were designed on basis of the four fragments that were obtained from the VIDISCA as well as from the consensus LTR sequence derived from the alignment of other PTLV-1 (Tables G-H in S1 Text). PCR reactions were performed in a final volume of 50 μl. Each amplification reaction was performed in 1x DreamTaq buffer containing 200 μM of each dNTP, 50 pmol of each primer, and 1.25 U DreamTaq DNA polymerase (Thermo Fisher Scientific). The amplification reactions were performed for 35 cycles consisting of a 30 s denaturation step at 94°C, a 30 s annealing step at 55°C and an elongation step of 150 sec at 72°C. Amplicon purification and sequencing was performed essentially as described above, but by using a primer-walking sequencing strategy. Sequences were assembled with the SeqMan Pro software (DNASTAR, Inc.,Madison, USA). The resulting contig was further analyzed using the MacVector software package (MacVector, Inc., Cambridge, UK). The complete genome of STLV-1 Mra18C9 has been deposited at GenBank with accession number MK639100.
LTR (701 bases) and env (483 bases) sequences were aligned together with most sequences available on GenBank using DAMBE [41]. An alignment of concatenated gag-pol-env-tax sequences (5820 bases) was also generated with the sequences of complete PTLV-1 genomes available on GenBank.
Phylogenetic trees resulted from analyses using the neighbor-joining method performed with the PAUP* v4.0b10. The final alignment was submitted to the Modeltest program (version 3.6) to select, according to the Akaike information criterion, the best model to apply to phylogenetic analyses. The selected substitution models were: Tamura-Nei for env and concatenated gag-pol-env-tax, and general time-reversible (GTR + γ) for the LTR. To test the robustness of the tree topologies, 1,000 bootstrap replicates were performed. Bayesian approaches were inferred with the MrBayes 3.2.7 program and robustness was tested with posterior probabilities. Both methods raised similar phylogenetic tree topology.
The analysis was performed following the alignment of complete genomes available in GenBank (Table C in S1 Text). Splicing acceptor and donor sequences were previously described for ATK [13]. Frameshift sites had been previously identified [42].
|
10.1371/journal.pntd.0002044 | Molecular and Functional Characterization of a Trypanosoma cruzi Nuclear Adenylate Kinase Isoform | Trypanosoma cruzi, the etiological agent of Chagas' disease, is an early divergent eukaryote in which control of gene expression relies mainly in post-transcriptional mechanisms. Transcription levels are globally up and down regulated during the transition between proliferating and non-proliferating life-cycle stages. In this work we characterized a nuclear adenylate kinase isoform (TcADKn) that is involved in ribosome biogenesis. Nuclear adenylate kinases have been recently described in a few organisms, being all related to RNA metabolism. Depending on active transcription and translation, TcADKn localizes in the nucleolus or the cytoplasm. A non-canonical nuclear localization signal was mapped towards the N-terminal of the protein, being the phosphate-binding loop essential for its localization. In addition, TcADKn nuclear exportation depends on the nuclear exportation adapter CRM1. TcADKn nuclear shuttling is governed by nutrient availability, oxidative stress and by the equivalent in T. cruzi of the mammalian TOR (Target of Rapamycin) pathway. One of the biological functions of TcADKn is ribosomal 18S RNA processing by direct interaction with ribosomal protein TcRps14. Finally, TcADKn expression is regulated by its 3′ UTR mRNA. Depending on extracellular conditions, cells modulate protein translation rates regulating ribosome biogenesis and nuclear adenylate kinases are probably key components in these processes.
| Infection with Trypanosoma cruzi produces a condition known as Chagas disease which affects at least 17 million people. Adenylate kinases, so called myokinases, are involved in a wide variety of processes, mainly related to their role in nucleotide interconversion and energy management. Recently, nuclear isoforms have been described in several organisms. This “atypical” isoform in terms of primary structure was associated to ribosomes biogenesis in yeast and to Cajal body organization in humans. Moreover nuclear adenylate kinases are essential for maintaining cellular homeostasis. In this manuscript we characterized T. cruzi nuclear adenylate kinase (TcADKn). TcADKn localizes in the nucleolus or cell cytoplasm. Nuclear shuttling mechanisms were also studied for the first time, being dependent on nutrient availability, oxidative stress and by the equivalent of the mammalian TOR pathway in T. cruzi. Furthermore we characterized the signals involved in nuclear importation and exportation processes. In addition, TcADKn expression levels are regulated at an mRNA level, being its 3′UTR involved in this process. These findings are the first steps in the understanding of ribosome processing in trypanosomatids.
| Trypanosoma cruzi, the causative agent of Chagas' disease, is a protozoan parasite with a complex life cycle which involves two intermediary hosts, triatomine insects and mammals and three main parasite stages, epimastigotes and amastigotes which replicate in the insect vector and mammalian host respectively; and trypomastigotes the non-replicative form [1]. The complexity of its life cycle involves multiple morphological and metabolic changes that are possible due to a strict control of gene expression [2]. Early eukaryotes from the order Kinetoplastida, transcribe their genes as large polycistronic arrays and therefore rely on post-transcriptional mechanisms for gene expression regulation [3], [4], [5], [6], [7], [8], [9], [10].
Furthermore trypanosomatids present unique characteristics regarding ribosome structure [11], [12] and ribosomal locus organization [13]. Instead of having the typical ribosomal locus organization which consists of ribosomal promoter, ETS1 (external transcribed spacer 1), 18S rDNA, ITS1 (internal transcribed spacer 1), 5,8S rDNA, ITS2 (internal transcribed spacer 2), 28S rDNA, ETS2 (external transcribed spacer 2), ribosomal terminator and the 5S rDNA, they present the 28S rDNA fragmented in 7 small rDNAs [13]. There is almost no information about ribosome biogenesis in trypanosomatids, but their extremely divergent ribosomal locus suggests that they might present unique characteristics in ribosome biogenesis and assembly. For example in T. brucei ribosomal 5S rRNA biogenesis involves proteins which are exclusively found in trypanosomatids [14], [15], [16].
In the last few years, an atypical nuclear adenylate kinase (ADK, ATP∶AMP phosphotransferase, EC: 2.7.4.3) isoform has been characterized in several organisms, such as Drosophila melanogaster [17], Saccharomyces cereviciae (FAP7) [18], Caenorhabditis elegans [19] and Homo sapiens (hCINAP) [20], [21]. ADKs are mainly involved in maintaining the adenine nucleotide pool, which includes ATP synthesis from ADP [22]. They are distributed in all kind of organisms, from bacteria to vertebrates, presenting conserved motifs, structures and functions. However, nuclear ADKs present unique characteristics and differ enormously in terms of primary structure and function from other previously characterized ADKs. It has been shown that all nuclear ADKs, present phosphotransferase activity in vitro, furthermore the human and yeast variants also present ATPase activity [17], [19], [23], [24]. In S. cereviciae, FAP7 has shown several diverse functions; first of all it has been related to oxidative stress response by the activation of the transcription factor POS9 [18], secondly overexpression of FAP7 confers, resistance to arsenite exposure, a powerful oxidant [25], [26]. Finally FAP7 has been related to ribosome biogenesis; being involved in the final step of maturation of the 20S pre-rRNA, which corresponds to the cleavage at “site D” by direct interaction with Rps14, a ribosomal protein that is found near the 3′ end of the 18S rRNA [23], [25]. Interestingly, conserved residues predicted to be required for nucleoside triphosphate (NTP) hydrolysis are essential for FAP7 function in vivo [18], [23]. Furthermore the human isoform (hCINAP) has also been vastly characterized and is involved in Cajal body organization [27]; transcription process and cell cycle progression [28].
In trypanosomatids ADKs have been identified in Leishmania [29], Trypanosoma [30], [31] and Phytomonas [32]. In T. brucei [31] and T. cruzi [30] several isoforms have been characterized with different subcellular localization including, flagella, glycosomes, mitochondria, and cytoplasm [30], [31], [33], [34], mainly related to energy balance maintenance.
In the following work we characterized T. cruzi nuclear ADK isoform, showing that it is involved in ribosome processing and presents unique characteristics being completely different from the other isoforms found in these parasites.
Stock cultures of T. cruzi epimastigotes of the Y strain were maintained in axenic conditions at 28°C in BHT (Brain Heart Triptose) media (pH 7) supplemented with 10% fetal calf serum, 100 U.mL−1 penicillin, and 100 mg.L−1 streptomycin [35]. Transfected parasites were maintained in the same media containing 500 µg.mL−1 of G418 and 10% fetal calf serum. Parasites were counted in a Neubauer hemocytometer chamber.
T. cruzi TcADKn gene (Systematic ID: Tc00.1047053507023.280) was amplified from genomic DNA of epimastigotes from the Y strain and cloned in the pRSET-A vector (Invitrogen) by digesting with HindIII/XhoI and TcRps14 (Systematic ID: Tc00.1047053506945.230) was amplified from genomic DNA of epimastigotes from the Y strain and cloned in the pGEX vector (GE Healthcare) digested with BamHI/XhoI. The sequence coding for full-length T. cruzi TcADKn was cloned in the pTEX-eGFP expression vector by digesting with HindIII/SalI. The pTEX-eGFP plasmid was constructed by cloning the eGFP into the pTEX-TAP vector, kindly provided by Dr. Esteban Serra (IBR, Rosario). A total of 108 parasites of the Y strain, were grown in BHT medium at 28°C, harvested by centrifugation, washed with PBS, and resuspended in 0.35 mL of electroporation buffer (PBS containing 0.5 mM MgCl2, 0.1 mM CaCl2). The cell suspension was mixed with 50 µg of plasmid DNA in 0.2 cm gap cuvettes (Bio-Rad). Parasites were electroporated with a single pulse of 400 V, 500 µF with a time constant of about 5 ms. Stable cell lines were achieved after 30 days of treatment with 500 µg.mL−1 G418 (Sigma) [36]. For deletion analyses TcADKn segments were amplified by PCR from the pTEX-ADKn-eGFP plasmid, cloned into pGEM T-easy vector (Promega) and subcloned in the pTEX-OMNI-eGFP vector. The pTEX-OMNI vector derives from the pTEX-GFP vector [37], by the addition of the 3-FLAG (Sigma), HA (influenza virus hemagglutinin), and aT (C-terminal alpha tubulin) epitopes present in the pDIY cloning vector (GI:374430409). TcADKn locus was amplified from genomic DNA, the 3′ UTR was cloned in the pTREX-OMNI-eGFP vector which contains the same epitopes as the pTEX-OMNI vector but in a pTREX backbone. The pTEX-Dhh1-eGFP plasmid was kindly provided by Dr. Alejandro Cassola (IIB-UNSAM). For oligonucleotide sequence refer to Table S2.
Freshly grown trypanosome samples were washed twice in PBS. After letting the cells settle for 30 min at room temperature in poly-L-lysine coated coverslips, parasites were fixed at room temperature for 20 min with 2% formaldehyde in PBS, followed by a cold methanol treatment for 5 min. Afterwards, all the samples were treated with anti-TcADKn (1∶200), anti-GFP antibody (1∶500) (Invitrogen) or anti-TcPABP1 (1∶500) for 1 h followed by 1 h incubation with anti-mouse (daylight 488, Jackson Immuno Research) or anti-rabbit (daylight 594, Jackson Immuno Research) secondary antibody. Slides were mounted using Vectashield with DAPI (Vector Laboratories). Cells were observed in an Olympus BX51 fluorescence microscope. Images were recorded with an Olympus XM10 camera. Images were analyzed with MBF ImageJ for microscopy bundle.
Fusion proteins were expressed in Escherichia coli BL21 (DE3) or DH5α. Cells were grown in Luria broth medium (LB) at 37°C with ampicillin to an optical density of 0.4 to 0.5 measured at 600 nm (OD600). Protein expression was induced with 1 mM of isopropyl-β-D-thiogalactoside (IPTG) for 16 to 20 h at 37°C. Cells were harvested by centrifugation, and pellets were resuspended in 5 to 8 volumes of breaking buffer (350 mM NaCl, 50 mM Tris-HCl, pH 7.4, 0.5 mM EDTA) containing a protease inhibitor PMSF (10 µg/ml) and disrupted by sonication. Extracts were clarified by centrifugation at 12,000 rpm for 20 minutes. To purify recombinant 6x-His-TcADKn protein, clarified extracts were incubated with Ni-nitrilotriacetic acid beads (1 mL beads for 5 g of cell pellet; QIAGEN) for 12 h at 4°C. Proteins were eluted with 5 bead volumes of breaking buffer containing 200 mM imidazole.
Eluates containing nearly homogenous recombinant protein were pooled and dialyzed overnight in breaking buffer containing 20% glycerol and stored at −80°C. This procedure yielded 90%-pure recombinant protein, as judged by SDS-polyacrylamide gel electrophoresis (SDS-PAGE) and Coomassie brilliant blue staining. To purify glutathione S-transferase (GST) fusion proteins, GST-TcRps14 and the GST epitope, clarified extracts were incubated with 2 ml of glutathione-Sepharose beads (Amersham) for 4 h at 4°C. The beads were washed extensively with breaking buffer, and proteins were eluted with the same buffer containing 20 mM of reduced glutathione (Amersham). Eluates were dialyzed overnight in breaking buffer containing 20% glycerol and stored at −80°C.
For ADK activity, 50 µg of purified recombinant protein fraction were added to the reaction mixture (100 mM Tris-HCl buffer pH 7.5, 20 mM glucose, 5 mM MgCl2, 100 mM KCl, 2 mM dithiothreitol, 1 mM NADP+, 5 U.mL−1 and 2 U.mL−1 glucose-6-phosphate dehydrogenase) in a cuvette in a final volume of 0.5 mL. After 5 min incubation at 35°C the reaction was started by the addition of a small volume of ADP to a final concentration of 10 mM. ADK activity was calculated by measuring the increase in absorbance at 340 nm that accompanied the reduction of NADP+ [30]. For ATPase activity a sample of 50 µg of protein was added to the reaction mixture (100 mM Tris-HCl, pH 7.5, 60 mM KCl, 5 mM MgCl2, 5 U.mL−1 of polynucleotide kinase, 5 U.mL−1 of lactate dehydrogenase, 20 mM phosphoenolpyruvate, 1 mM NADH). After 5 min incubation at 35°C the reaction was started by the addition of a small volume of ATP. ATPase activity was calculated by measuring the decrease in absorbance at 430 nm that accompanied the oxidation of NADH. The results were plotted and the slope was used to calculate specific activities.
1×108 epimastigotes, were harvested by centrifugation, washed with PBS, resuspended in 1 mL of Tri-Reagent (Sigma), mixed by inversion and 200 µL of chloroform were added followed by centrifugation at 12,000×g at 4°C. The supernatant was transferred to a clean test tube with 500 µL of isopropanol, after 10 min incubation at room temperature; they were centrifuged at 12,000×g for 15 min at 4°C. The pellet was washed with ethanol 75%, left to dry and resuspended in 20 µL of RNase-free water. RNA concentrations were determined spectrophotometrically, purity was confirmed by gel electrophoresis. 3 µg of RNA were used for retrotranscription, which were previously treated with DNaseI (Sigma) in order to eliminate any DNA contamination. TcADKn mRNAs were isolated by RT-PCR cloned in pGEM T-easy vector (Promega) and sequenced. TcADKn differential mRNA expression along the epimastigote's growth curve was quantified by SYBR green-based real-time PCR in a Real-Time PCR system (Bio-Rad) using default protocols. Data were relativized to 18S expression. Three independent experiments were carried out. eGFP expression along the growth curve of parasites expressing the pTREX-OMNI-eGFP-LAN or pTREX-OMNI-eGFP constructions was quantified by SYBR green-based real-time PCR in a Real-Time PCR system (Bio-Rad) using default protocols. GFP expression was relativized to neomycin (Neo) expression (present in the pTREX-OMNI-eGFP vector). Three independent experiments were carried out.
1.25×108 epimastigotes from day 2 of culture were harvested washed with PBS, resuspended in buffer A (20 mM Tris-HCl, pH 7.6, 2 mM MgCl2, glycerol 10%, Nonidet P-40 0.5%, 1 mM EDTA, 1 mM DTT, 0.25 M sacarose, 50 mM KCl, 1 mM E64 and RNAse inhibitor from Sigma) and incubated for 30 min in ice. They were harvested at 10,000×g for 15 min at 4°C, the supernatant was transferred to a clean tube containing 20 mg of protein G-agarose (Sigma) and 10 µl of preimmune serum the mixture was left in agitation for 1 h at 4°C. This fraction corresponded to the clarified extract. In parallel 20 mg of protein G-agarose were blocked with 100 µg of BSA in buffer A for 2 h at room temperature. After the pre-blocking 10 µl of anti-TcADKn serum were added and incubated for 2 h at 4°C, washed three times with 500 µL of buffer A and 400 µL of clarified epimastigotes extract was added and left in agitation for 1 h at 4°C. After incubation beads were washed three times with buffer A and five times with PBS 1×. The pellet was resuspended in 800 µL of TriReagent (Sigma) for RNA extraction. A small fraction was separated for protein analysis.
10 µg of recombinant 6x-His-tagged T. cruzi arginine kinase (TcAK, Tc00.1047053507241.30) and 6x-His-tagged TcADKn were incubated with 10 µg of GST or 5 µg of GST-TcRps14, in buffer K (150 mM NaCl, 50 mM Tris-HCl pH 7.4, 0.5 mM EDTA) containing protease inhibitor PMSF (10 µg.mL−1) in a final volume of 30 µL. After 1 h of incubation in ice, 190 µL of buffer K were added to the mixtures, 20 µL were removed for analysis (10% of the input), and the remainder was incubated with 10 µL of glutathione-Sepharose beads (Amersham) for 1 h in ice with regular agitation. Beads were spinned down by centrifugation and 20 µL of the supernatant were subsequently removed for analysis. The beads were washed three times with 1 mL of ice-cold buffer K. Bound proteins were extracted by boiling the beads in SDS-PAGE loading buffer (output) and resolved in 12% polyacrylamide denaturing gels. Proteins were identified by Western Blot analysis.
Western Blots were performed using total T. cruzi extracts fractioned by electrophoresis in polyacrylamide denaturing gels and transferred to polyvinylidene fluoride (PVDF) membranes. The PVDF membranes were treated for 1 h with 5% non-fat dry milk in PBS and then incubated with the primary antibody ON, using anti-TcADKn diluted 1∶5000, anti-His 1∶3000 (Sigma) or anti-GST diluted 1∶2000 (Invitrogen), anti-GFP diluted 1∶2500 and anti-α-tubulin diluted 1∶2000 (Abcam). Membranes were washed and incubated with the corresponding secondary antibody for two hours (anti-mouse HRP 1∶2500, anti-rabbit HRP 1∶2500, Vector Labs). Detection was done by chemiluminescence (Pierce).
The 3HA-FAP 7 strain (mat α ura3-52 lys2-80 ade2-101 trp163 his3-200 leu2-1) was kindly provided by Dr. Baserga. Strains were grown in YPG (1% yeast extract, 2% peptone, 2% galactose) until transformation. Yeasts were transformed as explained in http://home.cc.umanitoba.ca/~gietz/. The genes of TcADKn, TbADKn, FAP7, E. coli ADK, TcADK6, TbADKF were amplified and cloned in the p416 vector [38], kindly provided by Dr. Cecilia D'Alessio, Fundacion Instituto Leloir. Transformed yeast were grown in minimum medium with galactose for selection and afterwards shifted to minimum medium with glucose for complementation assays [23].
Exponentially growing T. cruzi epimastigotes were treated with different drugs: actinomycin D (Sigma) 10 µg.mL−1 for 4 h, cicloheximide (Sigma) 50 µg.mL−1 for 4 h, puromycin 200 µg.mL−1 4 h, starvation in PBS 24 h, leptomycin B (Sigma) 0.1 µg.mL−1 for 5 h, rapamycin (Sigma) 100 µM for 6–8 h, phleomycin 150 µg.mL−1 for 4 h, hydrogen peroxide 200 µM for 1 h. After leptomycin and rapamycin treatment fluorescence was quantified for forty treated and untreated parasites. In each parasite the fluorescence from the green (GFP) channel was quantified in an area selected according to blue signal (DAPI) fluorescence (nucleus) using the RGB plugin in ImageJ (http://rsb.info.nih.gov/ij/). Cytoplasmic fluorescence was quantified in the same way selecting the brightest perinuclear areas in the green (GFP) channel. Selection criterion was the same for all transfected parasites. Cytoplasmic fluorescence determinations included the previously selected nuclear areas and fluorescence values, which were afterward subtracted, resulting in the area and values corresponding to the cytoplasm. For each parasite, the relationship between fluorescence/area was obtained for the nucleus and cytoplasm and then the ratio between the nuclear and cytoplasmic values was calculated. For media supplementation experiments BHT medium of parasites in day 10 of culture was supplemented with glucose or proline 2% and were incubated for 24 h. Results were monitored by immunofluorescence. For RNAseI treatment, epimastigotes in day 2 of culture, were harvested, washed with PBS, resuspended in buffer (50 mM Tris-HCl pH 7.8) and broken with liquid nitrogen, samples were treated for 2 h at 37°C with 20 mg.mL−1 of RNaseI. Samples were boiled in SDS-PAGE loading dye and analyzed by Western Blot. For native gel analysis, loading buffer (Tris-HCl 100 mM pH 8, sucrose 2%, BPB 0,05%) was added to protein samples and analyzed by Western Blot.
18S rRNA precursors were isolated using a simple adapter ligation protocol. A standard oligonucleotide blocked in its 5′ end was adenylated using the New England Biolabs (NEB) adenylation kit following manufacturer's indications. Epimastigotes RNA was extracted as explained above, for each ligation reaction three RNA samples were pooled. The 3′ adenylated adapter was ligated in the absence of ATP using the T4 RNA ligase 2 truncated (NEB) using manufacturer's indications. The final ligation products were reversely transcribed into cDNA using a complementary oligonucleotide to the adapter. PCRs were performed using specific primers for the 18S and ITS region. PCR products were cloned in the pGEM T-easy vector (Promega) and submitted for sequencing. The same strategy was used for RNA extracted from immunoprecipitates against TcADKn. RT PCR were done using oligonucleotides, for the ITS region, TcH2B (Systematic ID: Tc00.1047053511635.20) and TcNDPK3 (Systematic ID: Tc00.1047053510879.210). A standard PCR protocol was used, 5 minute denaturation at 95°C and 30 cycles: 1 minute denaturation at 95°C, 1 minute at the corresponding annealing temperature, and 1 minute at 72°C; finally 10 minutes at 72°C. Controls without retrotranscriptase were done to eliminate any DNA contamination possibilities.
Sequences were obtained from the TriTrypDB (http://tritrypdb.org/tritrypdb/). Assembly and sequence data analysis, including ORFs prediction, were carried out using the software package Vector NTI 10.3.0 (Invitrogen) and the online version of BLAST at the NCBI (http://blast.ncbi.nlm.nih.gov/Blast.cgi). Sequence analysis and nuclear localization signals and nuclear exportation signal were carried out using the online predictors http://www.psort.org/ [39], http://www.cbs.dtu.dk/services/NetNES/,and http://psort.hgc.jp/, respectively [40].
Adenylate kinases have been mainly related to nucleotide interconversion and energy management. [74]. In 2005 the first nuclear ADK isoform was found [21], later they were characterized in several organisms [17], [18], [19]. This “atypical” isoform in terms of primary structure was associated to ribosomes biogenesis in yeast [23] and to Cajal bodies organization in humans [20], [27], [71]. In these enzymes the P-loop domain, responsible of nucleotide binding in phosphotransferases, could be involved in other functions such as protein interactions, endonucleolytic RNA cleavage, RNA-protein interaction or RNA metabolism [17], [18], [23], [28]. Even though the function of nuclear adenylate kinases is not completely understood it must be extremely important as they are essential for cell viability [17], [19], [23], [28].
In this paper we report the existence of the 7th ADK variant in T. cruzi, which corresponds to the nuclear isoform. We studied its nuclear shuttling and characterized its non-canonical nuclear localization signal, being one of the few atypical NLS that involves the catalytic site of the protein (Walker domain or P- loop). Probably its enzymatic activity is not essential for its nuclear importation as fusion proteins did not yield a higher-than-expected-band in western blotting suggesting that they might be inactive. We postulate that TcADKn enters the nucleus in an unfolded conformation, being the nuclear localization signal within the P-loop, once it enters the nucleus it folds correctly regarding the active site inside the protein. The available data does not allow us to conclude how the importation process takes place; TcADKn could be forming a complex with other proteins, which are recognized by the importin and then enter the nucleus or it could be recognized directly by the importing complex. Further experiments should be carried out in order to understand the nuclear importation mechanism. We could also relate its nuclear exportation to the CRM1 exportin adapter [49], being one of the few proteins in T. cruzi which has been reported to use this transporter.
T. cruzi ribosomes have been studied for a long time because they exhibit unique characteristics which are absent in higher eukaryotes and that could be capitalized for therapeutic drug design [12]. Scientific studies have focused on ribosomal structure rather than in its biogenesis. In trypanosomatids there is almost no evidence about ribosome processing sites or the proteins involved in each step. Proteomic data has revealed that many ribosomal proteins and accessory non-ribosomal proteins are conserved in T. cruzi [11], however their function has not been determined [12]. TcADKn homolog in yeast has been related to ribosome processing, being associated to the final cytoplasmic step of maturation of the 18S rRNA by direct interaction with TcRps14. [23]. By yeast complementation assays we could postulate that TcADKn could be involved in ribosome 18S rRNA processing. This idea is reinforced with the fact that we detected in vitro interaction between TcADKn and TcRps14 and moreover we detected ribosomal precursors in TcADKn immunoprecipitates. These data suggest that ribosome biogenesis in T. cruzi presents conserved characteristics with yeast. However it also presents similar characteristics to mammals. In yeast the ribosomal precursors of the 18S rRNA subunit presents only one intermediary after A2 cleavage within the ITS1, while mammals present two intermediaries [75]. In our experiments we could detect two 18S pre-rRNA precursors indicating that the 18S rRNA processing presents both mammalian and yeast characteristics. So in T. cruzi 18S rRNA biogenesis would be unique as it combines characteristics of both mammals and yeast.
Finally, TcADKn nuclear shuttling is regulated by nutrient availability, ribosome biogenesis, DNA integrity, oxidative stress and probably by the equivalent of the mammalian TOR pathway in T. cruzi. Furthermore the results obtained after puromycin and cicloheximide treatment suggest that ribosome assembly might be necessary for TcADKn nucleolar localization as this one is lost when ribosomes cannot reassemble after cicloheximide treatment. On the contrary in puromycin treatment in which protein synthesis is blocked but ribosomes can re-assembly, nucleolar localization is not disrupted [56], [57], [58], [59]. Similar regulation mechanisms have been observed for other ribosomal proteins [76]. The existence of tight regulation mechanisms gives us an idea of the complexity of ribosome biogenesis and the susceptibility of this process to environmental changes and unfavorable conditions.
Figure 7 summarizes the role of TcADKn in the formation of the 18S ribosomal subunit and its regulation mechanisms. We hope that the information presented encourages the study of ribosome biogenesis in these divergent organisms.
|
10.1371/journal.pgen.1002083 | Meiotic Recombination Intermediates Are Resolved with Minimal Crossover Formation during Return-to-Growth, an Analogue of the Mitotic Cell Cycle | Accurate segregation of homologous chromosomes of different parental origin (homologs) during the first division of meiosis (meiosis I) requires inter-homolog crossovers (COs). These are produced at the end of meiosis I prophase, when recombination intermediates that contain Holliday junctions (joint molecules, JMs) are resolved, predominantly as COs. JM resolution during the mitotic cell cycle is less well understood, mainly due to low levels of inter-homolog JMs. To compare JM resolution during meiosis and the mitotic cell cycle, we used a unique feature of Saccharomyces cerevisiae, return to growth (RTG), where cells undergoing meiosis can be returned to the mitotic cell cycle by a nutritional shift. By performing RTG with ndt80 mutants, which arrest in meiosis I prophase with high levels of interhomolog JMs, we could readily monitor JM resolution during the first cell division of RTG genetically and, for the first time, at the molecular level. In contrast to meiosis, where most JMs resolve as COs, most JMs were resolved during the first 1.5–2 hr after RTG without producing COs. Subsequent resolution of the remaining JMs produced COs, and this CO production required the Mus81/Mms4 structure-selective endonuclease. RTG in sgs1-ΔC795 mutants, which lack the helicase and Holliday junction-binding domains of this BLM homolog, led to a substantial delay in JM resolution; and subsequent JM resolution produced both COs and NCOs. Based on these findings, we suggest that most JMs are resolved during the mitotic cell cycle by dissolution, an Sgs1 helicase-dependent process that produces only NCOs. JMs that escape dissolution are mostly resolved by Mus81/Mms4-dependent cleavage that produces both COs and NCOs in a relatively unbiased manner. Thus, in contrast to meiosis, where JM resolution is heavily biased towards COs, JM resolution during RTG minimizes CO formation, thus maintaining genome integrity and minimizing loss of heterozygosity.
| Cell proliferation involves DNA replication followed by a mitotic division, producing two cells with identical genomes. Diploid organisms, which contain two genome copies per cell, also undergo meiosis, where DNA replication followed by two divisions produces haploid gametes, the equivalent sperm and eggs, with a single copy of the genome. During meiosis, the two copies of each chromosome are brought together and connected by recombination intermediates (joint molecules, JMs) at sites of sequence identity. During meiosis, JMs frequently resolve as crossovers, which exchange flanking sequences, and crossovers are required for accurate chromosome segregation. JMs also form during the mitotic cell cycle, but resolve infrequently as crossovers. To understand how JMs resolve during the mitotic cell cycle, we used a property of budding yeast, return to growth (RTG), in which cells exit meiosis and resume the mitotic cell cycle. By returning to growth cells with high levels of JMs, we determined how JMs resolve in a mitotic cell cycle-like environment. We found that, during RTG, most JMs are taken apart without producing crossovers by Sgs1, a DNA unwinding enzyme. Because Sgs1 is homologous to the mammalian BLM helicase, it is likely that similar mechanisms reduce crossover production in mammals.
| Recombination has a major role during meiosis, as it is necessary for accurate homolog segregation at the first meiotic division [1]. Meiotic recombination is initiated by DNA double strand breaks (DSBs) that are formed by the Spo11 nuclease [2], [3]. Single stranded DNA, produced at break ends by 5′ to 3′ resection [4], then interacts with complementary sequences on the homolog or on the sister chromatid [5], [6]. Some interhomolog recombination events produce a noncrossover (NCO), in which both interacting chromosomes retain parental flanking sequence configurations, whereas other events produce a reciprocal exchange of flanking sequences, or crossover (CO). COs, in combination with sister chromatid cohesion, form the inter-homolog linkage that is required for proper homolog segregation [1]. In Saccharomyces cerevisiae, COs comprise about one half of all interhomolog recombination events [7]. Meiotic COs are produced by the resolution of joint molecule (JM) intermediates [8]–[10], most of which contain two Holliday junctions [11], here called double Holliday junction JMs (dHJ-JMs).
In most organisms, including S. cerevisiae, meiotic DSB formation and recombination are also necessary for progressive colocalization and alignment of homologs during prophase. This process culminates at pachytene, where homologs are joined at sites of recombination and linked tightly along their entire length by a meiosis-specific tripartite protein structure called the synaptonemal complex (SC; [12]).
Although genome-wide programmed DSB formation is central to normal meiosis, it does not usually occur during the mitotic cell cycle. During the budding yeast mitotic cell cycle, most breaks are repaired by recombination between sister chromatids [13]–[15], and the inter-homolog homologous recombination (HR) events that do occur during the mitotic cell cycle produce COs less frequently than in meiosis [13], [16].
The lower yield of COs during mitotic recombination, as compared to meiotic recombination, can be explained in two ways. First, fewer dHJ-JMs are produced per DSB repair event during mitosis than during meiosis [15], and it is possible that most mitotic DSB repair does not involve dHJ-JM formation. Second, it is possible that JMs are produced at significant levels during mitotic HR, but are resolved differently than are JMs produced during meiosis. In S. cerevisiae, most meiotic JMs are resolved as COs [8]–[10] in a process that most likely involves endonuclease cleavage of Holliday junctions, and that is triggered by Cdc5, the budding yeast polo-like kinase homolog [17], [10]. Much less is known about JM resolution during the mitotic cell cycle, since the products of intersister recombination cannot be distinguished from the precursor molecules.
Several structure-selective nucleases have been suggested as having a role in JM resolution by Holliday junction cleavage [18]. The most extensively studied of these is a structure-selective heterodimeric endonuclease, hereafter called the Mus81 complex, that contains the conserved Mus81 nuclease in complex with a second protein, called Mms4 in S. cerevisiae and Drosophila, and Eme1 in fission yeast, mammals and plants [19]–[21]. Meiotic progression defects are evident in S. pombe and S. cerevisiae mutants lacking the Mus81 complex, but the nature of these defects differs in the two organisms. In S. pombe, mutants lacking the Mus81 complex show a strong CO defect and accumulate unresolved JMs [19],[22]–[24], while in S. cerevisiae, mus81 or mms4 mutants show only a minor CO loss and resolve the vast majority of JMs [25]–[29]. Thus, in budding yeast, most meiotic JMs must be resolved by other, yet unidentified endonucleases. It also is not clear whether or not the Mus81 complex resolves JMs that form during the mitotic cell cycle. A recent study of I-Sce1 endonuclease-promoted mitotic recombination in S. cerevisiae suggested redundant roles for the Mus81 complex and for the Yen1 endonuclease in interhomolog CO formation [30], but it remains to be established that these crossovers are produced by dHJ-JM resolution.
dHJ-JMs can also be resolved by an endonuclease-independent process, called dissolution, that uses a RecQ-family helicase and a type 1 topoisomerase to disassemble JMs and to produce only NCOs [31]–[34]. Dissolution has been demonstrated in biochemical studies of the human BLM helicase combined with the TOPOIIIα/BLAP75 heterodimer, and of the corresponding budding yeast proteins Sgs1 and Top3/Rmi1 [35], [33], [36]. Dissolution has not yet been directly demonstrated in vivo, but is consistent with observations that loss of BLM or Sgs1 helicase activity is accompanied by a substantial increase in mitotic sister chromatid exchange [37]–[39], and that sgs1 mutants show increased JM accumulation and CO formation during mitotic DSB repair [16], [15]. During meiosis, sgs1 single mutants show only a slight increase in COs, but produce “abnormal” JMs involving 3 or 4 chromatids at elevated levels [40], [41]. In addition, the CO and JM formation defects of mutants lacking SC components are partially suppressed by sgs1 mutation [40], [42], [41]. These findings are consistent with the suggestion that the Sgs1/BLM helicase prevents COs by reducing JM levels. However, because this helicase also has the potential to disassemble early strand invasion intermediates that are precursors to JMs [43], [44], it remains to be determined if Sgs1/BLM act primarily to prevent JM formation, or to disassemble JMs once they form.
Finally, JMs that form during the G1 phase of the mitotic cell cycle can, in theory, also be resolved passively by chromosome replication [45], producing a CO if the original JM contains an odd number of HJs and an NCO if the original JM contains an even number of HJs.
In the current study, we present experiments aimed at examining how JMs are resolved during the S. cerevisiae mitotic cell cycle. Although several groups have detected JMs in S. cerevisiae undergoing vegetative growth [46], [47], [15], definitive study of their resolution has been precluded by their relatively low levels and by the fact that most form between sister chromatids. However, interhomolog JMs can be recovered at high levels during meiosis, especially in cells that lack Ndt80, a transcription factor required for expression of many mid- and late-meiosis proteins, including the Cdc5 polo-like kinase which is required for meiotic JM resolution [48], [17]. ndt80 mutant cells arrest at the pachytene stage of meiosis, with duplicated but unseparated spindle pole bodies [49], with homologs tightly paired by SC [49], and, most important to this study, with a high level of unresolved JMs [8]. To examine resolution of these JMs in a cellular environment that mimics the mitotic cell cycle, we used a singular property of S. cerevisiae, called return to growth (RTG). When cells in meiosis I prophase are shifted to rich medium, they rapidly exit meiosis, adopt a G1-like transcription pattern, and ultimately resume the mitotic cell cycle [50]–[58].
We report here the first molecular characterization of JM resolution during RTG. We show here that, unlike in meiosis, most JMs are resolved after RTG in a manner that does not produce COs. Examination of JM resolution in sgs1 and in mus81 mutants suggest that, during RTG of wild-type cells, the majority of JMs are resolved by Sgs1-mediated dissolution, with a minor fraction of JMs being resolved by Mus81 complex-dependent cleavage to produce both CO and NCO products.
To determine how JMs are resolved after RTG, we used ndt80Δ mutant cells, which arrest at pachytene with fully-formed SC and high levels of JMs [49], [8]. In general, RTG experiments involved incubating ndt80Δ cells in nutrient-poor sporulation medium (1% potassium acetate) for 7 hr to allow cells to initiate meiosis and arrest at pachytene, and then shifting cells to nutrient-rich growth medium (YPD) to induce RTG. We confirmed that ndt80Δ cells retain viability after RTG [49]; virtually all cells produced colonies when a culture incubated 7 hours in sporulation medium was plated on YPD agar plates (colonies/visible cells = 1.0+/−0.1; strain MJL3164—see Table S1). To examine the timing and efficiency of RTG in greater detail, we monitored progression of the first cell cycle after RTG (Figure 1). Budded cells were first observed 2 hr after RTG, and half of the cells had produced a bud by 2.5 hr. Nuclear division occurred about 1 hr after bud emergence, with half of the cells having undergone nuclear division by 3.5 hr after RTG. By 4 hr after RTG, virtually all cells had undergone nuclear division, consistent with the high viability seen in plating experiments.
Cells of the SK1 strain background used here complete a mitotic cell cycle every 80 minutes while growing in YPD (M. L., unpublished data), whereas in the current experiments, the first cell division did not occur until at least 2.5 hr after the shift from sporulation to YPD growth medium (Figure 1b). This difference might be explained if nuclear division during RTG was delayed by the presence of unresolved interhomolog connections that were formed during meiosis. To test this suggestion, we examined RTG in spo11 mutant cells (strain MJL2807), which do not initiate recombination or produce SC [59], [60]. Bud emergence and nuclear divisions occurred at times similar to those seen in SPO11 cells (Figure 1b), indicating that the extended gap phase seen upon RTG is not caused by a need to resolve recombination-dependent meiotic chromosome structures.
ndt80Δ cells arrest with chromosomes that are fully paired by SC [49]. It was previously shown that the SC formed in NDT80 cells breaks down rapidly after RTG [56]. We confirmed this observation in ndt80Δ strains by staining surface-spread nuclei for Zip1, a central component of the SC [61]. Most cells lose full-length linear SC within 15 minutes of transfer to YPD, and less than 30% of cells contained even residual (dotty) Zip1-containing structures 1.5 hr after RTG, before bud emergence and well before nuclear division (Figure 1c, 1d).
The first nuclear division of meiosis involves segregation of homologs (reductional division), whereas during mitotis, sister chromatids separate from each other (equational division). To determine if the first nuclear division after RTG is reductional or equational, we used a TRP1/trp1 heterozygous strain. TRP1 is tightly linked to the centromere of chromosome IV (<0.5cM; [62]), so chromosome segregation in the first division after RTG can be determined by examining TRP1 allele segregation (Figure 2a). If the first division is reductional, one daughter cell will inherit both copies of the TRP1 allele, whereas the other will inherit both copies of the trp1 allele, resulting in a sectored Trp+/Trp− colony. If the first division is equational, both daughter cells will inherit one TRP1 and one trp1 allele, resulting in a uniform Trp+ colony. A TRP1/trp1 ndt80Δ/ndt80Δ diploid (strain MJL3163) was induced to undergo meiosis for 7 hr, returned to growth by plating on YPD, and the resulting colonies were replica plated onto medium lacking tryptophan. Only one colony in 2767 was sectored, and the rest were uniformly Trp+ (Figure 2b). Thus, the first nuclear division after RTG involves a mitosis-like equational chromosome segregation.
Because DNA replication can resolve JMs, it was important to determine whether or not cells undergo replication before the first division after RTG. During the mitotic cell cycle, bud emergence is closely followed by initiation of DNA replication [63]. We asked if bud emergence after RTG was also associated with DNA replication. ndt80Δ cells arrest after meiotic DNA replication, and thus have a 4C DNA content. Therefore, DNA re-replication before the first division after RTG will result in tetraploid daughter cells. On the other hand, if DNA re-replication does not occur after RTG, diploid daughter cells will be produced. To determine whether DNA re-replication occurs after RTG, we monitored the copy number of chromosome V, using a centromere-linked array of tet operator (tetO) repeats that bind a constitutively-expressed tet repressor-green fluorescent protein fusion [64], [65], referred to here as CEN5-GFP. To check the efficiency of detection of individual CEN5-GFP signals, diploids that were hemizygous (strain MJL3312) or homozygous (strain MJL3313) for CEN5-GFP were grown to log phase, and the number of GFP dots per nucleus was scored in unbudded cells (G1-phase of the cell cycle). As expected, unbudded cells with a hemizygous CEN5-GFP showed one dot per nucleus (133/133). In contrast, 28/104 unbudded cells homozygous for CEN5-GFP showed two dots in their nuclei (Figure 2d), indicating that two copies of CEN5-GFP are detected with about 25% efficiency. The reduced efficiency of detection of two GFP spots is most likely a result of the limited separation of centromeres during interphase in yeast, due to the close attachment of centromeres to the spindle pole body [66].
Using this assay, we determined the number of GFP dots in unbudded cells produced from the first or second division after RTG of a diploid with a hemizygous CEN5-GFP (strain MJL3312). Re-replication followed by an equational division would result in each daughter cell inheriting two copies of CEN5-GFP, and two GFP dots will be observed in the nucleus (Figure 2c). However, if no re-replication occurs, each daughter cell will inherit one copy of CEN5-GFP, resulting in one GFP dot in the nucleus. All cells examined (282/282) showed only one dot in each nucleus. Thus, cells do not undergo DNA replication before the first nuclear division after RTG.
To confirm the conclusion that cells do not undergo DNA replication before the first nuclear division after RTG, we monitored the copy number of the loosely centromere linked MAT locus. Re-replication, followed by an equational division, would result in most daughter cells being MATa/MATa/MATα/MATα tetraploids. However, if no re-replication occurs, most daughter cells will be MATa/MATα diploids. Sporulation of MATa/MATa/MATα/MATα tetraploid cells would frequently produce MATa/MATα nonmating diploid spores. On the other hand, sporulation of MATa/MATα diploid cells will only produce haploid spores with a single MATa or MATα allele (Figure S1).
To sporulate cells that are phenotypically Ndt80−, we used a strain (strain MJL3430, pGPD1-GAL4-ER pGAL1-NDT80; [67], [68], [10]) where NDT80 is normally not expressed, but where NDT80 expression can be induced by the addition of estradiol (ED). Seven independent segregants from RTG performed without NDT80 expression (without ED) were induced to undergo a second meiosis with NDT80 expression (with ED), and tetrads produced by these strains were dissected. All spores from 4 spore-viable tetrads (at least 10 tetrads per primary segregant; n = 400) were either MATa or MATα maters, and none were MATa/MATα nonmaters, confirming the conclusion that re-replication does not occur before the first nuclear division after RTG.
Since unresolved JMs are expected to interfere with chromosome segregation at mitosis, the observation that most ndt80 mutant cells retain viability after RTG ([49]; see above) suggests that meiotic JMs must be resolved before the first cell division after RTG. During meiosis, JMs are mainly resolved to produce COs [8]–[10]. To ask if JMs are resolved similarly after RTG, we monitored segregation of the recessive cycloheximide–resistance allele, cyh2-z, in a cyh2-z/CYH2 heterozygous diploid. In wild-type meiosis, 66% of cells undergo second division segregation for cyh2-z, resulting from crossing over between the CYH2 locus and the centromere of chromosome VII (CEN7; see Materials and Methods). If JMs are similarly resolved as COs during RTG, 66% of cells are expected to have a CO between CYH2 and CEN7. Assuming random sister chromatid segregation at the first division after RTG, as it is in mitosis [69], half of the cells with a CO between CEN7 and CYH2 will produce cycloheximide-resistant cyh2-z/cyh2-z daughter cells (33% of total colonies; Figure 3a).
To directly compare JM resolution after RTG and during meiosis, we used an ndt80Δ/ndt80Δ CYH2/cyh2-z strain that contains an estrogen-inducible CDC5 gene (ndt80Δ pGPD1-GAL4-ER pGAL1-CDC5; strain MJL3267), to allow conditional JM resolution [10]. In the absence of inducer (-ED), cells accumulate in pachytene with unresolved JMs. ED addition induces CDC5 expression, and cells exit from pachytene and resolve JMs to produce COs, but do not progress further through meiosis [10]. Thus, if CDC5 is expressed before RTG, JMs will be resolved and COs will be produced at a level similar to that seen in meiosis. Thus, 33% of colonies are expected to be cycloheximide resistant (Figure 3a). Cells were induced to undergo meiosis for 7 hr, and then aliquots were plated on YPD to undergo RTG (Figure 3b). The remainder of the culture was incubated for another 4 hr in sporulation medium, either with ED to induce pachytene exit, or in the absence of ED as a control, and aliquots were plated on YPD. Colonies on YPD were replica plated onto YPD with cycloheximide to score for sectored colonies produced by crossovers. Only a small fraction of the RTG colonies from samples taken before mock or CDC5 induction contained cycloheximide-resistant sectors (3.9% and 2.6%, respectively, Figure 3c, 3d), and cells plated after a 4 hr incubation without ED also produced few cycloheximide-resistant sectors (4.6%, Figure 3e). In contrast, when CDC5 was expressed and JMs resolved as COs, 30% of colonies contained cycloheximide-resistant sectors (Figure 3f). The relatively low frequencies of colonies with cycloheximide-resistant sectors in all samples that underwent RTG without CDC5 induction indicates that the majority of JMs are not resolved as COs after RTG.
Reduced CO formation after RTG was confirmed by molecular analysis. To allow direct comparison between events that occur during meiosis and during RTG, we used a recombination-reporter strain, described below, that also contained the estrogen-inducible NDT80 allele described above (strain MJL3430) that confers reversible pachytene arrest [68]. Pachytene-arrested cells can be transferred to YPD without estradiol addition to undergo RTG in the absence of NDT80 expression. Alternatively, they can be kept in sporulation medium, and by adding ED to induce NDT80 expression, be made to complete meiosis (Figure 4a, 4b). Meiotic NDT80 expression resulted in meiotic divisions (Figure 4d), spore formation (data not shown), and the rapid expression of CDC5, a known target of Ndt80 [70]. Cdc5 was detected one hr after addition of ED to meiotic cultures, whereas Cdc5 was not present in RTG cultures until 2–2.5 hr after the shift to YPD, about 30 min before nuclear division (Figure 4c, 4e). The mitotic cyclin Clb2, which is not produced during meiosis [71], was observed only in the RTG culture, at about the same time as Cdc5 (Figure 4c).
Recombination intermediate resolution and recombinant product formation were monitored at the molecular level, using a recombination reporter system [7] (Figure 4f). JM resolution initiated at similar times in both ED-induced meiotic and RTG cultures (Figure 4g). However, the two cultures differed markedly in terms of CO production. JM resolution in the meiotic culture was accompanied by a marked increase in crossovers in the same time interval, and was complete by 1.5 hr after Ndt80 induction (Figure 4h). In contrast, no increase in COs was seen in the first 2 hr after RTG, during which JMs decreased by five-fold. After two hr, a time that corresponded to the time of bud emergence (Figure 4e), resolution of the remaining JMs was accompanied by a modest increase in COs (Figure 4h). NCO products were produced in meiotic and in RTG cultures at similar levels (Figure 4i). Similar results were observed in RTG experiments using ndt80Δ cells lacking the inducible NDT80 system (strain MJL3164; Figure S2).
The data presented here support the conclusion from genetic experiments described above, that most JMs are resolved after RTG without producing COs. The CO increase seen after 2 hr indicates that surviving JMs can be resolved as COs during the later stages of RTG.
The Mus81 complex plays a major role in JM resolution during meiosis in S. pombe and a less prominent role in meiotic JM metabolism in S. cerevisiae [19], [26], [20], [27], [22], [24], [72], [28]. To determine if the Mus81 complex resolves JMs after RTG, ndt80Δ mus81Δ cells (strain MJL3389) were induced to undergo meiosis for 7 hr and then transferred to YPD. Bud emergence and nuclear division occurred at times similar to those seen in ndt80Δ MUS81 cells (Figure 5a, compare to Figure 1b). JMs were resolved completely after RTG (Figure 5b). A modest net increase in noncrossovers was seen (Figure 5d), similar to that seen in MUS81 cells (see Figure 4i). Unlike in wild-type, where JM resolution after two hr was accompanied by an increase in COs, no significant CO increase was observed after RTG in mus81Δ mutants (Figure 5c). These data indicate that the Mus81 complex is not required for JM resolution after RTG, but it may play an important role in the limited JM resolution as COs that occurs at later stages.
The BLM and Sgs1 helicases, in combination with topoisomerase III and Rmi1/BLAP45, resolve dHJs in vitro as NCOs [33], [36]. To ask if Sgs1 has a similar role in JM resolution after RTG, we used an sgs1 mutant allele (strain MJL3388; sgs1-ΔC795) that expresses only the first 652 amino acids of the protein [73], and which lacks both the helicase domain and a region (the HRDC domain) which in BLM interacts with Holliday junctions [74]. Although bud emergence occurred at a similar time after RTG in sgs1-ΔC795 and in SGS1 cells, nuclear division was 1.5–2 hr later in sgs1-ΔC795 than in SGS1 (Figure 6a, compare to Figure 1b). A recombination-null ndt80Δ sgs1-ΔC795 spo11 triple mutant (strain MJL3428), which does not produce JMs, underwent nuclear division without this delay (Figure 6a), suggesting that the delay in nuclear division seen in sgs1-ΔC795 might result from a delay in JM resolution.
To ask if JM resolution is delayed in ndt80Δ sgs1-ΔC795 cells, we monitored JMs and recombination products, using the molecular assay system described above. As was previously described [41], ndt80Δ sgs1-ΔC795 cells accumulate high levels of intersister JMs, and JMs with more than two chromatids (multi-chromatid JMs; mcJMs), in addition to the dHJ-JMs that accumulate in ndt80Δ SGS1 cells (Figure 6b). Resolution of all JM species was delayed by about 1 hr in sgs1-ΔC795 as compared to SGS1. While the vast majority of JMs resolved in SGS1 by about 2.5 hr after RTG (Figure 4g), more than half of total JMs remained unresolved in sgs1-ΔC795 at the same time, although all JMs resolved by 4 hr (Figure 6b). Thus, loss of the Sgs1 helicase results in a substantial delay in JM resolution after RTG.
Delayed JM resolution after RTG in sgs1-ΔC795 was accompanied by altered recombinant product formation. COs increased only slightly in the first 1.5 hr after RTG (Figure 6c), but there was also only a slight increase in NCOs during the same period (Figure 6d). After 1.5 hr, JM resolution was accompanied by an increase in both COs and NCOs (Figure 6c, 6d). Thus, in both SGS1 and in sgs1-ΔC795, few COs are produced during the first 1.5–2 hr after RTG, with substantially greater CO formation at later times. However, unlike in SGS1, where most NCOs appear in the first 1.5–2 hr after RTG, NCO production in sgs1-ΔC795 is delayed until the time that COs also appear.
Most JM intermediates formed during budding yeast meiosis are produced by interhomolog recombination and are resolved as COs, and the majority of meiotic COs derive from interhomolog JMs [8], [9], [17], [10]. In contrast, interhomolog JMs and COs are less prominent during the mitotic cell cycle. Most JMs produced during mitotic DSB repair involve sister chromatids [15], and only a minor fraction (typically 5–10%) of mitotic recombination involves crossing-over, as would be expected if interhomolog JMs are rarely resolved as COs during the mitotic cell cycle [16], [75]. Testing this suggestion has, to date, been limited by the very low levels of interhomolog JMs produced in vegetatively-growing cells, even when initiating DSBs occur at levels similar to those seen in meiosis [15].
In this paper, we used RTG as an alternate approach to the study of JM resolution during the mitotic cell cycle. Although aspects of RTG have been examined in many studies [50]–[58], interpretation has been complicated by the relatively poor synchrony of yeast meiotic cultures. Thus, RTG samples from normal meiotic cultures can contain cells with unrepaired DSBs, cells with repaired DSBs but unresolved recombination intermediates, and cells where intermediates already have been resolved. To avoid complications inherent in the analysis of such a complex mixture, we performed RTG using meiotic cultures of ndt80 mutant cells, which arrest at a single stage of meiosis (pachytene), with chromosomes fully paired by synaptonemal complex and with high levels of interhomolog JMs. This has provided insight into features of the mitosis-like cell cycle that immediately follows exit from meiosis, and into mechanisms of the recombination intermediate resolution.
When transferred from sporulation to growth medium, yeast cells degrade most meiotic transcripts within 20 min, and return to a pattern of gene expression that roughly resembles the G1 phase of the mitotic cell cycle [57]. Despite this rapid change in transcription patterns, cells spend an extended lag period (1.5 to 3 hours, equivalent to one or two normal mitotic cell cycles) before they undergo bud emergence, the first outward sign of resumed growth (Figure 1). Although cells disassemble synaptonemal complex and resolve meiotic recombination intermediates during this period ([56], this work), a similar lag before bud emergence is seen in spo11 mutants (this work), and also if SC disassembly and JM resolution occur before RTG, by virtue of Cdc5 induction in ndt80Δ CDC5-IN cells (Y.D. and M.L., unpublished observations). It is therefore likely that this extended gap phase represents the time needed for metabolic adjustment to the shift from acetate to glucose, and from nitrogen-depleted to nitrogen-rich medium, rather than the time needed to disassemble meiosis-specific chromosome and DNA structures.
During the mitotic cell cycle, bud emergence is accompanied by the initiation of chromosome replication [63], but this is not the case during RTG. We used two different approaches to confirm that bud emergence occurs without DNA replication after RTG [53]. This could be the consequence of a failure to express completely the ensemble of proteins necessary for DNA replication. While some replication protein-encoding genes are transcribed after RTG ([57], Lea Jessop and M. L., unpublished observations), transcripts of DBF4 and CDC7, which encode a kinase critical for replication origin firing, are rapidly reduced upon RTG [57]. Re-replication may also be blocked if cyclin-dependent kinase remains at post-S phase levels throughout RTG, which would prevent origin re-licensing [76]–[78].
We also find that the first nuclear division after RTG involves an equational division, unlike the reductional division that occurs during meiosis I. Reductional division at meiosis I requires the loading, at kinetochores, of the meiosis-specific protein complex monopolin, which promotes co-orientation of sister kinetochores towards a single spindle pole [79], [80]. Monopolin contains a meiosis-specific protein, Mam1, and two nucleolar proteins, Csm1 and Lrs4, whose kinetochore localization requires Cdc5 activity [79], . Meiotic CDC5 transcription requires NDT80, and MAM1 transcripts are reduced in ndt80 mutants [70] and rapidly decline upon RTG [57]. In addition, monopolin loading at kinetochores requires active Cdc7/Dbf4 kinase [82], which is most likely not produced after RTG [57]. Therefore, it is unlikely that monopolin is loaded at kinetochores during RTG of ndt80Δ cells, and thus it is not surprising that the first nuclear division after RTG is equational.
Most of the Holliday junction-containing JMs that accumulate during meiosis in ndt80 mutants are resolved as COs upon restoration of either NDT80 or CDC5 gene expression ([10], this work). In contrast, our genetic and molecular analyses show that most of the JMs that form during wild-type meiosis are resolved without crossover formation during RTG. This indicates that mechanisms of JM resolution that operate during RTG differ from those that operate during meiosis.
There are three general mechanisms for dHJ-JM resolution: endonuclease cleavage; helicase/topoisomerase-mediated dissolution; and replication (Figure 7a–7c). Of these, replication and dissolution produce only NCO products, while endonuclease cleavage can, in principle, produce either COs or NCOs, depending upon the orientation of the two cleavage reactions. Since most dHJ-JMs resolve as COs during meiosis, meiotic resolution must involve endonuclease cleavage, and this cleavage must be constrained so that the two Holliday junctions are usually cut in opposite directions (see Figure 7a).
In contrast, JM resolution during RTG appears to occur in two phases with different outcomes (Figure 7d–7f). In wild-type cells, about 80% of JMs disappear during the first 1.5–2 hr after RTG. Few COs are produced during this period, and NCOs increase to near-final levels. The greatest net increase in COs occurs at 2 hr and later (Figure 7e), when the remaining 20% of JMs are resolved (Figure 7d). Thus, RTG appears contain an initial period (hereafter called early RTG) that precedes bud emergence, during which SC breaks down (Figure 1c) and the majority of JMs resolve without CO formation (Figure 7d, 7e). During the second period (hereafter called late RTG), between bud emergence and nuclear division, JM resolution is accompanied by CO formation.
JM resolution without CO formation, which predominates during early RTG, could occur by endonucleolytic cleavage that is constrained to produce only NCOs, by dissolution, or by replication (Figure 7a–7c). Resolution by replication is unlikely, since all available evidence indicates that the first cell division after RTG occurs without prior replication (this work, [53]). Both JM resolution and NCO formation are significantly reduced during early RTG in sgs1-ΔC795 mutant cells (Figure 7d, 7f), which lack both the helicase and Holliday junction-binding domains of this RecQ helicase [73], [74]. The most parsimonious interpretation of these data is that, in wild-type cells, JM resolution during early RTG occurs primarily by dissolution, catalyzed by Sgs1 and Top3/Rmi1, as has been observed in vitro [36]. However, it is formally possible that other activities are responsible for the initial phase of JM resolution in wild-type, and that, unlike in wild-type, the majority JMs that form during sgs1-ΔC795 meiosis have structures that are refractory to resolution by these hypothetical activities.
During budding yeast meiosis, the Sgs1 helicase acts with Mus81/Mms4 to prevent the accumulation of abnormal recombination intermediates [28], [29]. Normal JM intermediates are protected from Sgs1 by components of the synaptonemal complex, and sgs1-ΔC795 partially suppresses the JM deficit observed in mutants lacking SC components [40], [42], [41]. These and other observations have been interpreted as indicating that Sgs1 acts primarily to prevent JM formation during meiosis. Our current data indicate that, in addition to preventing JM formation, Sgs1 can also dissolve JMs in vivo, but is prevented from doing so during meiosis by the SC. This suggestion is also supported by the finding that most JMs are resolved without CO production upon Cdc5-independent SC breakdown in pachytene-arrested meiotic cells (Anuradha Sourirajan, Arnaud de Muyt and M. L., unpublished observations).
While JM resolution during early RTG is rarely accompanied by CO production, JMs that survive this initial phase appear to be resolved frequently as COs. This is seen in wild-type, but is most evident in sgs1-ΔC795 mutant cells, where an increase in the rate of JM resolution during late RTG is accompanied by a marked increase in both CO and NCO recombinants (Figure 7e, 7f). Because COs can only be produced by endonuclease-mediated JM cleavage, this suggests that a Holliday junction resolvase is activated 1.5–2 hr after RTG, a time that is also marked by bud emergence. We do not know the regulatory change that is responsible for this change in modes of JM resolution, but it is worth noting that both Cdc5 and the G2/M phase cyclin, Clb2, are first produced at this time (Figure 4c).
During meiosis, the Cdc5 kinase triggers JM resolution as COs [10], suggesting an obligate cleavage of JM Holliday junctions in opposite directions (Figure 7a). In contrast, JM resolution during late RTG of sgs1-ΔC795 mutants produces both COs and NCOs (Figure 7e, 7f), as would be expected for the mixed parallel and opposite cleavage patterns contained in the original DSBR model ([83], see Figure 7a). This apparent difference in resolution mechanisms may reflect the chromosome environment in which intermediates reside. While JM resolution during late RTG occurs in the absence of detectable SC, crossover-designated meiotic JMs are thought to reside in SC-associated structures, called late recombination nodules, that contain the Holliday junction-binding proteins Msh4/Msh5 and associated Mlh1, Mlh3 and Exo1 proteins [84]–[86]. In mlh1, mlh3, and exo1 mutants, meiotic JM levels are normal but crossover formation is reduced roughly two-fold [87], [88], consistent with the suggestion that the Mlh1/Mlh3/Exo1 components of late recombination nodules direct nuclease-mediated meiotic JM resolution towards a crossover-only outcome. In the absence of such specialized chromosome structures, nuclease-mediated JM resolution may be more evenly divided between COs and NCOs, in both mitotic and meiotic cells.
Although the nuclease(s) responsible for dHJ resolution during either meiosis or during RTG remain to be determined, it is worth noting that CO formation during RTG is even more reduced in mus81Δ mutants than in wild-type (Figure 7e), and the increase in COs seen during late RTG in wild-type and in sgs1-ΔC795 is not seen in mus81Δ mutants. In many organisms, including S. cerevisiae, the Mus81 nuclease complex is dispensable for most meiotic COs [26], [89]–[91], and the majority of meiotic JMs resolve in a timely manner in S. cerevisiae mus81 or mms4 mutants [27], [28]. In addition, it has been reported that intact Holliday junctions are a relatively poor substrate for the Mus81/Mms4 nuclease, while junctions with one nicked strand are resolved efficiently [92], [22], [93]. On the other hand, MUS81 is required for timely disappearance of X-shaped DNA molecules that form in methyl methanesulfonate-treated rmi1-ts cells [94]. This would suggest a role for Mus81/Mms4 in resolving these JMs, whose structure remains to be determined.
Our data suggest that Mus81/Mms4 has a role in resolving the JMs that survive until late RTG, but it does not appear to be active during early RTG. It is possible that either Mus81/Mms4 or a junction nicking activity that converts HJs into a Mus81/Mms4 substrate are absent during early RTG. Alternatively, the Mus81 complex may be modified during late RTG so that it resolves intact Holliday junctions unassisted. The latter suggestion, if correct, might explain the failure to observe robust Holliday junction resolution activity in most biochemical studies [95].
In this work, we have shown that Holliday junction-containing recombination intermediates, formed during meiosis, are resolved during RTG in a manner that substantially reduces CO production. To the extent that recombination is regulated similarly during RTG and during the mitotic cell cycle, and to the extent that similar recombination intermediates are present, this finding can help explain the relatively low yield of COs during mitotic recombination. In particular, our findings reinforce the identification of the BLM family of RecQ helicases as playing an important role in suppressing CO recombination during the mitotic cell cycle [38]. Our findings also suggest that the Mus81 complex is the primary nuclease responsible for mitotic CO recombination [30]. Our finding, that these two enzymes act during different phases of the period before the first cell dvision after RTG, raises the intriguing possibility that the mitotic cell cycle may be similarly partitioned. It is attractive to suggest that helicase-mediated dissolution predominates during most of the mitotic cell cycle, with endonuclease-mediated JM cleavage being activated at the end. This would minimize the potential for CO-mediated loss of heterozygosity and chromosome entanglement, while preserving the ability to resolve JMs that escape dissolution before the initiation of mitosis.
In applying conclusions regarding JM resolution during RTG to the mitotic cell cycle, it should be kept in mind that these processes are not identical. For example, RTG involves the disassembly of chromosome structures that are not present during the mitotic cell cycle, as well as S-phase bypass, and both of these differences have the potential to affect modes of JM resolution. It will be of considerable interest to examine, during RTG, patterns of expression and modification of proteins involved in recombination, repair, and cell cycle progression during meiosis and the mitotic cell cycle.
Strains are listed in Table S1 and are SK1 derivatives [96]. The URA3-ARG4 recombination interval has been described [7]; cyh2-z is a spontaneous cycloheximide resistance mutation (CyhR); spo11-Y135F [97] was a gift from S. Keeney; mus81Δ and sgs1-ΔC795 have been described [42], [28]. Strains with estrogen-inducible CDC5 and NDT80 alleles (pGPD1-GAL4-ER pGAL1-CDC5 and pGPD1-GAL4-ER pGAL1-NDT80, respectively) have been described [10]. Strains were constructed by genetic crosses, or by transformation. Media formulae were as described [98], [99].
Sporulation was as described [99] using 400 ml cultures in a 2.8 liter baffled Fernbach flask (BellCo Glass) with a cell density of 2x 107 cells per ml at the beginning of sporulation. For RTG experiments, cells were induced to undergo meiosis for 7 hr, harvested by centrifugation, resuspended in an equal volume of liquid YPD (prewarmed to 30°C) and aerated with vigorous shaking at 30°C in conditions similar to those used for sporulation. For plating experiments, samples were sonicated twice for 5 seconds at baseline power (Microson XL 2005), diluted appropriately and then plated on YPD plates. To determine colony-forming units, samples were counted in a hematocytometer and the concentration of cells was determined; cells with unseparated buds were counted as a single entity. For Ndt80 or Cdc5 induction, β-estradiol (ED; Sigma; 5 mM stock in ethanol) was added to a final concentration of 1 µM. For no Cdc5-indcuation control experiments, the same amount of ethanol (without ED) was added. For RTG after Cdc5 induction during meiosis, cells were washed twice with sporulation medium lacking ED at 30°C before resuspension in YPD.
Unless stated otherwise, all data presented are the average of two independent experiments; error bars in plots indicate standard error.
To score bud emergence and nuclear division, 1 ml of a culture was mixed with 1 ml of ethanol and stored at 4°C. Just before examination, 1 µl of 1 mg/ml 4′,6-diamidino-2-phenylindole (DAPI) was added and samples were left for 5 min at room temperature, washed once with an equal volume of water and resuspended in 0.5 ml water. Cell morphology was scored using phase contrast or differential interference contrast microscopy and nuclear morphology by DAPI epifluorescence microscopy, using a Zeiss Axioplan 2 epifluorescence microscope and a QICAM camera. Images were acquired using QCapture 3.1.1 and processed with Adobe Photoshop CS3.
GFP chromosome dot visualization was done using cells fixed in 3.7% formaldehyde as described [65]. Vectashield with DAPI (Vector Laboratories) was used to simultaneously stain DNA. Cells were counted as having two GFP dots if two separated GFP dots could be clearly visualized. Sample fluorescence was visualized using a Zeiss Axioplan 2 epifluorescence microscope and a Micromax 1300 CCD camera. Images were acquired using IPlab 3.7 and processed with Adobe Photoshop CS3.
Nuclear spreads were performed and stained as described [100] using cells from 5 ml of culture. Zip1 was detected using anti-Zip1 rabbit polyclonal sera (a gift from G.S. Roeder, 1∶100 dilution) as the primary antibody and Alexafluor 488 conjugated goat anti-rabbit IgG (Molecular Probes #A11034) at 1∶100 as the secondary antibody. To visualize DNA, 40 µl of Vectashield with DAPI (Vector Laboratories) was added. Sample fluorescence was visualized using a Zeiss Axioplan 2 epifluorescence microscope and a Micromax 1300 CCD camera. Images were acquired using IPlab 3.7 and processed with Adobe Photoshop CS3.
During RTG, cells lose synchrony and continue to further cell cycles, complicating calculation of a cumulative cell division curve. We assumed that bud emergence and nuclear division occur with the same relative timing in the first and second cell division after RTG. To distinguish between daughter and mother cells, we took advantage of the fact that after RTG, ndt80Δ cells produce an elongated bud that can be easily distinguished from the round mother cell (Figure 1a). The fraction of cells that had not yet budded (unbudded cells) was calculated according to the equation: unbudded cells = (X1-Y1)/Z1 where X1 = unbudded round cells (i.e. cells before the first mitotic division), Y1 = unbudded elongated cells (i.e. products of the first mitotic division) and Z1 = total cells counted. At late times, due to continuous division of the cells, the number of cells that have already undergone the first mitotic division (Y1) can exceed the number of cells that have not undergone a mitotic division (X1). In such a case, (X1-Y1) was set to zero.
The fraction of cells that had undergone the first nuclear division (post-division) was calculated according to the equation: post-division = X2/Y2 where X2 = round cells that were undergoing mitosis (detected as budded with a nucleus stretched between the mother and daughter cells) plus all elongated cells with a nucleus (i.e. cells that have already completed the first mitotic division) and Y2 = all round cells. At late times, due to continuous cell division, X2 may be greater than Y2. In such a case, the fraction of post-division cells was set to one.
DNA preparation and analysis on Southern blots were as described [101], [8]. XhoI and XmnI digests were probed with ARG4 coding sequences (+165 to +1413). XhoI/EcoRI double digests were probed with HIS4 coding sequences (+538 to +718).
Protein was prepared from 4 ml of sporulating culture by TCA precipitation [102]. 5 µl samples of each extract were displayed on 7.5% polyacrylamide Tris-Glycine pre-cast gels (Bio-Rad) and electroblotted to a PVDF membrane (Invitrogen), using an iBlot Dry Blotting System (Invitrogen) as recommended by the manufacturer. Blots were washed for at least one hr on an orbital shaker at room temperature in blocking buffer, 0.2% I-block (Tropix) in PBST (0.15 M NaCl, 0.053 M Na2HPO4, 0.008 M KH2PO4, 0.05% v/v Tween-20, pH 7.4). Primary antibody, diluted in blocking buffer, was added to the blot and incubated on an orbital shaker at room temperature for at least one hr. Blots were washed four times for 15 min with blocking buffer, incubated with secondary antibody for one hr with shaking at room temperature, and wash steps were repeated. Signal was developed using the chemiluminescent CDP-star substrate (Applied Biosystems), detected using a Fuji LAS3000 CCD camera, and quantified using ImageGauge V4.22 software (Fuji). Blots were stripped with OneMinute Western Blot Stripping Buffer (GM Biosciences) and reprobed for Arp7 as a loading control. Primary antisera were as follows: Arp7 – goat polyclonal (Santa Cruz Biotechnology, Inc; Sc-8961), 1∶500; influenza hemagglutinin (HA) – mouse monoclonal (5 µg/µl; Roche Applied Science; 12CA5), 1∶10,000; Cdc5 – goat polyclonal (Santa Cruz Biotechnology, Inc; Sc-6733), 1∶500; Ndt80 – rabbit polyclonal (a gift from K. Benjamin), 1∶10,000; Clb2 – rabbit polyclonal (Santa Cruz Biotechnology, Inc; Sc-9071), 1∶500. Secondary antibodies were alkaline phosphatase conjugates of goat-anti-mouse (Sigma, A3562), goat-anti-rabbit (Sigma, A3687) and rabbit-anti-goat (Sigma, A4187), all used at 1∶10,000.
To measure the frequency of recombination between the CYH2 locus and the centromere of chromosome VII, we measured second division segregation pattern of the TRP1 and CYH2 alleles in dissected tetrads from strain MJL3548 (CYH2/cyh2-z TRP1/trp1), using TRP1 as a centromere-linked marker [62]. Of 72 tetrads with 4 viable spores, 12 tetrads were parental ditypes, 12 were non-parental ditypes and 47 were tetratypes. One tetrad had gene conversion of cyh2-z and was not counted. Thus, as expected for a locus far removed from its centromere, the vast majority of cells undergo at least one crossover between CYH2 and CEN7, and about two thirds of cells produce spores with a crossover between the CYH2 locus and its centromere.
|
10.1371/journal.pcbi.1001079 | First Principles Modeling of Nonlinear Incidence Rates in Seasonal Epidemics | In this paper we used a general stochastic processes framework to derive from first principles the incidence rate function that characterizes epidemic models. We investigate a particular case, the Liu-Hethcote-van den Driessche's (LHD) incidence rate function, which results from modeling the number of successful transmission encounters as a pure birth process. This derivation also takes into account heterogeneity in the population with regard to the per individual transmission probability. We adjusted a deterministic SIRS model with both the classical and the LHD incidence rate functions to time series of the number of children infected with syncytial respiratory virus in Banjul, Gambia and Turku, Finland. We also adjusted a deterministic SEIR model with both incidence rate functions to the famous measles data sets from the UK cities of London and Birmingham. Two lines of evidence supported our conclusion that the model with the LHD incidence rate may very well be a better description of the seasonal epidemic processes studied here. First, our model was repeatedly selected as best according to two different information criteria and two different likelihood formulations. The second line of evidence is qualitative in nature: contrary to what the SIRS model with classical incidence rate predicts, the solution of the deterministic SIRS model with LHD incidence rate will reach either the disease free equilibrium or the endemic equilibrium depending on the initial conditions. These findings along with computer intensive simulations of the models' Poincaré map with environmental stochasticity contributed to attain a clear separation of the roles of the environmental forcing and the mechanics of the disease transmission in shaping seasonal epidemics dynamics.
| Nonlinearity in the infection incidence is one of the main components that shape seasonal epidemics. Here, we revisit classical incidence and propose a first principles derivation of the infection incidence rate. A qualitative analysis of the SIRS model with both the classical and the proposed incidence rate showed that the new model is physically more meaningful. We conducted a statistical analysis confronting the SIRS and SEIR models formulated using both incidence rate functions with four data sets of seasonal childhood epidemics. Two data sets were hospital records of cases of syncytial respiratory virus (RSV). The other two data sets were taken from the well-known UK measles epidemics database. We found that seasonal epidemics is better explained using our incidence rate model embedded in a Poisson sampling process. The results presented here are not by any means an exhaustive exploration of the interplay between nonlinear dynamics and stochasticity. Our results may be viewed as the starting point of multiple research avenues. Three such research topics could be: the first-principles derivation of non-linear incidence rate functions, the role of bistability and demographic stochasticity for disease persistence and the simulation of environmental and demographic stochasticity in the Poincaré map.
| A plethora of deterministic epidemic models involving susceptible , infected and recovered individuals have been proposed [1], [2], carefully analyzed [3]–[8] and confronted with data sets in the biomathematics and ecology literatures [9]–[12]. A well defined topic within this mathematical ecology research area is the study of -type models with seasonal forcing [13]–[16]. These models have proved to be useful for understanding the observed patterns and the natural processes behind human and non-human epidemics [17]–[21]. Here, we restrict our attention to the and models in which we introduce seasonal forcing while varying the structural form of the incidence rates. Two hypotheses pertaining the RSV and the measles transmission mechanisms were modeled with two simple functional forms of the incidence rates. We show that in doing so, we are able to attain a clear separation of the roles of the environmental forcing and the mechanics of the disease transmission in shaping the epidemics dynamics.
The construction of deterministic incidence rates functions is a critical building block of epidemiological modeling. In a seminal paper, Hethcote [1] showed that because there are many choices for the form of the incidence, demographic structure and the epidemiological-demographic interactions, there really is a plethora of incidence rate functional forms to choose from. Not surprisingly, the biomathematics literature abound in qualitative mathematical analyses of many of these functional forms [22]–[26]. However, biological first principles derivations of incidence rate functional forms are not too common. As we show in this study, using such first principles derivations greatly enrich the reaches of the practice of confronting models with data while testing biological hypotheses. Thus, despite the big amount of available functional incidence rates forms [1], we believe that the set of models chosen to be confronted with data should be restricted to those forms derivable from first principles. To illustrate this argument, in this study we first show that a simple probabilistic setting wherein infectious encounters are modeled with a pure birth stochastic process leads to a general nonlinear incidence form proposed previously by Liu [24] and later analyzed by Hethcote and Van Den Driessche [23] (hereafter we refer to the Liu, Hethcote and Van Den Driessche incidence rate as the LHD incidence rate). The LHD incidence rate leads to models with qualitatively different dynamics compared with the ones obtained using the classical incidence rate.
In the SIRS model with either incidence rate and seasonal forcing, becomes a periodic function of time and the trajectory “pursuits” a moving target thus giving rise to limit cycles. That moving target is the former endemic equilibrium that bounces back and forth between two points. In either model, the target switches between that moving point and the disease free equilibrium when crosses 1, giving rise to a period doubling bifurcation. In the SIRS model with classical incidence rate this mechanism does not depend on the initial conditions. In this work we show that the disease free equilibrium (DFE) is unconditionally an attractor in the SIRS model with LHD incidence rate. This leads to a scenario where two regions of attraction can coexist. The trajectory will either reach the disease free equilibrium or have periodic solutions depending on the initial conditions. Furthermore, after carrying a formal model selection we show that the SIRS model with LHD incidence rate leads to a significant fit improvement over the classical SIRS model with the same seasonal forcing. Finally, we compared the applicability and generality of the classical and LHD incidence rates functions by fitting them to two measles time series data sets. Using the later function leads to a vast improvement of model fit in both cases. Since we were fitting a deterministic SEIR model, we chose to use the data from the two largest cities in the measles data set (London and Birmingham, see http://www.zoo.cam.ac.uk/zoostaff/grenfell/measles.htm), where the effects of demographic stochasticity are expected to be less influential in the dynamics of the epidemics [10].
Varying the form of the contact rate function while including environmental stochasticity in the SIRS and SEIR models leads to a better understanding of the dynamics of an infectious disease transmission. Depending on the model and contact rate, the disease free equilibrium (DFE) is either a saddle point or an attractor. In the first case, if a trajectory located originally in the basin of attraction of the endemic equilibrium (EE) basin of attraction is perturbed with environmental noise, it may transiently visit the DFE stable submanifold and then return to the EE basin of attraction. If however the DFE and the EE coexist as stable equilibria, a trajectory initially at the EE basin of attraction may end up in the DFE basin of attraction. The interaction between stochasticity and the different contact rate models was studied using computer intensive simulations of the Poincaré map [27].
The classical model has been extensively studied in order to predict and understand various disease dynamics behaviors, as well as their spread and persistence [28]. For many infectious diseases, the pool of susceptible individuals is replenished due to the waning of immunity [17], [18]. To account for the lost of immunity, the classical susceptible , infected and recovered model is adjusted by allowing a fraction of the recovered individuals to move back into the susceptible pool at a rate . This susceptible, infected, recovered and susceptible model is expressed as(1)(2)(3)where is the rate of loss of infectiousness and the total population size remains constant (i.e. ). The constant represents both, the birth and mortality rates. Assuming that birth and mortality rates are equal is justified on the grounds that the annual infection rate is considerably higher than the population growth. The constant is the contact rate, the average number of individuals with whom one infected individual makes sufficient contact to pass on the infection [29]. The fraction represents the average number of infections per susceptible individual and hence represents the expected number of infections when susceptible individuals are available [5]. Note that the above definition of as a per individual constant leads to a consistency of the units within each of the model equations and assumes homogeneous mixing. In the following sections we will discuss different ways to model the incidence rate.
The equations for the classic SEIR (Susceptible-Exposed-Infectious-Recovered) model are as follows [30]:(4)(5)(6)(7)where represents both, the birth and mortality rates per capita. The mean latent and infectious periods of the disease are and . As written, the SEIR model has a stable endemic equilibrium provided . Further biological realism to model recurrent epidemics can be incorporated to both this SEIR model and the SIRS model above by assuming that the transmission rate varies seasonally. Indeed, Earn et al [30] study the range of the dynamical behavior of the SEIR model with seasonality and find it useful for explaining the measles numerous transitions between regular cycles and irregular, possibly chaotic epidemics. Also, Alonso et al. [31] show that noise amplification provides a possible explanation for qualitative changes from regular to irregular oscillations of lower amplitude. In this paper, we follow the suggestion made by Hethcote [1] and couple Liu, Hethcote and Van Den Driessche's incidence rate with seasonal forcing in both the SIRS and SEIR models.
To incorporate the claim that epidemics of recurrent infections is driven by seasonality, it is customary to depart from the standard incidence rate by assuming that the average number of incidences sufficient for transmission per infected individual , is a periodic or quasi-periodic function of time (). Often, the incidence rate is assumed to have a sinusoidal form of the type(8)where stands for the strength of the seasonality and year. Various authors have shown that such a generic description of the seasonal variation in transmission rates is not as revealing as a detailed description of the actual processes underlying the seasonal drivers of transmission through mechanistic seasonal forcing functions [11], [18], [30], [32], [33]. However, as we show in the results section, in some cases this sinusoidal function may unequivocally represent a linear transformation of a weather covariate. Although other authors have used a more flexible Haar step function for the seasonal forcing (e.g. [30]), we restrict ourselves to the incorporation of the sinusoidal form above (eq. 8) as the seasonal forcing. This has the advantage of ease of interpretation and qualitative analysis. In any case, the main purpose of incorporating the forcing is to explore the main qualitative characteristics of coupling the seasonally varying disease transmission and different incidence rate functional forms.
Brauer [34] generalizes the incidence rate definition in the following way: if the average member of the population makes contacts in one unit of time with , and if is the probability of choosing one infected individual from the population at random, then is the rate of new infections per unit of time. The mass-action incidence rate model is recovered using and the classic incidence rate is recovered by picking . A general incidence rate function was proposed by Hethcote and van den Driessche [23]:where and are constants. Consider the special case where and . Using Brauer's generalization and idea, Hethcote and van den Driessche's model is recovered using the function . Then, the incidence rate function becomeswhere and . Although the mathematical properties of the general function are known in general [23], [35], [36] a mechanistic, first principles derivation of it is still lacking. Such a derivation can be obtained using a probabilistic reasoning analogous to the argument used by [37] to model the Allee effect through stochastic mating encounters:
Through physical movement or any other means of dispersion, an infected individual will have contact with a given number of susceptible individuals in the population. The potential to effectively disperse the disease (virus) could be thought of as being proportional to that number of susceptibles with whom the infected individual makes contact: indeed, the more contact the infected individual has with susceptibles, the more likely he is to effectively transmit the disease. It then follows that the magnitude of the realized disease dispersion could be measured for example, in terms of the dispersion ability (i.e. vagility) of the infected individual. Accordingly, every infected individual will be expected to realize a certain virus (or micro-parasite) dispersion potential. Let the realized disease dispersion made by one infected individual be denoted by . Then, the number of successful transmission encounters per infectious individual can be modeled with a random variable . By writing , we are stressing the fact that the infection process is a function of the magnitude of the realized dispersion. Furthermore, we assume that the probability that an infected individual encounters and infects a susceptible individual given a realized change in dispersion is proportional to the previous number of successful infection encounters times a function of the number (or density) of the infected individuals in the population. Often [7], a non-linear function is chosen to account for factors such as crowding of infected individuals, multiple pathways to infection, stage of infection and its severity or protective measures taken by susceptible individuals. These assumptions allow us to specify a new infection event as the conditional probability(9)where is a non-negative function such that is a constant. Towards the end of this section we discuss possible functional forms for . We remark that if counts the number of successful transmission encounters of an infected individual that recently invaded a population consisting only of susceptible individuals, then the expected value of is in fact equal to the mean number of secondary infections in the context of the SIRS model. If the SEIR model dynamics is in place, then, when there is only one infected individual in the population, . Assuming that the probability that more than one successful infectious encounter occurs after an extra dispersion amount is negligible, then can be modeled using a simple homogeneous birth process where the quantity being born is the number of successful virus transmission encounters. The probabilistic law of this stochastic process is completely defined by the terms . To solve for these terms, first note that according to eq. (9)which leads toIn the limit when , the above equation leads in turn to the following system of differential equations:Then, it is well known [38] that solving this system of equations leads to(10)(11)Furthermore, approximating using a Taylor series expansion around leads to specific quantitative definitions of the stochastic process . For example, if or if , the one-step transition probability mass function (pmf) of adopts the negative binomial and Poisson forms respectively [37]. The Negative Binomial transition pmf would bring into the picture over-dispersion (higher variance to mean ratio) as a key qualitative property of the moments of the pure birth process describing the evolution of the number of successful transmission encounters. In any case however, the probability that one infected individual successfully passes on the infection isThis expression is readily interpretable: for a fixed value of , the probability of successfully passing on the infection converges to as the product grows large. Therefore, in this expression we are recovering the model property that the probability of successfully passing on the infection increases with the realized disease dispersion effort . Each individual's realized dispersion is in turn related to the individual's ‘effort’ to transmit the infection. In a given population, the magnitude of the realized disease dispersion for each infected individual can be expected to vary widely. Accounting for this demographic source of heterogeneity could be achieved by assuming that each individual's dispersion ability is drawn from a given probability distribution. That is, we would be modeling the variation in disease dispersion per infected individual with a random variable whose pdf has support on . Without loss of generality, here we model randomness in the product instead of just in the realized disease dispersion . Then, the probability that an infected individual chosen at random from the population realizes more than one successful secondary infection is found by averaging over all the possible realizations of . That is,A suitable probabilistic model for with empirical and theoretical support can be difficult to find (see for instance the models in [39]). A flexible positive, continuous distribution such as the gamma distribution could therefore be used. Here, we assume that the magnitude of the disease dispersion brought about by an infected individual is distributed according to a special case of the gamma pdf, the exponential distribution. Accordingly, letting we get that the probability of successfully transmitting
the infection is(12)As mentioned before, various biological hypotheses pertaining the behavior of the transmission as a function of the abundance of infected individuals have been advanced to justify various functional forms of . Suitable candidates for should satisfy the conditionsThese conditions guarantee the basic requirement that the probability of a new infective encounter (eq. 9) is null in the absence of infected individuals and that the overall chance that a new infection occurs increases proportionally with when is small. Furthermore, if such proportionality decreases in magnitude as grows large (that is, is concave down). Consider the following two functional forms:
Many other functional forms for the incidence rate could be derived using the above arguments. If for instance other heavy-tailed distributions are used instead of the exponential distribution, other incidence rate functional forms will arise and this could certainly be the topic of further research. However, in this work we limit ourselves to the exploration of the reaches of using the LHD model because it explicitly incorporates heterogeneity in transmission potential, because of its bi-stability properties (see “qualitative analysis of the SIRS models” section) and to formally test if it arises as a better explanation for bi-annual epidemic patterns using data from different localities and diseases. Thus, from this point on, in this work we will only consider the LHD incidence rate function and the classical incidence rate . In his seminal paper, Hethcote [1] also mentions that the LHD general incidence rate function could be eventually coupled with any seasonal forcing function. Motivated by this comment, in the results section we explore the reaches of doing so.
The two different SIRS models were fitted to time-course data of reported cases of syncytial virus infections. The data come from Gambia and Finland (Figure 1). Two ML formulations were used. The first one consisted of a Poisson likelihood that only required the available observed counts of infected individuals (eq. 13). The second formulation consisted of the joint likelihood of the counts and of the observed weather covariate and thus used information present on the time series of reported cases and on the corresponding time series of mean monthly temperature range for both locations. The ML estimates according to the first formulation for each model and data set combination are displayed in Table 1. Both information criteria used indicate that for Finland, the best model was the SIRS model with LHD incidence rate function. For Gambia, both information criteria for the SIRS model with classic incidence rate function are lower by three points approximately. This implies that given the data and the two information criteria ways of penalizing the likelihood score, both models are nearly indistinguishable for any practical purpose [49]. In Gambia, the extra parameter introduced by the LHD model is penalized: given the data set at hand, incorporating one extra parameter does not lead to a clear improvement In Figure 2 we plotted the model predicted number of infected individuals versus the observed values for the classical and the LHD SIRS model respectively. Note that, even though the best model is deterministic, the dynamics displayed by the data (small epidemics followed by a big epidemic peak) is very well recapitulated by the predicted solutions.
The results of the second ML formulation are qualitatively identical to the results with the Poisson likelihood (see table in the Text S1). For Finland, the BIC statistic for the classical model was 10376.2000 and for the LHD model 9893.5780. For Gambia, the BIC for the classical model was 729.1133 whereas the LHD model had a BIC of 733.2750. Hence, here again, for Finland the LHD is the best model whereas for Gambia, the classic model is better. Because the BIC can be used only to compare models for which the numerical values of the dependent variable are identical for all estimates being compared, it cannot be used to select between the two ML formulations. Indeed, in the second likelihood formulation the data fitted consist not only of the time series of infected counts but also of the monthly temperature range, thus it uses twice as much data for parameter estimation. Zeng et al [55] mention that an indication of which likelihood formulation is better can be obtained by comparing the per datum BIC score. Take for instance the BIC for the LHD model for Finland, 9893.5780. Dividing that BIC by the total number of data points used (, we get a per datum BIC of 48.4979. Now, the BIC for the LHD model for the Poisson likelihood formulation is (Table 1) 3196.9330. Dividing that number by the number of data points used () we get 31.34248. Thus, the Poisson likelihood formulation yields a better per datum BIC for Finland. For Gambia, the Poisson likelihood formulation seems to be better than the Poisson-Normal sampling model: for the classic model with Poisson likelihood this statistic is , whereas for the classic model with Poisson-Normal likelihood it is .
The SEIR model with classic and LHD incidence rate were fitted to measles time series data from London and Birmingham. In both cities, the SEIR-LHD model was selected as best (see Table 2). Notably, the difference in AIC and BIC is at least 2000 points in each case. The predictions for each model and city combination are shown in Figure 3. We remark that assessing and comparing the quality of the model predictions visually may be misleading. Indeed, according to our likelihood formulation, the parameter estimation process does not weight equally a deviation from the model prediction at low and high infected counts. In fact, the variance of the Poisson sampling error varies according to the mean predictions .
In this section we discuss the differences in the qualitative behavior of the SIRS model (1)–(3) with both classical and LHD incidence rates with and without seasonal forcing. We refer the interested reader to the Text S1 for proofs of the following claims. By construction, the set is a positively invariant set of the SIRS model (1)–(3). If we set the coefficients constant, the Dulac criterion guarantees that the SIRS model with neither the classic nor the LHD incidence rate function has periodic solutions in . Regarding the classical incidence rate, the SIRS model has two stationary solutions: a disease free equilibrium () and an endemic equilibrium (). It is well known that is a threshold for this model: If the disease remains endemic, while implies that the disease dies out. On the other hand, the SIRS model with LHD incidence rate has one disease free equilibrium and two endemic equilibria and . The is unconditionally a local attractor. However, only one of the endemic equilibria denoted as , lies inside the positively invariant set . If the endemic point is locally an attractor. Thus, when the LHD model exhibits bi-stability.
Introducing seasonal forcing has the following effects on the SIRS dynamics with classic incidence: first, it is well known that by letting the contact rate to be a periodic function of the form (8) where is small, the SIRS model with classical incidence rate has a periodic solution with period . This behavior is shown in Figure 4 A and B. Also, when seasonal forcing is introduced, the basic reproductive number becomes a periodic function of time, , that oscillates between the values and . The endemic point also becomes a periodic function of time that bounces back and forth between two extreme points, and . The expressions for and are given in the Text S1. The associated limit cycle of the model's solution inherits the stability behavior of the endemic point: if , then the limit cycle is asymptotically stable. A stable limit cycle is displayed in Figure 4 C. Because the function can cross the boundary of periodically depending on the value of , the dynamic behavior of the model's trajectory with respect to the nature of the endemic point (stable/unstable) can be described with a race analogy: The model's solution can be thought of as a hopeless ‘pursuer’ engaged in a race against the endemic solution who plays the role of the fast ‘leader’ that cannot be caught upon. Just as in a cycling race, as soon as the leader changes its strategy, so does the pursuer behind the leader. In that way, if is such that and only while , the leader () is deemed as stable and the solution's trajectory pursues the endemic point . As soon as becomes less than , the leader ‘changes its strategy’ and is deemed unstable whereas the becomes stable. At that moment, the trajectory switches its objective and pursues the and keeps doing so while . That sudden change of objective gives rise to a period doubling bifurcation of the limit cycle as seen in Figure 4 D. This change of objective (period doubling bifurcation) happens as grows large. We remark that at least one route to chaos in the associated Poincaré map of this model when is taken as the bifurcation parameter has been shown [53], [56], [57].
Finally, in the SIRS model with LHD incidence rate (see Figure 5 A and B), if we let the contact rate to be a periodic function of the form (8), a limit cycle also arises (see Figure 5 C). Here again, as increases, the trajectory engages in the same pursuer/leader dynamics and the limit cycle undergoes a period doubling bifurcation (Figure 5 D). However, contrary to what happens in the classical SIRS model with seasonal forcing, periodicity or extinction of the epidemics depends also on the initial conditions: if the initial proportion of infected individuals is too high, the disease will die from a subsequent depletion of the susceptible pool of individuals. Only if the epidemic begins with a small number of individuals will it slowly work its way up and attain a persisting limit cycle.
Multiple lines of evidence show that the forced SIRS and SEIR models with LHD incidence rate function constitute a better explanation of the seasonal epidemic patterns than the corresponding classical models with seasonal forcing, for the data sets and cases explored here. The first line of evidence is statistical in nature: when confronted with different time series of seasonal epidemics, the LHD model was selected as best in three out of four cases and in the fourth case, the LHD model was nearly indistinguishable from the classic model. By formulating the fitting and the model selection problems using likelihood-based inference and information theoretic model selection criteria we were able to conclude that given the data and the models at hand our model embodies the most likely explanation of how the observed data arose. Our model's nonlinear incidence rate takes into account heterogeneity in the ability to transmit the infection while modeling the infectious process as a pure birth stochastic process and hence, it is a more realistic model formulation. This new level of model complexity was achieved by incorporating only one extra parameter. The emphasis we give to a first principles derivation that hinges on interpretability and simplicity is not always sought in other SIR-type model formulations and modeling exercises [6], [17], [18], [24], [58]. Hence, our results show that a careful exploration of other incidence rate functions before resorting to mathematically more complex, high-dimensional models may bring new insights into the current understanding of the functioning of epidemics.
Another line of evidence in favor of the LHD model comes from its qualitative predictions. The classical SIRS model without the seasonal forcing predicts somewhat artificially that regardless of the initial proportion of infected and susceptible individuals, provided , the endemic equilibrium will be reached [28]. On the other hand, the LHD model without seasonal forcing predicts that the disease-free equilibrium is always an attractor, thus exhibiting bi-stability (see qualitative analysis section). Hence, if the initial proportion of infected individuals is too high, the disease will die from a subsequent depletion of the susceptible pool of individuals, contrary to what the classical model predicts. For the disease to persist in the population, the initial proportion of infected individuals has to be very low. Only then the infection process will proceed steadily to the endemic solution. This qualitative prediction matches the virus transmission strategy that the syncytial virus seems to have evolved: recall that in our model the extra parameter is the density of infected individuals at which the probability of successfully transmitting the infection is . In every locality, the ML estimates of were in the order of to , thus indicating that a very low density of infected individuals is needed in order to maximize the transmission rate of the measles and RSV diseases.
Incorporating weather covariates to our nonlinear SIRS model further improves the biological insights that can be concluded from the parameter estimation and model analysis exercises. A simple look at the strong auto-covariation patterns and at the pure weather trends, in particular for Gambia (Figure 1) indicate that modeling weather and weather effects with a sinusoidal function seems a natural add-on to the classic SIRS model, for this data set. For Gambia, the fact that the per datum BIC for the LHD model with the joint Poisson-Normal likelihood is very similar to the per datum BIC for the classic model indicates that the weather can indeed be viewed as a simple rotation and translation (eq. 15) of the weather effects (eq. 8). Thus eq. 15 may not always be viewed only as a phenomenological artifact [18]. For Finland, however, this was not the case. The per datum BIC favors much more clearly the Poisson likelihood formulation. Hence, we consider that in Finland the weather effects model (eq.8) would be better expressed as some unknown nonlinear transformation of the weather. In other words, in this country with more extreme weather, a change in the temperature range of a certain size is not translated as an equivalent change in the weather effects in the transmission rate. Also embedded within our weather effects model formulation (eq.8) is the hypothesis that weather affects incidence rates in a nonlinear fashion. In particular, when the strength of seasonality is high enough, the limit cycles predicted by both weather forced models undergo a period doubling bifurcation such that relatively small epidemic outbreaks are followed by big ones. Notably, these effects of the strength of seasonality were detected in Finland, the locality where the amplitude of the relative weather oscillation is larger.
The model selection exercise should by no means be the ending point of the analysis. Instead, if appropriateness of one model vs. the other cannot be resolved, a near-tie in a model selection situation should lead to the search and reformulation of each model's scientific predictions in a way that can be clearly tested in further experiments. Hence, the model selection results presented here should be rather viewed as the starting point of further analyses (see [59]).
Even for simple deterministic models, parameter estimation for dynamic data can be non-trivial. Dynamic models often present multimodal likelihoods thus complicating the parameter estimation process [42]. In these cases, the type of inferences possible is limited due to the presence of wide confidence sets that include parameter values with different qualitative predictions. If for instance the ML estimate of a bifurcation parameter is in a 2 limit-cycles region but its confidence interval includes parameter values for which these cycles do not appear, then there is not enough evidence in the data at hand to properly infer something about the size of the parameter of interest and hence, about the dynamic properties displayed by the data. In our case however, the precision of our parameter estimates and in particular, of the bifurcating parameter (Tables 1 and 2) is enough to identify the bifurcation region where the strength of seasonality lies for the data at hand.
Although in the two models studied here a period doubling bifurcation appears in the limit cycle, the LHD incidence rate model still provides very different qualitative predictions. In the classical model, the value of the basic reproduction number as a function of time acts as a stability switch for the DFE, so that any trajectory that begins with biologically realistic initial conditions will eventually enter the limit cycle. This is not the case for the LHD model, for which the periodicity or extinction of the epidemics depends very naturally on the initial conditions. Other studies have incorporated seasonal forcing in SIRS-type models [17], [60], [61], but since all have used the classical incidence rate function, they constrain their disease persistence and epidemics predictions to whether the basic reproductive number can or cannot be periodically above 1.
Nonlinear incidence rate forms derived from first principles constitute a promising starting point to review the interaction between demographic and environmental stochasticity and nonlinear seasonal effects. Indeed, recent studies have considered including in the classic SIRS model stochasticity in the seasonal process, besides sampling and/or observation error [17]. After showing that a simple pure observation error fit of our LHD model brings about a considerable fit improvement, we explored the qualitative differences between the models by coupling the deterministic skeletons with environmental noise. In Figure 6, the depicted stochastic trajectories show that in the classical model increasing the environmental noise results in transient visits to the disease free equilibrium stable submanifold (panel c)), whereas in the LHD model, with a large enough perturbation the trajectory visits the disease free equilibrium basin of attraction and remains there. Hence, the fact that regardless of the value of the basic reproduction number the DFE is always an attractor opens the door to stochastic phenomena whereby the trajectory exits the endemic solution basin of attraction and hits just by chance the DFE basin of attraction, only when the LHD incidence rate is used. By the same token, the trajectory periodically wanders in the direction of the DFE stable submanifold (similar to the “saddle fly-by” reported by Cushing et al [54]).
The results presented here are not by any means an exhaustive exploration of the interplay between nonlinear dynamics and stochasticity, both critical factors shaping seasonal epidemic patterns. However, our results may be viewed as the starting point of multiple research avenues. Three such research topics could be: first-principles derivation of non-linear incidence rate functions, the role of bi-stability and demographic stochasticity for disease persistence and the simulation of environmental and demographic stochasticity in the Poincaré map.
|
10.1371/journal.pbio.0050076 | Superfamily Assignments for the Yeast Proteome through Integration of Structure Prediction with the Gene Ontology | Saccharomyces cerevisiae is one of the best-studied model organisms, yet the three-dimensional structure and molecular function of many yeast proteins remain unknown. Yeast proteins were parsed into 14,934 domains, and those lacking sequence similarity to proteins of known structure were folded using the Rosetta de novo structure prediction method on the World Community Grid. This structural data was integrated with process, component, and function annotations from the Saccharomyces Genome Database to assign yeast protein domains to SCOP superfamilies using a simple Bayesian approach. We have predicted the structure of 3,338 putative domains and assigned SCOP superfamily annotations to 581 of them. We have also assigned structural annotations to 7,094 predicted domains based on fold recognition and homology modeling methods. The domain predictions and structural information are available in an online database at http://rd.plos.org/10.1371_journal.pbio.0050076_01.
| The three-dimensional structure of a protein can reveal much about that protein's evolutionary relationships and functions. Such information about all the proteins in an organism—the proteome—would offer a more global view of these relationships, but solving each structure individually would be a formidable task. In this study, we have parsed all Saccharomyces cerevisiae proteins into nearly 15,000 distinct domains and then used de novo structure prediction methods together with worldwide distributed computing to predict structures for all domains lacking sequence similarity to proteins of known structure. To overcome the uncertainties in de novo structure prediction, we combined these predictions with data on the biological process, function, and localization of the proteins from previous experimental studies to assign the domains to families of evolutionarily related proteins. Our genome-wide domain predictions and superfamily assignments provide the basis for the generation of experimentally testable hypotheses about the mechanism of action for a large number of yeast proteins.
| The yeast Saccharomyces cerevisiae is one of the most widely studied organisms, yet a large fraction of its proteins are of unknown structure and/or unknown function. Knowledge of the structure of a protein is critical to understand how it functions, and hence, a complete set of protein structures for yeast is desirable, but difficult to accomplish experimentally.
The accuracy of de novo structure prediction methods, although far from the accuracy of experimental structures, has improved in recent years. The Rosetta de novo structure prediction method [1–4] is currently one of the best methods available for predicting the structure of proteins lacking obvious homology to known structures [5–8]. Application of Rosetta to genome-wide annotation has been limited by the difficulty of distinguishing accurate from inaccurate predictions and the computational cost associated with scaling the procedure to whole genomes. Initial results have been encouraging, showing promise on subsets of protein families and prokaryotic genomes [9,10]. We have previously [11] predicted structures for short Pfam families without structural information, and showed that a simple confidence function could partially separate correct structure predictions from incorrect predictions.
There is a rich body of work on the relationship between superfamily (encoded in databases such as SCOP [12–14] and CATH [15]) and function (encoded in databases such as Kyoto Encyclopedia of Genes and Genomes [KEGG] [16] and Gene Ontology [GO] [17]). Although many superfamilies have been shown to carry out multiple functions, Hegyi and Gerstein [18] found that the majority of structure superfamilies carry out one or a few molecular functions, and conversely, that the majority of functions are carried out by one or a few SCOP superfamilies. This relationship can be exploited when predicting to which structure superfamily a protein belongs [9].
We describe an integrated approach for assigning protein domains to structure superfamilies that combines de novo structure predictions with GO function, process, and component annotations. We first parse all yeast proteins into putative structural domains using the Ginzu method [7,19]. Ginzu predicts domain boundaries by applying a hierarchy of sequence-based methods beginning with searching for homologs of known structure using PSI-BLAST [20] and ending by parsing block patterns in multiple sequence alignments (MSAs). After running Ginzu on the full proteome, we applied the Rosetta structure prediction method to domains shorter than 150 amino acids for which no homolog of known structure was found. The top structure predictions were compared to protein domains of known structure using the MAMMOTH protein structure comparison program [21]. The reliability of an assignment to a protein structure superfamily derived from these structure comparisons was evaluated using a logistic regression–based confidence function optimized on a large training set of Rosetta models for proteins of known structure. Superfamily predictions of increased accuracy were obtained by integrating GO function, component, and process annotations [17,22] from the Saccharomyces Genome Database [23] with the structure prediction data using a simple Bayesian approach. We predicted structures for 3,338 domains and have annotated 581 of them with novel SCOP superfamily assignments. The domain predictions, the predicted structures, and superfamily assignments are accessible at http://rd.plos.org/10.1371_journal.pbio.0050076_01.
A total of 6,238 open reading frames (ORFs) were parsed into structural domains using Ginzu [7,19]. Ginzu was used successfully in Critical Assessment of Techniques for Protein Structure Prediction 6 (CASP6) to delineate domains within query proteins by sequentially searching for (1) sequence-detectable homology to the Protein Data Bank (PDB) using PSI-BLAST [20], (2) more-remote fold recognition hits to PDB structures [24,25], (3) hits to Pfam-conserved sequence family domains [26,27], and (4) block patterns in MSAs. This hierarchical application of methods is organized so that methods providing more reliable information are applied first, thus accuracy is not sacrificed as we apply multiple methods in an attempt to maximize comprehensive coverage of the genome. A total of 14,934 domains were predicted, of which 38% had a sequence-detectable homolog of known structure, and an additional 9% could confidently be annotated by fold recognition methods. A summary of the genome-wide domain parses is presented in Table 1, and a complete list of domain predictions are presented in Table S1.
Although the confident fold recognition results generated as part of this study are not the main focus of this paper, they provide a wealth of information on proteins for which there are no detectable sequence homologs of known structure. The results for 1,361 domain annotations using fold recognition are detailed in Table S1 and are available at http://rd.plos.org/10.1371_journal.pbio.0050076_01.
A total of 4,006 yeast protein domains shorter than 150 amino acids (a practical length limit for the Rosetta method) and not linked to known structures by PSI-BLAST or fold recognition methods were identified by Ginzu: 668 of these contained predicted transmembrane helices and were omitted; the remaining 3,338 domains were folded using the Rosetta de novo method. Ten thousand structure models were generated for each of these remaining 3,338 domains using the Rosetta de novo method [1,2,28] and then condensed to 30 representative models by clustering. The size of the calculation is significant and is estimated at 12 million CPU hours, or 1,350 CPU years. This calculation was performed on the World Community Grid (WCG) parallel grid computing facility provided by IBM (http://wcgrid.org).
The 30 representative models for each domain were compared to a database of experimentally determined protein structure domains (based on ASTRAL; see Materials and Methods) with representatives from all SCOP (version 1.67) superfamilies and evaluated using a confidence function (referred to as the MAMMOTH Confidence Metric [MCM]) described below. The confidence of a given prediction for a given protein-domain is estimated based on features resulting from the Rosetta structure prediction, clustering, and structure–structure matching steps (using MAMMOTH [21]). The primary improvement in the confidence function over our previous work [11] is the inclusion of the contact order (CO; average sequence separation of contacting amino acids [29]) of the residues superimposed in the MAMMOTH structure–structure alignment of the predicted structure with the matched structure; this CO term penalizes less-significant matches dominated by local contacts such as single long alpha helices. Figure 1 describes the performance of the MCM on a large benchmark set developed for this study (see Materials and Methods). The MCM score, PMCM, ranges between 0 and 1 and is an estimate of the probability of the identification of the correct structure superfamily identification. A total of 404 domains in the yeast dataset (see Table 2) have a PMCM above 0.8 (considered significant for the purpose of this discussion) and can be found in Table S2 (additional domains are annotated with superfamily via integration with GO as described below).
The confidence estimates derived from our SCOP benchmark set are likely to be somewhat inflated when applied to the yeast protein set for two reasons; first, as discussed in the following section, the domain boundaries are derived directly from experimental structures in our SCOP benchmark, but are subject to error for the yeast proteins, and second, in the SCOP benchmark set, there is by construction always at least one closely related structure in the correct superfamily, whereas proteins with novel folds in yeast may not belong to any pre-existing superfamily. Below and in Materials and Methods, we describe tests on two additional validation sets that include the above sources of error (and thus allow for the estimation of the effects of such errors on structure superfamily prediction). Although there is a non-negligible presence of errors in domain parsing and superfamily assignment, our results show that the superfamily assignments generated herein (see Table S2) should be valuable for stimulating the generation of experimentally testable hypotheses about the structure and often the mechanism of action of these proteins.
There is a strong relationship between the function of a protein and its structural superfamily [18]. Most commonly, proteins in the same superfamily carry out one or a few functions. The reverse is also true; often only one or a few superfamilies are found to carry out a specific function. We derived probability distributions, P(GO|SF), that relate SCOP superfamily (SF) to molecular function, biological process, and cellular component (GO). We also constructed probability distributions, P(SF|D), that give the probability of a given superfamily, given the predicted structures (D), that is derived from the distributions of PMCM for a target, as described in Materials and Methods. These distributions were integrated to determine the degree to which a superfamily prediction is simultaneously compatible with the structure predictions and the functional annotation available for a given protein, using:
where P(SF|D,GO) is the probability that the domain belongs to SCOP superfamily SF, given the predicted structures, D, and the GO terms, GO, for the protein. The independence assumption underlying Equation 1 is described in Materials and Methods.
The superfamily distributions derived from the structure prediction data alone (P(SF|D)), the GO annotations (P(SF|GO)), and from the two together (P(SF|D,GO)), are compared in Figure 2 for four proteins for which the true SCOP superfamilies are known, showing the synergy between the two sources of information. The ambiguities in P(SF|D) (red line) and P(SF|GO) (blue line) are reduced upon integration P(SF|D,GO) (black line), resulting in less ambiguous predictions for many difficult-to-annotate domains. The overall performance for the P(SF|D,GO) over the benchmark set (see Materials and Methods) is shown in Figure 3. A total of 177 yeast domains (see Table 2) were assigned a structural superfamily with a P(SF|D,GO) over 0.8 (Table S3).
True performance of these technologies cannot be assessed on the benchmark dataset because the domain boundaries of this set are perfect (derived from known structures in the ASTRAL database). A subset of the proteins without links to known structure at the start of this project now have strong homology to a structure that has since been solved, see Figure 4 for examples. These recently solved structures give us an opportunity to assess the performance of our technology without bias in the selection of the proteins, with real domain prediction error incorporated, and without the contamination of the results by weak homology to known structures. Twenty-seven domains from this project now have a homolog of known structure (see Materials and Methods for homology definition) and three of the 11 predictions with a PMCM of 0.8 or higher are correct. Five of seven proteins from this set with a P(SF|D,GO) ≥ 0.8 are correct; five of which (three correct) also have a PMCM of 0.8. The small sample size represented by this dataset makes it difficult to assess accurate upper and lower bounds of the estimated error. We have also generated a much larger dataset (see Materials and Methods), as part of the Human Proteome Folding Project (HPF). This dataset, containing proteins from over 150 organisms, was derived without the use of fold recognition (and thus is not identical to the protocol for yeast), but provides valuable information as to the effect of domain prediction on our procedure. A total of 44% of the 207 predictions that were made for recently solved structures with PMCM above 0.8 in this dataset were correct; 84% of the 51 predictions with P(SF|D,GO) above 0.8 were correct; and 31 of these predictions had both a PMCM and a P(SF|D,GO) above 0.8, and 27 of these were correct. Over all three validation sets, more than 40% of the predictions with a PMCM above 0.8, and more than 75% of the predictions with a P(SF|D,GO) above 0.8, are correct, illustrating the value of data integration in this work.
Importantly, we were able to use these sets of recently solved proteins to better characterize the errors associated with different confidence Ginzu domain predictions. We found that a subset of the incorrect domain parses which significantly diminish the chances of correctly predicting fold and function are easily removed using a simple filter (described in Materials and Methods). This domain-prediction filter allows us to recover more-accurate predictions for multi-domain proteins. We were able to classify 50% of the amino acids from the 6,238 attempted ORFs to SCOP superfamilies which is significantly higher than the 35% coverage achieved by a sequence-based hidden Markov model approach [30].
In this section, we discuss several protein complexes with components assigned to superfamilies by both GO-integration and MCM approaches. These predictions and the much larger set of predictions in the database accompanying this paper provide a basis for hypothesis generation and experimental testing, but it must be borne in mind that there is a significant probability that any single prediction is incorrect, as indicated by our estimates of error.
The mediator complex, a large complex containing 24 polypeptides [31], has been shown to be required for transcriptional activation in many eukaryotic organisms and play key roles in transmitting regulatory information to the pre-initiation complex. During transcriptional initiation, it interacts with the RNA polymerase II holoenzyme and the promoter region. The role of the mediator complex in transcriptional regulation, and the complete makeup of this complex and its dynamic composition throughout different cell and developmental states (in response to specific regulators) are active areas of research. To date, several studies have explored the overall makeup of the complex by probing protein–protein interactions [31] and by electron microscopy of purified mediator complex, but to our knowledge, this complex has eluded higher resolution methods such as crystallographic analysis. Although there exists an extensive body of work on the overall function of this complex, the roles, positions, and structures of most of the individual polypeptide components remain undetermined.
We find confident superfamily predictions for several proteins within this complex that were not structurally annotated prior to this work. Table 4 outlines these predictions, as well as their sources and confidence estimates. Several proteins in the Mediator head domain are predicted to contain DNA-binding domains. In addition, multiple head domain proteins are predicted to be long helical bundles, potentially serving as scaffolds. ROX3 contains two predicted domains, see Figure 5A. The first domain is predicted to belong to the Homeodomain-like superfamily (PMCM = 0.43; GO-term: transcription from RNA polymerase II promoter; P(SF|D,GO) = 0.83); implying DNA binding. MED4 (Figure 5B), in the middle region of the mediator complex, is also a two-domain protein with a N-terminal homeodomain-like superfamily assignment (PMCM = 0.65; GO-term: transcription from RNA polymerase II promoter; P(SF|D,GO) = 0.98). Several superfamily predictions for this complex (such as hits to superfamilies like spectrin-like and Rossmann folds) are difficult to interpret unambiguously due to the large number of functions compatible with each of these superfamilies, but are not incompatible with DNA-binding functions. Gal11 (Figure 5C) contains a diverse mix of predicted domain structures: for the first domain, we find a PSI-BLAST match to the motor domain of myosin, and the second domain shows a strong fold recognition hit to a structural domain from sec24 (a component of the secretion system). Rosetta models for the third domain match the spectrin-like superfamily, and the models of the fourth domain match the DNA-binding lambda-repressor–like fold (with a competing hit to the tRNA-binding fold). The fifth domain shows a match to the RNA pol II–like fold (RPB1) by confident fold recognition. Although these domain and structure predictions are insufficient by themselves to localize specific molecular function to components of the mediator complex, it is encouraging that we can make some headway in localizing specific superfamilies and functions to components of this large complex.
TIF35 (Figure 5D) is a subunit of the translation initiation factor complex, a complex essential for translation [32]. We predict three separate domains exist within this protein. The middle and C-terminal domains of this protein are both strongly predicted to belong to the RNA-binding domain, RDB superfamily (d.58.7; identified by PSI-BLAST). The N-terminal 65 amino acids lack sequence-detectable homologs of known structure, but the Rosetta-generated models and GO-selected structure prediction for this protein shows a strong match to the Translation proteins SH3-like domain (PMCM = 0.44; GO-term: translation initiation factor activity; P(SF|D,GO) = 0.92).
The mitochondrial ribosome, or the mitoribosome, shares a number of protein components with bacterial ribosomes, but it is believed that the mitoribosomes have comparatively more proteins than their bacterial counterparts; many of the proteins associated with the mitoribosome have no detectable sequence similarity to other mitochondrial proteins [33]. We have predicted the structure for two components known to be associated with the mitoribosome [34,35]. MRPL37 (Figure 5E) is predicted to belong to the Ribosomal protein L6 superfamily (PMCM = 0.31; GO-term structural constituent of ribosome; P(SF|D,GO) = 0.86), a superfamily involved in RNA binding. We predict that MRPL44 (Figure 5F) belongs to the dsRNA-binding domain–like superfamily (PMCM = 0.78; GO-term: structural constituent of ribosome; P(SF|D,GO) = 0.86). Overall, the structure predictions for these mitoribosome proteins suggest that they belong to superfamilies compatible with known, although highly diverged, components of both the bacterial and eukaryotic ribosome.
INH1 (Figure 5G) is an ATPase inhibitor predicted to be a member of the ARM repeat superfamily (PMCM = 0.73; GO-term: ATP synthesis couple protein transport; P(SF|D,GO) = 0.82). Inh1 dimerizes and binds to the F1 complex of the ATPase, thereby inhibiting its function [36,37]. Many members of the ARM repeat superfamily are involved in protein and peptide binding, which is consistent with both the dimerzation and the binding to the ATPase.
All data are accessible via the Yeast Resource Center (YRC) public data repository [38] at http://rd.plos.org/10.1371_journal.pbio.0050076_01. The data will also be made available in other formats upon request.
Comprehensive generation of three-dimensional structures with resolution or reliability of those determined by X-ray crystallography or nuclear magnetic resonance (NMR) is currently beyond the capabilities of any protein structure prediction method; these methods can, however, play an important role in generating structural annotations for whole genomes due to the much lower investment of resources required per protein domain. In this work, we have shown that it is possible to: (1) generate protein structure models on a genome-wide scale, (2) automate the assessment of the structure prediction quality, (3) convert the results into pre-existing encodings of structure in the form of SCOP superfamily classifications, and (4) augment the model-based assignment of SCOP superfamily by integrating with pre-existing function, process, and component information encoded in the GO database.
We were able to assign SCOP superfamilies to 7,094 of the 14,934 predicted domains in yeast using PSI-BLAST and fold recognition methodology. A total of 4,006 of the remaining 7,840 domains were short enough (less than 150 amino acids) for de novo structure prediction. Of these, 668 were omitted because they contained at least one predicted transmembrane helix. Low-resolution structure models were built for the remaining domains using Rosetta; of these, 404 were assigned to superfamilies with confidence using MCM, and an additional 177 were assigned with confidence after integrating with GO process, component, and function annotations.
A significant challenge in carrying out this work was the magnitude of the computation required for generating de novo structure predictions for large numbers of domains. Robust and fast methodology, efficient data storage, analysis tools, and data organization were required. Our use of distributed computing (http://wcgrid.org), innovative database architecture [39,40], and fully automatic methods were essential for this full-genome annotation. Yeast is particularly interesting because it is the focus of a vast global research effort. Future work will include an ongoing effort to scale this procedure to over 150 completely sequenced genomes as well as to employ recently developed higher resolution structure prediction methods [41] that produce more-accurate and reliable models, but require significantly greater computational resources per protein domain.
The information content in the predicted structures may be further leveraged by integration with other data such as global quantitative measurements of mRNA, protein expression levels, DNA–protein, and protein–protein interactions. Such datasets are available for yeast and several other organisms as part of ongoing functional genomics efforts, and integration of these data types with the predicted structures should contribute to the annotation of protein functions.
Two representative domains from each SCOP [12–14] superfamily were folded using the Rosetta de novo method [1,2,28]. Superfamilies without members shorter than 200 amino acids were excluded, as were proteins for which Rosetta failed to produce predictions within a reasonable time. One thousand models were generated for each domain. This resulted in structure predictions for 998 domains for which the structures have been experimentally determined. The predicted structures were clustered by root mean square deviation (RMSD), and the centers of the top 30 clusters were compared to a domain database generated from ASTRAL 1.67 (reduced to 40% sequence identity) [42,43] using a modified version of MAMMOTH [21] that calculates the contact order of the aligned regions of the predicted structure and the ASTRAL domain. An overview of the statistics is presented in Table 5, and a detailed description of the results in Table S4.
The MCM estimates the probability that the MAMMOTH match between predicted structure and the ASTRAL domain (see previous section) has identified the correct superfamily and is based on the closeness of match (MAMMOTH Z-score), the length of the two proteins involved, LAstral and Lpredicted, the CO of the region of the predicted structure that was superimposable on the experimental structure, and the degree to which Rosetta converged during the generation of the set of predicted conformations (converg below; estimated during the clustering step). The general formula for the confidence functions is given in Equation 2, and the weights of the parameters (a, b, c, d, and the constant C) for the three models described in the following paragraph are presented in Table 3.
This model is similar to that used in previous studies [11], with two improvements. First, we have fit three separate logistic regression models, one for all alpha proteins, one for all beta proteins, and one for alpha and beta proteins; the size of the benchmark set and the fact that we are fitting a small number of parameters allows for this trifurcation of the benchmark set. Second, we compute the CO [29] over the matched region. This penalizes the scenario in which small numbers of long secondary structure elements (usually helices) are aligned; the CO term as well as the length ratio corrects for the overly confident score we would otherwise calculate based on convergence and MAMMOTH Z-score alone. We used 5-fold cross-validation to fit each of the three secondary structure class–specific confidence functions. For selecting between the three models for a query protein, we use secondary structure content predicted by PsiPred [44]. The alpha model is used for proteins with over 15% predicted alpha-helical content and under 15% beta-sheet content. The beta model is used for protein with more than 15% predicted beta strand and less than 15% alpha helical. The alpha/beta model was used for all other domains.
Given a set of predicted structures D for a given protein, we estimate the probability the protein belongs to superfamily, SF, P(SF|D) as follows. Each superfamily is initially assigned a probability corresponding to the maximum PMCM value for that superfamily over the top five PMCM values for all predicted conformations for the query protein; probabilities less than 0.2 are set to zero. If the sum of the raw probabilities is greater than 0.8, they are scaled linearly so that the sum is 0.8. Because of the uncertainties of de novo structure prediction, these scaled probabilities, Pscaled(SF|D), are then linearly combined with the background superfamily distribution, P(SF) (Equation 3):
The final distributions, P(SF|D), are guaranteed to have non-zero probabilities for every superfamily, and to sum to 1. The background distribution P(SF) ensures that (1) we do not disregard useful functional information at the integration with GO stage and (2) that we do not over interpret the confidence values derived from the benchmark training set.
We obtain P(SF|D,GO) of a superfamily, SF, given both protein structure prediction, D, and GO annotations, GO, using Bayes' rule and the assumption that P(GO,D|SF) ∼ P(GO|SF)*P(D|SF):
We obtain P(D|SF) via Equation 5:
After substituting Equation 5 into Equation 4, both P(SF) and P(D) cancel. P(SF|D) is computed as described in the previous section, and P(GO|SF), P(GO), and P(SF) are computed from proteins in the PDB that are annotated with GO function, component, or process and also classified in SCOP. To deal with cases in which there is a single function annotation for a given superfamily, we allow for the possibility that the uniqueness of this mapping is due to under-sampling of superfamily space (as represented by the PDB) or function space (as represented by GO) by adding pseudo counts distributed according to the background superfamily distribution, Pastral95(SF), computed from ASTRAL 1.67 culled so that no sequences are more then 95% identical.
The parameter M (a regularization parameter controlling the relative contribution of our pseudo-counts) was estimated by carrying out function assignment given the superfamily over the benchmark set: we chose M to minimize the classification error estimated using 10-fold cross-validation. The overall procedure was relatively insensitive to the value of M ranging from one to ten with an optimal value of four. The P(SF|GO) are too diffuse for confident superfamily prediction from GO annotations alone, hence the integration with the structure prediction data is critical for accurate superfamily predictions.
Equation 6 relies on the assumption that the functional annotations are independent and mutually exclusive, which is not the case. (GO is a directed acyclic graph [DAG], with an implicit conditional dependence of lower nodes on parent nodes.) Nodes can have multiple parents, thus the probability of the child nodes of a more general term are not guaranteed to sum to the probability of the parent term. To circumvent this problem, we assigned the combined probability for each superfamily by taking the maximum probability for that superfamily given the predicted structures and all functions, i.e., Equation 7:
Finally, the sum of P(SF|D,GO) for any given protein domain is normalized to sum to one; thus confident assignments are not made when there are strong matches to more than one superfamily.
The performance of the MCM and the GO integration was evaluated on two independent datasets. The first dataset, from HPF project, consists of 768 predicted domains that now have a homolog with a known structure that is classified in SCOP 1.69. The homologs were identified by blasting predicted domains against all sequences from ASTRAL 1.69 and selecting those with a PSI-BLAST e-value less than 1 × 10−3. We also require that the shorter of the two sequences is more than 80% of the length of the longer one, and that 60% or more of the predicted domain is aligned with the ASTRAL domain. These domains are part of an ongoing project in which we predict structures for over 150 genomes; although domains with any homology to known structures are excluded, a number of structures have been solved and classified in SCOP during the 18 mo the project has been running. The scope of this separate project prohibited us from carrying out fold recognition calculations on these domains, and since domains that can be assigned using fold recognition methods will on average have higher MAMMOTH structural similarities to known structures than domains that cannot be assigned, results from this dataset represent an upper bound on performance on the dataset in this paper.
The second dataset was generated the same way the HPF set was generated, but limited to yeast domains. The proteins from which these domains are derived have been subjected to fold recognition and hence give a better estimate of the true performance. This dataset is, however, too small for statistically significant conclusions to be made.
Based on inspection of the results on the HPF dataset, domains from predicted two-domain proteins are excluded if both the domains are predicted using less-confident methods (MSA, unassigned, or Pfam domains), or if the domain under consideration is an MSA domain regardless of the neighboring domain type. A large fraction of these proteins have single domains, and correct superfamily matches are quite unlikely when models are only generated from domain fragments.
The generation of structure predictions was divided into three completely automated steps: pre-processing, production (the running of Rosetta), and post-processing (clustering, superfamily assignment, and function integration). The pre-processing protocol includes domain prediction, prediction of secondary structure, disordered regions, trans-membrane helices [45], and signal peptides [46], and the local structure fragments and other files necessary for running Rosetta. This step was conducted in-house on two 64-CPU Linux clusters. The production step, generating 10,000 structure predictions, was completed in collaboration with IBM running Rosetta on the World Community Grid as part of a larger effort, and is estimated to have used 12 million CPU hours, or 1,350 CPU years. The post-processing step was performed in-house (using the same hardware as the pre-processing step), and included clustering and superfamily assignment by MCM and GO integration. The resulting dataset is complex, and is stored, queried, organized, and analyzed using an open-source software package, 2DDB [39,40] of our own construction.
|
10.1371/journal.pntd.0004982 | Translation Regulation and RNA Granule Formation after Heat Shock of Procyclic Form Trypanosoma brucei: Many Heat-Induced mRNAs Are also Increased during Differentiation to Mammalian-Infective Forms | African trypanosome procyclic forms multiply in the midgut of tsetse flies, and are routinely cultured at 27°C. Heat shocks of 37°C and above result in general inhibition of translation, and severe heat shock (41°C) results in sequestration of mRNA in granules. The mRNAs that are bound by the zinc-finger protein ZC3H11, including those encoding refolding chaperones, escape heat-induced translation inhibition. At 27°C, ZC3H11 mRNA is predominantly present as an untranslated cytosolic messenger ribonucleoprotein particle, but after heat shocks of 37°C—41°C, the ZC3H11 mRNA moves into the polysomal fraction. To investigate the scope and specificities of heat-shock translational regulation and granule formation, we analysed the distributions of mRNAs on polysomes at 27°C and after 1 hour at 39°C, and the mRNA content of 41°C heat shock granules. We found that mRNAs that bind to ZC3H11 remained in polysomes at 39°C and were protected from sequestration in granules at 41°C. As previously seen for starvation stress granules, the mRNAs that encode ribosomal proteins were excluded from heat-shock granules. 70 mRNAs moved towards the polysomal fraction after the 39°C heat shock, and 260 increased in relative abundance. Surprisingly, many of these mRNAs are also increased when trypanosomes migrate to the tsetse salivary glands. It therefore seems possible that in the wild, temperature changes due to diurnal variations and periodic intake of warm blood might influence the efficiency with which procyclic forms develop into mammalian-infective forms.
| When trypanosomes are inside tsetse flies, they have to cope with temperature variations from below 20°C up to 37°C, due to diurnal variations and periodic intake of warm blood. In the laboratory, procyclic forms (the form that multiplies in the midgut), are routinely cultured at 27°C. When procyclic forms are heated to temperatures of 37°C and above, they decrease protein production, and at 41°C, mRNAs aggregate into granules. We show here that quite a large number of mRNAs are not included in granules and continue to be used for making proteins. Some of the proteins that continue to be made are needed in order to defend the cells against the effects of heat shock. Interestingly, however, a moderate heat shock stimulates expression of genes needed for the parasites to develop further into forms that can colonise the salivary glands. It thus seems possible that in the field, temperature variations might influence the efficiency with which of trypanosomes in tsetse flies become infective for mammals.
| African trypanosomes, like all other organisms investigated so far, respond to heat shock by repressing general protein synthesis, while enhancing or retaining synthesis of proteins that are required to survive or recover from heat stress [1]. Unlike other organisms, however, trypanosomes lack the ability to control the transcription of individual protein-coding genes [2–4]. Polymerase II transcription is polycistronic, and monocistronic mRNAs are created by 5' trans splicing of a capped spliced leader (SL) and polyadenylation [5]. The selectivity of the heat shock response, like other changes in gene expression, therefore relies on post-transcriptional mechanisms. Trypanosoma brucei procyclic forms are the forms that grow inside the tsetse fly midgut. In natural infections, these forms migrate to the proventriculus, developing into epimastigotes, and from there to the salivary glands where they become metacyclic forms which are infective for mammals [6]. After a tsetse fly bites a mammal, long slender bloodstream forms proliferate in the new host's blood and tissue fluids. Upon reaching high density, the parasites differentiate into non-dividing short stumpy forms [7], which are pre-adapted for differentiation into procyclic forms upon uptake by tsetse [8].
Nearly all previous work on heat shock in T. brucei has concentrated on cultured procyclic forms subjected to a one-hour heat shock at 41°C [1]. This is on the upper edge of temperatures that can be tolerated by most tsetse species in the wild [9], since tsetse prefer to rest in the shade and to feed on parts of animals that are not exposed to full sunlight [10]. Nevertheless, after the 41°C treatment trypanosomes recover quite rapidly upon return to the normal culture temperature of 27°C [1]. Heating to 41°C inhibits trypanosome transcription initiation [11,12] and stimulates overall mRNA degradation [1], resulting in gradual loss of total mRNA [1]. In addition, translation of most mRNAs is suppressed. After an hour at 41°C, there is almost no mRNA in polysomes, while three types of messenger ribonucleoproten (mRNP) granules appear. These granules contain most of the mRNA [13] and various combinations of translation factors, the two poly(A) binding proteins PABP1 and PABP2, the helicase DHH1, the aggregation-prone protein SCD6, and the 5'-3' exoribonuclease XRN1 [1,14,15]. Cycloheximide treatment causes retention of mRNAs in polysomes at 41°C, inhibiting both mRNA degradation and granule formation [1]. Thus, as in other organisms, granules are locations for storage and/or degradation of non-translated mRNAs.
Despite the general shut-down in gene expression after heat shock, synthesis of proteins that are required for survival during, and recovery after, heat shock—such as protein refolding chaperones—continues. We previously showed that the zinc-finger protein ZC3H11 binds to the 3'-UTRs of chaperone mRNAs, and is required both for target mRNA retention and for cellular survival after heat shock [16]. ZC3H11 binds to MKT1 and to PBP1, which in turn recruits LSM12 and poly(A) binding proteins PABP1 and PABP2[17]. MKT1 and PBP1 remain distributed throughout the cytosol after heat shock. Starvation also causes the formation of mRNP granules, but in this case MKT1 and PBP1 colocalise with SCD6 in the granules [17].
Recently, we investigated how ZC3H11 itself is regulated [13]. ZC3H11 is barely detectable in both bloodstream and procyclic forms grown at their normal culture temperatures of 37°C and 27°C respectively [16]. When procyclic forms are incubated at 37–41°C, the level of ZC3H11 protein progressively increases. This is partly caused by a loss of protein degradation, but more prominently by translational control. At 27°C, the ZC3H11 mRNA migrates in sucrose gradients as a messenger ribonucleoprotein particle at or just above the small ribosomal subunits, but after a 1h heat shock at 37°C, 39°C or 41°C, nearly all of the ZC3H11 mRNA is in the polysomal fractions and ZC3H11 mRNA does not colocalise with heat shock granules [13]. In this paper we have examined whether other mRNAs show similar translation regulation after a 39°C heat shock, and identified additional mRNAs that escape sequestration into stress granules after a 41°C heat shock.
Trypanosome culture conditions were as described in [18]. Procyclic trypanosomes were grown in MEM-Pros medium at 27°C (unless stated otherwise) at densities lower than 6×106 cells/ml. All experiments were done with Lister 427 monomorphic procyclic form parasites expressing the Tet repressor.
3–5×108 procyclic cells were treated with cycloheximide (100μg/ml) for 5 minutes, harvested at room temperature by centrifugation (850g, 8min, 20°C), washed once in 1ml of ice-cold PBS and lysed in 300μl of lysis buffer (20mM Tris pH7.5, 20mM KCl, 2mM MgCl2, 1mM DTT, 1200u RNasin (Promega), 10μg/ml leupeptin, 100μg/ml cycloheximide, 0.2% (vol/vol) IGEPAL) by passing 20–30 times through a 21G needle. After pelleting insoluble debris by centrifugation (17000g, 10min, 4°C) and adjusting to 120mM KCl, the clarified lysate was layered onto a 17.5–50% sucrose gradient (4ml) and centrifuged at 4°C for 2 hours at 40000 rpm in Beckman SW60 rotor. Monitoring of absorbance profiles at 254nm and gradients fractionation was done with a Teledyne Isco Foxy Jr. system. RNAs from pooled fractions were purified using TriFast. To control for the efficiency of RNA isolation, equal amounts of a human β-globin in vitro transcript were sometimes added to each of the collected fractions before RNA purification.
Proteins were detected by Western blotting according to standard protocols. For detection of the endogenous ZC3H11 protein only cytoskeleton-free extracts were used. Antibodies used were to the ZC3H11 (rabbit, 1:10000, [13]), RBP6 [19] and PTP1 [20]. Detection was done using ECL solutions (GE Healthcare).
Granules from normal and heat-shocked procyclic cells were enriched as described previously [21]. 5×108 control or heat-shocked (1 hour at 41°C) procyclic cells were harvested at room temperature by centrifugation (1500g, 10min), washed in 1ml of PBS and lysed in 200μl of ice-cold buffer A (20mM Tris-HCl pH 7.6, 2mM MgCl2; 0.25M sucrose, 1mM DTT, 10% glycerol, 1% Triton X-100, 800u RNasin (Promega), 1 tablet Complete Protease Inhibitor Cocktail EDTA free (Roche)/10ml buffer) by pipetting. Lysis was confirmed microscopically. The lysate was clarified (20000g, 10min) and the supernatant (SN1) was transferred to fresh tube with 750μl of peqGOLD TriFast FL (Peqlab). All remaining supernatant was removed after one short centrifugation (3min, 20000g). The pellet was resuspended again in 200μl of buffer A by passing 30–40 times through a 21G syringe, vortexed and centrifuged (20000g, 5min). The supernatant (SN2) was taken and the pellet was resuspended in 200μl buffer A as above. The whole procedure was repeated one more time to obtain the supernatant SN3. Then the pellet was resuspended one more time in 200μl buffer A as above and microtubules were disrupted by the addition of 12 μl 5M NaCl (283mM final conc.), the samples were passed through 21G syringe, incubated on ice for 30 minutes with vortexing every 5 minutes, then centrifuged (20000g, 10min). The supernatant (SG) was removed and used to prepare the "small granule" RNA (SG). The pellet was washed once in 200μl of buffer A without resuspension (20000g, 10min) and finally resuspended in 750μl of TriFast FL to make the "large granule" (LG) RNA. Another 5×107 control or heat-shocked procyclic cells were taken to obtain total RNA.
Total RNA was incubated with oligonucleotides complementary to trypanosome rRNA and RNAse H, and mRNA integrity was checked by Northern blotting with a probe that detects the beta-tubulin mRNA. The samples were then subjected to high throughput sequencing such that most samples gave about 30 million aligned reads. Sequences were aligned to the latest available T. brucei TREU927 genome sequence using Bowtie [22], allowing for up to 20 sequence matches. Reads that aligned to open reading frames were then aligned using a custom script, again allowing for each read to align up to 20 times. To extract the reads for individual open frames, we used a modified version of the "unique open reading frame" list of Siegel et al. [23]. Reads per million and other routine calculations were done in Microsoft Excel. Differences in RNA abundance between conditions or fractions were assessed using DESeq [24]. Untranslated region sequences were downloaded from TriTrypDB and sequence motifs searched using DREME and MEME [25]. Other statistical analyses were done in R. Functional gene classes were assigned manually using a combination of automated annotations and publications. All raw sequence data are available at Array Express with accession numbers E-MTAB-4555 (polysomes) and E-MTAB-4557 (granules).
The polysome gradient data are available under submission numbers E-MTAB-4555 and E-MTAB-4575. The heat shock granule results are available under submission number E-MTAB-4557.
No ethical approval was reqiuired for this work, which did not involve either animals or human subjects.
The first part of our study concerned the movement of mRNAs into, and out of, the polysomal fraction after a one-hour heat shock at 39°C. We were particularly interested in knowing which mRNAs show regulation similar to that of ZC3H11, since we hoped in that way to identify conserved sequence motifs. We chose 39°C because preliminary results showed that the treatment was sufficient to move ZC3H11 mRNA into the polysomal fraction, while only partially inhibiting overall translation. It is also a treatment that could be tolerated by tsetse flies [9]. Lysates from procyclic-form trypanosomes with or without heat shock were fractionated on sucrose gradients, which were then divided into free (F), subunit (S), monosome plus light polysome (L) and heavy polysome (H) fractions (Fig 1A). The 39°C heat shock caused a shift of the ribosomes from the polysomal towards the free subunit and monosome fractions (Fig 1A). To find the proportion of mRNA that was in each fraction, we analysed samples by Northern blotting, using the spliced leader as probe (Fig 1B) and including inputs (non-fractionated samples) as controls. The total amount of mRNA from the 39°C-treated cells was 57% of that from the non-shocked samples, but the sucrose gradient distribution of mRNA was similar to that of the non-shocked parasites (Fig 1C and S1 Table, sheet 2). This suggests that loss of translation is associated with mRNA degradation. We could not tell from our results which effect happened first: decreased translation might cause mRNA decay, but conversely initial decay events such as decapping would prevent translation initiation.
All samples were subjected to RNASeq (S1 Table, sheet 3, and S1 Fig).
To find out the proportions of each mRNA in the sucrose gradient fractions, we normalised the read counts / million reads (S1 Table, sheet 3) according to the spliced leader signals (S1 Table, sheets 4 and 5). We then calculated the percentage of each mRNA that was in the different sucrose gradient fractions (S1 Table, sheet 6). (A summary of the control results at 27°C was included in [4].) The correlation coefficients for percentage in polysomes between replicates ranged from 87% to 99% (S1 Fig). In the following discussion we will assume that mRNAs that migrate in the denser part of the gradient are being actively translated. However, there are two caveats to this. First, binding of ribosomes to an mRNA does not necessarily mean that the ribosomes are active in translation elongation. Second, although there are no microscopically visible granules at 39°C [13], some association of mRNAs with smaller aggregates cannot be ruled out. The percentages in polysomes ranged from 50–70% for most mRNAs (Fig 2A, 'All’), and there was a statistically significant (but very small) increase in these percentages after heat shock.
There were 80 transcripts for which the percentage in polysomes decreased by a factor of 1.25 or more after heat shock (S1 Table, sheet 8, S2 Fig). For this group, the percentage on polysomes was overall higher than average at 27°C, and lower than average after an hour at 39°C. A subset of these mRNAs was distinguished by poor translation even before heat shock (S2 Fig, subset A): it includes RBP33, cis-spliced poly(A) polymerase, PAG2 and PAG4 mRNAs. We placed these mRNAs into functional classes based on their encoded proteins. Transcripts encoding ribosomal proteins were notably enriched in the set of mRNAs with decreased translation (Fig 2A–2C, S1 Table, sheet 9), and there was a slight over-representation of mRNAs encoding RNA-binding proteins (Fig 2C).
We next looked at mRNAs with a two-fold higher proportion in polysomal fraction after heat shock (S1 Table, sheet 7). 77 mRNAs, including that encoding ZC3H11, fell into this category. This group had almost universally been in the lighter fractions at 27°C and rather oddly, it was enriched in mRNAs encoding proteins of no known function (Fig 2D, S1 Table sheet 9). More detailed analysis of these mRNAs placed them into three categories (S3 Fig). At 27°C most of these mRNAs migrated in the free fraction, lighter than the subunits, and moved into the light polysomes after heat shock. Group (B) mRNAs started in the free fraction, but moved to both the subunit and light polysomal fractions after heat shock (S3 Fig). Group (A) mRNAs, which included ZC3H11, were distinguished from the others by the fact that they migrated mainly with the subunit fraction at 27°C. We have shown for ZC3H11 that this is not due to association with a small ribosomal subunit [13] and the reason for the different behaviour is unknown. The mRNAs that can bind to ZC3H11 showed slightly higher than average polysome loading at both 27°C and 39°C (Fig 2A).
To find changes in overall abundances of total and polysomal mRNAs, we compared the read counts from total RNA samples (S2 Table, sheets 1 and 2). 260 mRNAs were significantly (>2x, Padj<0.01) increased in relative abundance after heat shock (S2 Table, sheet 3). However, comparison of the mRNA yields (measured by spliced leader hybridisation as in Fig 1B) revealed that the total amount of mRNA had decreased by about 40% after heat shock. As a consequence, the numbers of copies per cell of most mRNAs were reduced (Fig 2B). Those mRNAs for which polysomal association increased tended to show less severe decreases after heat shock (Fig 2B). As noted above, it is not possible to assign cause and effect since translation could influence degradation and vice-versa.
The mRNAs that bind to ZC3H11 encode proteins that are always needed in high amounts, even at 27°C. These mRNAs were correspondingly strongly polysome associated at 27°C (Fig 2A). This is probably ZC3H11-independent because ZC3H11 is barely detectable at 27°C and RNAi has no effect on cell proliferation or morphology [16]. Association of ZC3H11 target mRNAs with polysomes was more marked at 39°C, when ZC3H11 is expressed (Fig 2A), but the relative increases were not significantly different from those of the bulk mRNA population (Fig 2A and 2B).
We now looked at the proteins encoded by mRNAs whose relative abundances increased at least 2-fold after one hour at 39°C, or which moved from non-translated to polysomal fractions. As expected, these included mRNAs encoding several chaperones, including two ZC3H11 targets (Tb927.10.16100 and Tb927.2.5980) (Table 1). There was also a moderate increase in the mRNA encoding the major cytosolic HSP70. The surprise came when we compared this group of mRNAs with transcriptomes from various developmental stages. We found a very significant overlap with mRNAs that are increased during the differentiation of procyclic-form trypanosomes to epimastigotes, metacyclic forms, and bloodstream forms (Fig 3A & 3B and S2 Table, Sheet 3). Even three of the chaperones were in this category. Notable among the epimastigote- or salivary-gland-specific genes were several that are associated with meiosis, MND1, HOP1, SPO11 and MSH5 (Table 1). The MND1 homologue (Tb927.11.5670) mRNA was not only increased in the total RNA, but also moved towards polysomes (45% in polysomes at 27°C, 74% at 39°C). The HOP1 homologue (Tb927.10.5490) mRNA showed a similar shift towards the polysomal fraction, but no RNA abundance change. YFP-tagged versions of both proteins are restricted to the nuclei of epimastigote-like cells in salivary glands [26]. SPO11 is probably also meiosis specific. MSH5 is annotated as a putative meiosis mismatch repair protein but there is no experimental evidence for this. Apart from these, mRNAs encoding several putative cell cycle regulators and a telomere-binding protein were increased in abundance or polysome association (Table 1).
Examination of mRNAs that decreased in abundance showed that they were spread over numerous functional categories. These mRNAs significantly overlapped with mRNAs that decrease during differentiation of procyclic forms to epimastigotes or bloodstream forms (Fig 3C). There was, in contrast, no significant overlap with mRNAs that decrease in stumpy bloodstream forms [27].
The numerous changes in developmentally regulated mRNAs after an hour at 39°C suggested that some regulatory proteins might also have been affected. Indeed, mRNAs encoding 9 potential RNA-binding proteins were increased after heat shock (Table 1). Of these, two mRNAs—DRBD6 and RBP6 –peak in salivary gland parasites [28]. Only 55% of the RBP6 mRNA was in polysomes at 27°C, but 87% was in the fraction after heat shock. Induced expression of RBP6 in procyclic forms is known to trigger the procyclic-epimastigote-metacyclic differentiation cascade [19]. RBP10 mRNA, which also increased after heat shock, is most abundant in growing bloodstream forms [30]. Expression of ZC3H14 and ZC3H45 proteins has not yet been detected but the ZC3H45 mRNA is preferentially translated in bloodstream forms [31]; and ZC3H46 protein is more abundant in bloodstream forms than procyclics [32]. RBP7 protein is present in slender bloodstream forms and increased in stumpy forms [33,34]. RNAi targeting RBP7 inhibits cAMP-induced stumpy-form differentiation, while overexpression of RBP7 causes G1/G0 cell cycle arrest and causes initial gene expression changes associated with procyclic differentiation [35].
In addition to these RNA-binding proteins, heat shock induced mRNAs encoding three protein phosphatases. Two of these have no known function, but PIP39 mRNA is higher in stumpy and procyclic forms than in bloodstream forms, and PIP39 becomes phosphorylated during stumpy-to procyclic differentiation [36]. Movement of PIP39, RBP6 and the Tb927.11.4990 kinetoplastid-specific phosphatase mRNA to polysomes, as well as some others with a similar pattern, was confirmed by Northern blotting (Fig 4A and 4B). Finally, to see whether a more moderate heat shock might also trigger changes in differentiation regulators, we grew procyclic forms at 37°C overnight. Indeed, the levels of both RBP6 and PIP39 proteins were increased (Fig 4C).
Our second series of experiments addressed mRNA targeting to heat shock granules, which form only at temperatures of at least 40°C [1]. Lysis of trypanosomes in the presence of 1% Triton X-100 results in trapping of structures with a diameter of more than 24nm within the microtubule corset [21]. (For comparison, a ribosome is just under 30nm across.) The trapped material can then be released with high salt, so that a further centrifugation yields a small granule (SG) supernatant and a large granule (LG) pellet. First, we examined cells growing at 27°C. The SG fraction contained about 3.2% of the mRNA, and the LG pellet just 0.8%, as judged by hybridisation with a spliced leader probe [13] (S4 Table, sheet 2). We subjected duplicate fractions to RNASeq (S4 Table, sheets3 and 4). To work out the proportion of each mRNA within the SG and LG fractions, we compared those results with those for total RNA (S4 Table, sheet 3). The replicates for total RNA of cells growing at 27°C did not correlate very well (S4A Fig), perhaps because the cells had somewhat different cell densities at the time of harvest (about 3.7 x106 and 5 x 106/ml). For the 27°C samples we therefore also compared the granule results for individual replicates (S4 Table, sheet 4) with those for the input in the polysome experiments (S4 Table, sheet 5, S4A Fig). Independent of the way the calculation was done, the proportion of mRNA that was trapped inside the microtubule corset in normally growing cells was determined mainly by the length of either the open reading frame or the complete mRNA (Fig 5A and S5A and S5B Fig). This suggests that the trapping was due simply to the size of the polysome and had nothing to do with regulation or granule formation. There was no significant correlation between the percentage in granule fractions and the percentage in polysomes at 27°C (Fig 5B). It was however notable that mRNAs encoding ribosomal proteins were not trapped in granule fractions at all. Even allowing for the short lengths of most ribosomal protein mRNAs (Fig 5A), their behaviour was anomalous (S5 Table).
We next examined the effect of a 41°C heat shock on the distribution of mRNAs in granule and non-granule fractions. First, we compared results for total mRNAs with those obtained at 39°C, and also with previously published results (S1C–S1E Fig). The variability in the 27°C dataset (S4A Fig) meant that P-values for the total RNAs were high (S5 Table) and the overall correlation between different experiments was poor. This probably reflects differences in cell density as well as temperature. However, a core set of mRNAs was increased in at least 2, and often all three, datasets (Table 1 and S2 Table, sheet 11). In addition to a few chaperone mRNAs, these once again included mRNAs indicative of developmental regulation. They encoded CYC7, CYC11 and CYC10; SPO11 and MND1; bloodstream-specific alternative oxidase, pyruvate kinase and GPI-PLC; 6 protein kinases; 3 protein phosphatases including PTP1; and 8 RNA-binding proteins including both RBP10 and RBP6.
After one hour at 41°C, 6% of the total mRNA was in the small granule fraction, and 19% in large granules (S4 Table, Sheet 2). At the level of mRNAs from individual genes, however, the distribution looked very different.This is because half of the sequence reads were contributed by the most abundant 10% of the transcripts. For most coding sequences, 20–60% of the mRNA was in one of the granule fractions, usually with the large granule fraction predominating (Fig 5C and 5D). In contrast, a subset of very abundant mRNAs was not associated with granules (Fig 5D). These included those encoding ribosomal proteins, procyclin, the major cytosolic HSP70, mitochondrial HSP60 and a DNAj (Figs 5D and S4C and S4D.). Other mRNAs that showed less than 20% association with heat shock granules were those encoding histones, alpha and beta tubulin, 10 additional chaperones, the cytochrome oxidase complex and a few proteins involved in ribosome assembly. An ANOVA test showed that the mRNAs encoding ribosomal proteins were the only functional category that showed unique behaviour with regard to heat shock granule association (P = .00015 with Bonferoni correction). In subsequent analyses we therefore treated this group separately.
We previously showed that ZC3H11 prevents degradation of bound mRNAs after a 41°C heat shock [16]. Correspondingly, mRNAs that co-purify with ZC3H11 [16] tend to escape granule association. For the 23 mRNAs that showed the strongest enrichment in the ZC3H11-bound fraction [16], a median of 20% was associated with total granules, whereas for unbound mRNAs the median was 40% (Figs 5D and 6A). A similar result was obtained if large granules alone were analysed (S4E Fig). As previously noted, the ability to bind ZC3H11 also correlated with higher association with polysomes (Figs 5E and 6B). These results suggest that ZC3H11-bound mRNAs are protected against mRNA degradation, translational inactivation, and incorporation into granules. To check this hypothesis, we prepared granule fractions from cells with and without heat shock and / or ZC3H11 RNAi. Without RNAi, two target mRNAs encoding HSP70 and an FKBP remained largely in the soluble fractions despite heat shock: as seen from the RNASeq results, only a tiny proportion was detected in the large granule fraction (Fig 7A). ZC3H11 depletion had very little effect on this distribution without heat shock (Fig 7B and 7C). After heat shock, however, granule-free HSP70 and FKBP mRNA disappeared but neither mRNA accumulated in the large granule fraction either: instead, the mRNAs were simply destroyed.
After heat shock, there was little overall correlation between the coding region length and association with either total granules (S4C Fig) or small granules alone (S4D Fig), but DeSeq analysis showed that granules were enriched in long mRNAs (median length 4 kb) including several encoding large cytoskeletal proteins. (S5 Table, sheets 3 and 5). There was no overall correlation between loading onto polysomes at 39°C and the percentage in granules at 41°C. Some potential regulators that showed reproducible mRNA abundance increases—CYC7, DRBD5, DRBD6 and RBP6—showed less than 30% granule incorporation, suggesting that they might in some way be implicated in recovery from heat shock.
The results from this study have confirmed that the ability of an mRNA to bind ZC3H11 correlates not only with stabilisation at high temperature, but also with continued translation and exclusion from heat shock granules. The first conclusion generalises results that were already seen for reporters with the HSP70 3'-UTR, while the second is consistent with previous published data indicating that mRNAs in stress granules are not translated [37–40]. The mRNAs that are bound by ZC3H11 are already quite well translated at 27°C, and become even more so after heat shock: it is possible that this high translation protects them from incorporation into heat shock granules; alternatively ZC3H11 and its associated proteins [17] might prevent sequestration of bound mRNAs into granules.
Our results show that heat shock granules are not identical to starvation granules, despite sharing some of the same proteins and mRNAs. Some mRNAs that were excluded (<20%) from heat shock granules were also similarly absent from starvation granules [21]. Presumably these encode products that are required to recover from both starvation and heat shock. Apart from ribosomal protein mRNAs, which are probably a special case and are discussed below, several chaperone mRNAs were in this category. In contrast, ZC3H11 is not implicated in the starvation response, and its mRNA was 25% in heat shock granules but 79% in starvation granules. Other mRNAs that showed a similar pattern encoded RBP3, ZC3H30, ZFP1, RBP6 and a histone H3 variant [21]. PABP1 may be important in protecting ZC3H11 target mRNAs, since it is recruited by the ZC3H11-MKT1-PBP1 complex [17]. The level of ZC3H11 protein is not increased after starvation, which explains why some ZC3H11 target mRNAs are incorporated into starvation granules (S4 Table, sheet 7).
The mRNAs encoding ribosomal proteins were almost completely excluded from both heat shock granules and starvation granules [21]. Only two annotated "ribosomal protein" mRNAs, Tb927.10.10010 and Tb927.11.6360, did not follow this pattern, but neither is a structural component of the mature ribosome. The extraordinary behaviour is therefore a universal characteristic of mRNAs that encode components of the mature ribosome. These mRNAs are also outliers in other ways: the mRNA levels are higher that would be predicted based on their half lives and gene copy numbers [4], and the average ribosome densities are relatively low (mostly less than 4 ribosomes/kb) [4,31] although the majority of the mRNAs are loaded onto polysomes [4]. Association with polysomes is also notably decreased after heat shock, without much loss of the mRNAs (Fig 2A and 2B): it looks as if the mRNAs are being conserved in some way other than granule sequestration. The ribosomal protein mRNAs are co-regulated during trypanosome differentiation, being decreased in stationary phase trypanosomes and increasing only 1h after addition of the differentiation stimulator cis-aconitate [41]; this is consistent with the fact that they mostly peak in the G1 phase of the cell cycle [42]. We examined the untranslated regions of these mRNAs for specific enriched motifs and found none. The only notable feature is that the 5'-UTRs are very short, with a median length of 22nt, as opposed to 108 for other mRNAs (mean±SDs are 33±34 as opposed to 203±274).Given the lack of conserved linear motifs, it is possible that secondary structures are important; or, more unusually, ribosomal protein mRNAs might be characterised by a lack of motifs required for recruitment of SCD6 [14] or other granule proteins. Alternatively they might be regulated via recognition of the nascent polypeptides.
Our investigation of polysome loading revealed interesting sets of mRNAs that were retained in polysomes and/or increased in abundance at 39°. Some, like ZC3H11, were rather poorly translated at the normal temperature; these migrated either near 40S, or somewhat above 40S. The reason for the difference is unknown but binding to the small subunit is unlikely [13]. The 7 mRNAs with patterns most similar to that of ZC3H11 encoded a protein kinase, a protein phosphatase, a DNAj-like protein, and 4 other proteins of unknown function. There is no evidence of any link between these proteins and ZC3H11 function: although ZC3H11 is phosphorylated, the most likely culprit is a different kinase, casein kinase 1.2 [13].
Perhaps the most interesting observation was that the mRNAs that showed increased translation or abundance at 39°C included mRNAs that are up-regulated in salivary gland trypanosomes (Fig 2). The mammalian body environment has a temperature of 37°C (possibly higher in organs) and a 10°C temperature decrease is known to be an important factor in the switch from bloodstream to procyclic forms. However, the mRNAs that increased were not necessarily bloodstream-form specific. Indeed, the mRNAs encoding three chaperones, two cyclins, the meiotic mRNA MND1, and the RNA-binding proteins DRBD6 and RBP6, are elevated in salivary-gland parasites but not bloodstream forms (Table 1). Induced expression of RBP6 in procyclic trypanosome cultures (at 27°C) causes differentiation to epimastigotes, and then to metacyclic forms: after 24h of RBP6 expression, about 10% of cells are epimastigotes, while metacyclics begin to appear after 5–6 days [19]. It is therefore formally possible that all of the polysomal RNA changes that we saw upon heat shock are caused by RBP6. However, this seems unlikely since the 1-h time frame is extremely short. For example, the trypanosome alternative oxidase protein appears after 2 days of RBP6 expression, but the mRNA (Tb927.10.9760) moves towards the polysomes after only an hour at 39°C (to 72% from 54%).
Differentiation of bloodstream forms to procyclic forms includes an intermediate called the short stumpy form. Stumpy forms are arrested in G1, and express some proteins of procyclic form metabolism. Further differentiation to procyclic forms is induced by addition of cis-aconitate and a decrease in temperature from 37°C to 27°C. Heat shock of procyclic forms resulted in increased polysomal levels of two mRNAs implicated in this process. The first was the protein phosphatase PIP39, which is essential for differentiation of stumpy forms to procyclic forms [36], and which was increased at the protein level by incubating the procyclic forms at 37°C. RBP7, a potential RNA-binding protein, is required for differentiation of bloodstream forms to stumpy forms [35], and this mRNA moved towards polysomes after heat shock of procyclics. Importantly, we showed that PIP39 and RBP6 proteins increased in procyclics incubated at 37°C, which is a temperature that is quite likely to occur in the wild. It is possible that both PIP39 and RBP7 have functions–perhaps linked to growth arrest—in both the stumpy->procyclic and in the procyclic->epimastigote transitions.
There are several indications that stress responses can promote trypanosome differentiation, but it is not always clear whether differentiation is a direct or indirect effect [43]. If cell cycle arrest is needed for alterations in signalling and re-programming of gene expression, and a stress causes cell cycle arrest, differentiation might be enhanced although the stress does not induce differentiation directly. For example, the differentiation of stumpy forms to procyclic forms can be enhanced or promoted by a variety of stressful treatments, including mild cold shock [44], glucose deprivation [45], mild acid [46], and protease treatment [47], as well as by cis-aconitate. It is not known which of these stresses is physiologically relevant in tsetse. We know even less—in fact, nothing—about the stimuli within the fly digestive tract that initiate the development of procyclic forms to epimastigotes and metacyclic forms. A heat shock is definitely not required since development happens in laboratory tsetse colonies in which temperatures are controlled below 30°C. In Africa, however, tsetse are very likely to experience higher environmental temperatures, and the developing trypanosomes are exposed to warm blood meals every 3–5 days [48]. It is therefore possible that in the wild, temperature fluctuations inside tsetse, or other stresses, could play a role in trypanosome life-cycle progression.
|
10.1371/journal.pcbi.1003740 | Algorithms to Model Single Gene, Single Chromosome, and Whole Genome Copy Number Changes Jointly in Tumor Phylogenetics | We present methods to construct phylogenetic models of tumor progression at the cellular level that include copy number changes at the scale of single genes, entire chromosomes, and the whole genome. The methods are designed for data collected by fluorescence in situ hybridization (FISH), an experimental technique especially well suited to characterizing intratumor heterogeneity using counts of probes to genetic regions frequently gained or lost in tumor development. Here, we develop new provably optimal methods for computing an edit distance between the copy number states of two cells given evolution by copy number changes of single probes, all probes on a chromosome, or all probes in the genome. We then apply this theory to develop a practical heuristic algorithm, implemented in publicly available software, for inferring tumor phylogenies on data from potentially hundreds of single cells by this evolutionary model. We demonstrate and validate the methods on simulated data and published FISH data from cervical cancers and breast cancers. Our computational experiments show that the new model and algorithm lead to more parsimonious trees than prior methods for single-tumor phylogenetics and to improved performance on various classification tasks, such as distinguishing primary tumors from metastases obtained from the same patient population.
| Cancer is an evolutionary system whose growth and development is attributed to aberrations in well-known genes and to cancer-type specific genomic imbalances. Here, we present methods for reconstructing the evolution of individual tumors based on cell-to-cell variations between copy numbers of targeted regions of the genome. The methods are designed to work with fluorescence in situ hybridization (FISH), a technique that allows one to profile copy number changes in potentially thousands of single cells per study. Our work advances the prior art by developing theory and practical algorithms for building evolutionary trees of single tumors that can model gain or loss of genetic regions at the scale of single genes, whole chromosomes, or the entire genome, all common events in tumor evolution. We apply these methods on simulated and real tumor data to demonstrate substantial improvements in tree-building accuracy and in our ability to accurately classify tumors from their inferred evolutionary models. The newly developed algorithms have been released through our publicly available software, FISHtrees.
| In this paper, we develop new methods to advance the theory of phylogenetic inference for reconstructing evolutionary histories of cell populations in solid tumors. The work is specifically designed for use in tracking tumor evolution by gain and loss of genomic regions as assessed by multicolor fluorescence in situ hybridization (FISH), which measures the copy numbers of targeted genes and chromosomes in potentially hundreds of individual cells of a tumor. This technology was the basis of the earliest methods for phylogenetic reconstruction of single tumors [1], [2]. FISH remains uniquely valuable for such studies because the large number of cells that FISH can profile makes it possible to collect data on enough tumors in enough detail to build cell-by-cell phylogenies for populations of tumors and begin to study the common features of these phylogenies. In the present work, we specifically extend our previously developed inference algorithms to encompass a more complicated but realistic model of evolution of FISH probe counts, accounting for gain and loss of genetic material at the level of single gene probes, multiple probes on a single chromosome, or a probe set distributed across the whole genome. We demonstrate the value of these algorithmic improvements to more accurate phylogenetic inference and improved effectiveness of the resulting phylogenies in downstream prediction tasks.
The present work adds to the growing list of phylogenetic methods in cancer modeling, which were reviewed through 2008 in [3]. These include methods for analyzing comparative genomic hybridization (CGH) or other genetic gain/loss data in a single tumor type [4]–[11], for defining the cell type lineage of single tumors [1], [2], [12], , for organizing a taxonomy of tumor types [14], for reconstructing a partial order of genetic changes in multiple samples from one patient [15], and for reconstructing progression from cell types inferred from bulk genomic assays [16]. Recent high-throughput sequencing studies have also used ad hoc phylogenetic methods to infer putative tumor progression scenarios, e.g., [17]–[20]. Like many of these methods, the present work is aimed at building tree models that provide a proposed partial order on the observed cell states, a strategy motivated originally by the work of Fearon and Vogelstein, proposing a linear order for four types of events in colorectal cancer and associating each event with a tumor stage [21]. Other ordering methods have been proposed, mostly for CGH or breakpoint data [15], [22]–[28] and, more recently, sequencing data [29], [30].
The present work specifically advances the reconstruction of phylogenetic histories of single tumors from intratumor cellular heterogeneity data. The use of phylogenetic methods to reconstruct histories of single tumors was first developed in our prior work [1], [2] by taking advantage of the ability of FISH to profile genetic changes in large numbers of single cells, allowing one to survey hundreds of cells per tumor in populations of tens of tumors [31]. This early work showed that even small numbers of markers could reveal numerous genetically distinct cell populations in single tumors, which could be resolved by phylogenetic inference to reveal multiple distinct pathways of progression between tumors and even within single tumors. Numerous studies since then, using multicolor FISH [2], [31]–[36] and, more recently, single-cell sequencing [19], [37]–[39] have greatly increased our ability to identify distinct cell populations and, in the process, revealed far more extensive intratumor heterogeneity than had been suspected prior to 2010 (reviewed in [40]). The repeated observation of intratumor heterogeneity has necessitated a reconsideration of Nowell's [41] theory that tumors evolve clonally, showing that a tumor may contain many subpopulations relevant to the clinical prognosis of the patient [42] and that rare subpopulations may be more relevant to prognosis than the most common ones [43]. Furthermore, a simulation study has suggested that methods based on average copy number data perform poorly when there is substantial intratumor heterogeneity [44]. Such findings suggest a need for improved methods for organizing the dozens or hundreds of observed cell states in single tumors to infer the evolutionary processes that produced them.
Despite extensive work on tumor phylogenetics, however, the study of algorithms for reconstructing tumor evolution from large numbers of single cells has lagged far behind advances in data generation. The standard in practice for single-cell tumor phylogenetics remains the use of simple generic phylogeny algorithms (e.g., neighbor-joining [45]) that are not designed to model the patterns of copy number changes one would expect from evolution by chromosome abnormalities that largely drive tumor evolution. Until recently, algorithms designed specifically for inferring phylogenies of single tumors from FISH data have been limited to just a few probes per cell and lacked robust, publicly available software implementations [1], [2], [34]. In prior work [46], we developed algorithms to find copy-number phylogenies for in principle arbitrary numbers of probes and cells. That work, however, was itself limited to a simple model in which tumor cells evolve by events of gain or loss of a single copy number of a single probe at each mutation step. In real tumors, gene copy numbers can change due to a variety of mechanisms, including:
These events are illustrated schematically in Figure 1. While more complex probabilistic models of tumor evolution have been developed for inference of small phylogenies, with approximately ten taxa per tumor corresponding to distinct biopsies (e.g., [47]), the class of inference algorithms such models require would not be expected to scale to phylogenies of hundreds of single cells per tumor such as those examined in the present work.
The work presented here seeks to fill this need for scalable phylogenetic algorithms capable of fitting more realistic models of tumor-like evolution to data sets of hundreds of single cells per tumor. We improve on our prior work for inferring tumor evolutionary models considering only SD events [46] to now include CD and GD events, which are also frequently observed in tumor progression. We specifically focus on the problem of accurately inferring evolutionary distances between distinct cells in terms of maximum parsimony combinations of SD, CD, and GD events. The major contributions of the work are:
The new methods are implemented in version 2 of our software FISHtrees (ftp://ftp.ncbi.nlm.nih.gov/pub/FISHtrees). The work addresses a critical need in modern cancer research for algorithms capable of inferring evolutionary trajectories of hundreds of single cells per tumor under plausible models of evolution including both gene-specific and chromosome abnormalities that are central drivers of true tumor evolution.
We used data collected from cervical cancer (CC) [49] and breast cancer (BC) [36] patients to evaluate our methods. Figure 2(A) shows a tumor progression tree inferred from one of the cervical cancer samples. For comparison, Figure 2(B) shows a progression tree inferred on the same sample using our prior SD model [46]. Visual inspection shows that large regions of the two trees are identical but that allowing CD and GD events leads to some rearrangement and a reduction in tree depth and overall size. Next we evaluate the changes induced by adding SD, CD and GD events, using simulated data to show effectiveness of the methods in finding more parsimonious solutions to the broader model and using the real CC and BC data to show the biological relevance of the improvements. We further show that our algorithms infer trees with higher accuracy than the prevailing alternative algorithms for single-tumor phylogenetic inference. Finally, we perform statistical experiments to evaluate the effects of tumor sample size on the performance of our tree building algorithm.
To measure accuracy of the methods for FISH datasets with a known ground truth, we generated a dataset of trees with six probes, two of which were treated as being on the same chromosome. Each tree was generated by starting from a diploid root node and executing a branching process in which each node was recursively assigned a number of children drawn from a geometrically distributed random variable with mean . Each child was distinguished from its parent by selecting an SD, CD, or GD event with probability for each of the six possible SD events, of a CD event, and of a GD event. This process terminated when all leaf nodes had been assigned zero children by the sampling. We then generated simulated FISH data for each tree by uniformly sampling cells from the nodes in this topology. The simulated data corresponds to counts of probes for each sampled cell in the tree. We applied Algorithm 3 (see Methods) to find a minimum-cost tree for each of four event models: (i) SD only, (ii) SD and CD, (iii) SD and GD, and (iv) SD, CD and GD.
We quantified the accuracy of tree inference by comparing each simulated true tree to its corresponding inferred tree derived from the sampled cells. This assessment was performed at the level of accuracy of tree edges by the following procedure:
Intuitively, this formula measures the fractional agreement between bipartitions of the trees relative to the total number of bipartitions. We use a matching-based formula, rather than the more familiar Robinson-Foulds metric [53], both because of its greater sensitivity to small changes in trees and because the Robinson-Foulds measure is not defined for trees with different node sets. We also note that we use a different normalization factor than in our prior work [46], normalizing essentially by the total number of edges between the two trees, to control properly for the fact that different inference methods may infer different numbers of tree edges. The reconstruction error ranges in value from , if the real and inferred trees are isomorphic, to an upper bound of in the limit of complete disagreement.
To illustrate the meanings of the terms of the equation for , we present a simple example using a hypothetical ground truth and an inferred tree presented in Figure 3(A) and Figure 3(B), respectively. The set of nontrivial bipartitions in the ground truth are
and the nontrivial bipartitions in the inferred tree are
If we apply the matching algorithm on these two sets of bipartitions, the first and second bipartitions in the ground truth tree are matched with the first and second bipartitions in the inferred tree, respectively. The weight of the matching is . The number of common taxa between these two datasets is . The total number of nontrivial bipartitions in the real and inferred trees are and . Plugging these values into the equation for , we calculate .
A comparison of the four models is presented in Figure 4. The SD model showed reconstruction error with standard deviation (s.d.) of across the trees. The SD+CD model yielded error with s.d. . SD+GD yielded error with s.d. . The full SD+CD+GD model yielded error with s.d. . Collectively, the results suggest that one can reconstruct reasonably accurate trees even from the SD-only model, despite the fact that the trees were generated from a model of all three event types, although accuracy improves with each event type added. Accounting for GD events made a larger difference in accuracy than accounting for CD events, presumably because a missed GD event might require many SD or CD events to explain it, while a missed CD event could be explained with just two SD events. The reconstruction error for the full model is reduced by more than 1.7-fold relative to the SD-only model considered in our prior work.
We further compared these results to those derived using generic phylogenetic methods that have been used in much of the single tumor phylogenetics work to date [16], [54]. We tested the accuracy of reconstruction of the simulated trees described above using generic neighbor joining (NJ) with Euclidean distance and pure maximum parsimony (MP) treating copy numbers as arbitrary characters, approaches chosen because they have been the primary alternatives to our specialized algorithms in the single-tumor phylogeny literature. We omit here comparison to more complicated Bayesian phylogenetic models (e.g., [47]) because such approaches are not scalable to the numbers of cells we examine. We then used the weighted matching based similarity method, described above, to calculate the mean percentage reconstruction error between the inferred and the ground truth trees. The mean reconstruction errors for NJ and MP were (s.d. ) and (s.d. ), respectively, in contrast to the error of (s.d. ) for the SD+CD+GD algorithm proposed here. The test thus demonstrates that when the underlying evolutionary process includes cancer-like chromosome abnormalities, errors are substantially reduced by using an algorithm designed for that model relative to standard off-the-shelf algorithms still widely used for single-tumor phylogenetics work.
We performed additional experiments to evaluate the effects of different evolutionary parameters on the accuracy of inference of tumor progression trees by FISHtrees. For this experiment, we selected five different combinations of probabilities of SD, CD and GD events for generating the ground truth trees and then used SD, SD+CD, SD+GD and SD+CD+GD models to infer the tumor phylogenies. These data sets again each used six probes with two of the six on a common chromosome. The selected five combinations of (SD,CD,GD) event probabilities are: , , , and . These combinations of event probabilities were chosen to yield trees of comparable complexity to the real data while producing test sets enriched in distinct combinations of the three event types. They thus allow us to consider how robust our algorithms are to contributions from each of the three event types, singly or in combination. We report the reconstruction error for trees for each of these combinations of event probabilities in Table 1. These results again show that accuracy improves with each event type added. When the probability of SD events is high (as in combination 3), the SD model results in highly accurate trees (mean reconstruction error of with s.d. ). Accounting for GD events in combination with SD events always result in larger improvement in the reconstruction error in comparison to the SD+CD models, even when the CD events are very frequent (as in combinations 2 and 4). Finally, accounting for GD events in combination with SD and CD events results in the largest improvements when the probability ratio of GD events to SD+CD events is highest, as can be seen from comparison of parameter sets 1 and 2.
Next, we performed simulation tests to evaluate the effects of non-uniform distributions of cells across different levels of the trees on the performance of our tree inference method. In our initial simulation experiments described above, we assumed that observed cells were sampled uniformly across clones. In real tumors, the distribution of cells would not typically be uniform due to differences in age and fitness of clones. In order to test robustness of our method to non-uniformity of clone frequencies, we sampled the cells following a non-uniform model in which the sampling frequency of a clone varies geometrically with its depth in the tree with a parameter . We used values of and for in our experiments. When , of the total cells are located in the first three levels of the trees, while for , this fraction is . We generated trees in each case with probabilities of SD, CD and GD events fixed at and . We again used SD, SD+CD, SD+GD and SD+CD+GD models to infer the tumor progression trees. We present the results from this experiment in Table 2, where we also show the results from the uniform sampling of the cells. Additionally, we report the results on the trees inferred using NJ and MP for these three different cell distributions. From the table, we can see that the reconstruction error increases with increasing for all methods. The SD+CD+GD model, however, shows the best performance among all the models for all three values of and the least loss of performance with increasing .
Finally, we performed simulation experiments to understand the effects of varying the numbers of chromosomes with multiple probes. We created a simulated dataset of trees with eight probes where two pairs of probes each reside on two different chromosomes and the remaining four probes reside on four separate chromosomes. The probabilities of each of the SD, CD and GD events were fixed at , and , respectively. We report the results from this experiment in Table 3, which compares the results from this experiment with our earlier result using only a single chromosome with two probes and four other probes located on separate chromosomes. The table shows that inclusion of the extra possible CD event results in higher accuracy for all the models except for the SD only model. The performance drop in the SD model is expected, as it would require more SD events to explain a greater number of missed CD events. The highest gain in performance is observed for SD+CD+GD model. These results show that our algorithm will tend to yield comparatively more advantage over the earlier work with more complicated scenarios of sharing probes across chromosomes, suggesting its utility will increase as improvements in technology allow for larger probe sets.
We applied the algorithm to two sets of real data:
Among the eight genes in the BC dataset, DBC2 and MYC reside on chromosome and HER-2 and TP53 reside on chromosome . The other four genes belong to distinct chromosomes. The oncogene Cyclin D1 (CCND1), which plays a role in many solid tumor types, is in both the BC and CC datasets. However, in some other tumor types, such as oral cancer, CCND1 is part of a larger region with recurrent copy number gains on chromosome and other nearby genes have also been suggested to play a role in oncogenesis [66].
We evaluated the SD+CD+GD method by its effectiveness in reducing the parsimony score (total number of mutation events) of the resulting trees relative to the prior SD-only model. With the primary CC samples, the SD+CD+GD method found a lower-cost tree in of cases, a tree of equal weight in cases, and a higher-cost tree in cases. In each case of increased weight, the increase was by and appears to result from the subtree regrafting heuristic used in handling GD events (see Methods). These results suggest that the heuristic tree search may more often yield a suboptimal result for the SD+CD+GD model than it does for the SD-only model. The benefit of the more realistic model, however, outweighs the cost of this suboptimality in a large majority of instances. For trees derived from metastatic samples, of trees had lower weight for the full SD+CD+GD model and the remainder all had equal weight for the two models. Metastatic data sets tend to have fewer distinct cell types than do primary trees and thus may represent an easier optimization challenge. For the BC samples, of DCIS (samples 1–13) and of IDC (samples 14–26) had lower weight for the full model, with the remaining one sample having equal weight. Parsimony scores by tree are provided in Figures 5 and 6.
We next evaluated effects of the improved model on overall tree topology, based on results of our prior work [46] that tree topology can significantly distinguish trees drawn from distinct progression stages of a given tumor type, with possible implications for the varying balance of diversification and selection acting on different stages of tumor progression. Figure 7 quantifies the topology for each sample set based on fractions of cells inferred at each tree depth from to . The figure shows similar qualitative trends for both SD and SD+CD+GD methods, although with small quantitative differences. For example, both SD and SD+CD+GD trees recapitulate a tendency for CC primary trees to show relatively broad topology (Figure 7(A)) while CC metastatic trees prune rapidly beyond the first few tree levels (Figure 7(B)). There is, however, an overall shift to lower depth in the SD+CD+GD trees. For CC primary trees, of cells are located in the first tree levels for SD versus for SD+CD+GD. For CC metastatic, of cells are located in the first tree levels for SD versus for SD+CD+GD. For BC, the comparable numbers of cells in depths are for SD versus for SD+CD+GD in DCIS and for SD versus for SD+CD+GD. These results suggest that the overall tree topology is not greatly sensitive to the combination of event types, although there is a noticeable shift towards lower depth in the full model.
An additional evaluation was possible for the BC trees, because for the BC data, a probabilistic model and expert annotation based on two additional centromere probes made it possible to estimate the cell ploidy [36], which we define as the mode among the number of copies of the twenty-two autosomal chromosomes in a cell. Each cell in that dataset is thus annotated with an expert-curated overall ploidy estimate. We used these ploidy estimates to validate our inference of GD events based on whether edges assigned to GD events in our trees correspond to doubling of annotated ploidy. The percentage agreement by edge between GD events and annotated doubling in ploidy is across DCIS trees and across IDC trees. In of all inferred GD events, at least one endpoint of the corresponding edge is a Steiner node, and the uncertainty among whether a GD event occurred prior to or after the emergence of the Steiner node may explain why the per-edge agreement is not higher. Nonetheless, the data support the conclusion that inferred GD events are correct in a majority of cases.
As a final step, we repeated an approach developed in our prior work [46] to both validate the biological relevance of the trees and develop a practical application of them by treating the trees as sources of features for classification tasks applied to the CC data. For this purpose, we developed several sets of quantitative features based on inferred trees as well as comparative features derived from raw FISH probe counts. We used the following set of tree-based features:
We omitted a third feature set, bin count, used in our prior work because it is not easily comparable between SD and SD+CD+GD trees. We compared these features to four features derived directly from FISH probe counts without reference to the trees:
We used each feature set as input to the Matlab support vector machine (SVM) classifier with a quadratic kernel using rounds of bootstrap replicates per test with leave-one-out cross-validation to compute mean and standard deviation of accuracy. We used Matlab functions “svmtrain” and “svmclassify” for training and testing of the SVM classifier.
We then applied these methods for three classification tasks: (i) distinguishing primary samples that progressed to metastasis from their paired metastatic samples, (ii) distinguishing all primary samples from all metastatic samples, and (iii) distinguishing primary samples that metastasized from primary samples that did not metastasize. The first two tasks are relevant to identifying features that help us understand the differences in evolutionary mechanisms of primary and metastatic samples. The third is intended to model an important practical problem in cancer treatment: determining whether a given primary tumor will metastasize.
Figure 8 shows results on each task. For task (i), allowing SD+CD+GD events increased accuracy relative to SD trees from to for edge counts and from to for tree level cell count. The SD+CD+GD tree level cell count was the most effective of all features, tree-based or not. For task (ii), we similarly saw a substantial improvement in prediction accuracy for SD+CD+GD trees relative to SD trees. Classification accuracy improved from to for edge count features and from to for tree level features. In this case, both SD+CD+GD tree feature sets outperformed all other features sets, tree-based or otherwise. These results provide an indirect validation that using a more general tree model gets closer to the biological ground truth. For task (iii), we saw no improvement, with identical results for SD and SD+CD+GD trees for either feature set. All tree-based feature sets significantly outperformed all non-tree-based feature sets for this task. We conclude that the more realistic evolutionary models appear not to reveal any more information to the classifiers for predicting which primary samples will go on to metastasize than the SD trees, which were already quite effective for that task.
A key advantage of FISH for profiling tumor heterogeneity is that it makes it cost-effective to profile much larger numbers of cells than alternatives such as single-cell sequencing. To assess the practical importance of this advantage, we asked two related questions: (1) how many cells do we need per tumor to accurately reconstruct single-cell phylogenies and (2) how many tumors do we need to examine to identify reproducible, statistically significant features across trees.
We first assessed the number of cells needed per tumor by using our first simulated dataset of trees described above with subsamples of varying numbers of cells per tumor, measuring reconstruction error of our SD+CD+GD algorithm with the weighted matching algorithm. The mean reconstruction errors calculated across cases for subsamples of , , , and cells were (s.d. ), (), (), (), and () respectively. We can thus conclude that accuracy improves noticeably with increasing numbers of cells to at least cells per tumor before plateauing at approximately error.
We next assessed numbers of tumors needed to identify meaningful statistically significant properties of tumor classes by analysis of the CC paired and primary samples. We randomly subsampled from among the pairs and, for each subsample, calculated the following three tree statistics on progression trees inferred from our SD+CD+GD algorithm:
We then compared distributions of each statistic on primary vs. metastatic trees by a Wilcoxon signed rank test. As the samples were selected randomly, no ordering among the samples was considered. Figure 9 shows the 1-sided p-values of the three statistical tests when the number of randomly selected samples are increased from to . The figure shows that ability to distinguish the two tumor subsets improves with increasing number of tumors. While the threshold for significance varies by statistic, each reaches weak significance (p0.05) between and tumors. We can thus conclude that finding reproducible features distinguishing the tree types requires on the order of tens of tumors, at least for the candidate probe sets examined here.
Taken together, these two results demonstrate that building accurate trees on a large enough scale to distinguish meaningfully primary from metastatic trees requires data sets with roughly the order of thousands of single cells (hundreds of cells per tumor for tens of tumors), a scale of data that has so far been achieved only by FISH studies of tumor heterogeneity. We note, however, that one would expect these numbers to vary depending on the degree of tumor heterogeneity, the classes of trees one wishes to distinguish, and the specific markers examined.
This paper has presented novel theory and algorithms for reconstructing evolutionary trajectories of gene copy numbers in solid tumors in terms of a model of tumor evolution incorporating changes at the scale of single gene probes, full chromosomes, or all probes in the genome. We have derived algorithms to reconstruct maximum parsimony sequences of events, and thus estimates of evolutionary distance, between pairs of cells assayed by FISH probes. We have further incorporated these inferences into a method for building phylogenies of hundreds of cells in single tumors. These methods have been added to FISHtrees [46], our software for inferring tumor phylogenies from single-cell copy number data. Experimental results on simulated data confirm the ability of the new methods to improve phylogenetic inference accuracy relative to simpler models by adding CD and GD events that model chromosome-scale and whole-genome copy number changes that are frequently observed in tumor evolution. Application to observed human tumor data shows that these extended evolutionary models are able to yield more parsimonious tree reconstructions and that the resulting trees lead to improved accuracy in prediction tasks related to diagnosis and prognosis.
In future work, we hope to extend the theory developed here to handle even more realistic models and more challenging data types. One important direction will be advancing the theory developed here to improve upon the heuristic approximations used in the Steiner tree inference to better approach the goal of finding globally optimal trees for the most computationally challenging FISH data sets. The evolutionary models, likewise, might be further extended to go beyond the three mutational event types considered here to better approximate the numerous distinct mutational mechanisms by which copy number profiles of tumor cells might evolve. The data sets studied here do not include geographical information about locations of individual cells in the tumor, but other data sets for analyzing tumor heterogeneity do include such geographical information [38], [68]. We expect it would be interesting to construct phylogenies with distance functions that combine spatial distance in three dimensions with combinatorial distance measures between the cell count patterns, as we have studied here. Further, while FISH for the moment retains a unique advantage in the large number of cells it can profile, one can reasonably anticipate that single-cell sequencing will eventually become practical for comparable cross-tumor studies. There would thus be value in extending the theory developed here to single-cell sequencing data, a goal that would pose substantial algorithmic challenges due to the much larger number and variety of markers it can reveal as well as the more complicated error models it would entail. Finally, we hope to make more use of these single-tumor phylogenetic models in clinically relevant prediction tasks and further explore the biological insights one can gain from more accurate tumor phylogenies.
Our main theoretical result is a method for inferring minimum distances between two states within a copy number phylogeny when duplication/loss of single genes (SD), duplication/loss of all genes on a common chromosome (CD), and duplication of all genes in the full genome (GD) events are possible. We first establish some mathematical results and then develop an algorithm for accurate distance computation. This algorithm then becomes a subroutine in a heuristic Steiner tree algorithm for inferring copy number phylogenies in the presence of SD, CD, and GD events. We introduce some notation required for specifying and proving the theoretical results:
We develop the theory for inference of the Steiner (unsampled or extinct cell configurations) nodes in the paths formed by the sequence of gene copy number gains and losses from an initial configuration to a final configuration . We first extend the prior theory to account for SD and CD events. Our model assumes that on division of a tumor cell, the configuration can change either by gain or loss of one copy of a single gene (SD event) or by gain or loss of one copy of each gene on a single chromosome (CD event). For example, a configuration of four genes with the first two genes on the same chromosome might evolve in a single mutational event to by an SD event or to by a CD event. We propose Algorithm 1, provided in Figure 10, to calculate the minimum number of steps required to transform into considering SD and CD events, where, without loss of generality, we assume that the genes on a common chromosome have consecutive indices in . Algorithm 1 also identifies a minimum-length sequence of events, although this sequence is not necessarily unique. For example, if there are four genes on one chromosome and we want to get from configuration to configuration , then a shortest sequence of SD and CD events would be CD to , SD to , SD to , and SD to . Other orders of the same four events are also possible.
The above example focuses on a single chromosome because as explained below, the problem of finding the shortest SD+CD path can be solved one chromosome at a time. We begin by establishing the following lemmas:
We now extend the theory from the prior section to include SD, CD, and GD events. We assume in the proofs and discussion below that , where denotes lexicographical ordering. This assumption reduces the number of cases in several proofs. If instead, , the proofs are identical or symmetric except that GD events may be used in the wrong direction (halving instead of doubling). The use of halving events is corrected heuristically by a procedure of subtree pruning and regrafting at line 24 of the pseudocode of Algorithm 3, described below, and in FISHtrees. We will produce the complete proof by deriving a series of lemmas for three cases that together will cover all possible and :
We provide an upper bound on the runtime of Algorithm 2 as a function of the number of genes and their copy numbers. Considering all three events, where , the maximum number of doublings required is , where denotes the copy number of the first gene where and . At each stage of the algorithm, the maximum number of nodes generated as a result of a operation is . SD and CD events are used to create each of those nodes in the case of an odd configuration. So, the maximum number of required operations is . Therefore, the number of operations performed during the execution of Algorithm 2 is .
We implemented Algorithm 2 and integrated it with our approximate median-joining-based algorithm from our prior SD-only FISHtrees [46] code. The key steps of this algorithm are summarized in Algorithm 3 (Figure 12), which we describe at a high level here. The phylogeny algorithm first relies on Algorithm 2 to derive a matrix of pairwise distances between observed cell configurations, which are treated as states on a truncated integer lattice of dimension with a maximum value (UB) set to 9 in the current code. It then repeatedly samples triplets of nodes, identifying as potential Steiner nodes those that agree in each dimension with at least one of the triplet. Those Steiner nodes that lead to reduced minimum spanning tree cost are added to the node set, with the process is repeated until there is no further improvement. Finally a series of post-processing steps are performed to prune Steiner nodes that are not needed for the final tree and to apply subtree regrafting to correct for a potential source of suboptimality arising from the fact that the core phylogeny algorithm assumes symmetric distances but GD operations are asymmetric.
Neighbor Joining (NJ) and Maximum Parsimony (MP) methods have been commonly used for building single-tumor phylogenies [16], [54] and we therefore compared their accuracy to that of our own methods in inferring copy number phylogenies. We applied these two traditional phylogenetic tree building methods to build tumor progression trees using the individual copy number profiles as taxa and compared them with the trees built using our algorithms. We used implementations of both approaches in MEGA version 6 [69]. For NJ, we used Euclidean distances between cell copy number profiles to build the pairwise distance matrix. For MP, we treated copy number profiles of the genes in individual cells as sequences of arbitrary phylogenetic characters. We used the “Close-Neighbor-Interchange on Random Trees” search method. For the parameters “Number of Initial Trees” and “MP search level”, we used values of and respectively.
|
10.1371/journal.pcbi.1002560 | Complementarity of Spike- and Rate-Based Dynamics of Neural Systems | Relationships between spiking-neuron and rate-based approaches to the dynamics of neural assemblies are explored by analyzing a model system that can be treated by both methods, with the rate-based method further averaged over multiple neurons to give a neural-field approach. The system consists of a chain of neurons, each with simple spiking dynamics that has a known rate-based equivalent. The neurons are linked by propagating activity that is described in terms of a spatial interaction strength with temporal delays that reflect distances between neurons; feedback via a separate delay loop is also included because such loops also exist in real brains. These interactions are described using a spatiotemporal coupling function that can carry either spikes or rates to provide coupling between neurons. Numerical simulation of corresponding spike- and rate-based methods with these compatible couplings then allows direct comparison between the dynamics arising from these approaches. The rate-based dynamics can reproduce two different forms of oscillation that are present in the spike-based model: spiking rates of individual neurons and network-induced modulations of spiking rate that occur if network interactions are sufficiently strong. Depending on conditions either mode of oscillation can dominate the spike-based dynamics and in some situations, particularly when the ratio of the frequencies of these two modes is integer or half-integer, the two can both be present and interact with each other.
| We develop and demonstrate a model that allows us to examine how the predictions of spiking and rate-based models of neurons and their interactions are related. First, the behavior of a chain of neurons is explored by simulating each spiking neuron and spike-mediated interactions between neurons individually. Second, the same chain is studied using approximations based on the firing rate of the neurons. The predictions for these two approaches are closely compared and it is found that the simpler, rate-based approach captures the major system behaviors of the spike-based approach, namely spiking rates and modulations in those rates. Strong interactions between these modes take place when the frequency of one mode is an integer or half-integer multiple of the frequency of the other mode.
| The brain is a multiscale system, whose dynamics spans from microscale structures, such as ion-channels and synapses, to emergent behavior, such as oscillations at the whole-brain scale. The problem then arises of how to simultaneously incorporate these diverse scales to make predictions about brain dynamics.
Neuronal dynamics has most often been studied by starting from single-neuron perspective via Hodgkin-Huxley equations [1] and their many variants for different neural types (e.g., [2], [3]), or via idealized models such as integrate-and-fire and binary neurons. Strong nonlinearities are responsible for spiking, with the spike cycle often described in terms of a nonlinear oscillator [4], [5]. Such approaches have been extremely successful in accounting for neural dynamics at the single- or few-neuron level.
Single-neuron approaches can also be applied to networks of many neurons by incorporating their synaptic interconnections. While very large networks can be simulated if sufficient computer power is available [3], [6], [7], the results of brute-force simulations can be difficult to interpret, especially when emergent network-level phenomena are involved. Moreover, common misconceptions that arise from the single-neuron viewpoint sometimes impede understanding of large-scale dynamics. For example, the starting-point picture of spiking being due to a nonlinear oscillator often leads to a focus on coupled-oscillator descriptions of neural interactions. If overemphasized, this can obscure the existence of (often linear, or near-linear) collective modes of oscillation in the network, which modulate spike rates at frequencies that are not related to the spike rate itself [8]–[10] — in general, both nonlinear-spiking and collective-oscillation phenomena exist. Some widespread errors in the literature that stem from this standpoint (when adopted naively) are: (i) that large-scale brain rhythms and electroencephalographic (EEG) oscillation frequencies must correspond to spike rates of specific neural “generators” or “pacemakers”, whereas they are quite different from spike rates in general, and (ii) that brain rhythms and EEG oscillations must be highly nonlinear because spikes are, whereas collective oscillations that modulate firing rates can actually be linear, or very nearly so [10]. Of course, collective oscillations can also have their own large-amplitude nonlinearities that survive averaging over spike generation, or arise through other effects [11]–[13].
An alternative starting point is to average over neural properties at the outset to obtain a neural field theory (NFT) [10], [14] in which the average dynamics of large numbers of neurons are modeled. In this case, instantaneous local firing rates are tracked, but individual neuronal spike dynamics are not. Such approaches are well suited to studying large-scale phenomena and bridging across scales and are much less computationally intensive than corresponding studies based on direct computation of single-neuron dynamics. However, as noted, they do not directly incorporate spiking dynamics of individual neurons.
Two aspects are of particular significance here. One is the internal dynamics of neurons. In this study, this is discussed in terms of a comparison between spike events described by changes in membrane potential (in the spike-based approach) and spike rates (in the rate-based approach). Communication between neurons is also critical. In the spike-based approach spikes travel between neurons that are coupled pairwise or via a field that carries spike profiles [14]; in the neural field theory communication is through propagation of fields that carry the spike rate only. In this work, we examine two limiting cases, one in which spiking neurons communicate via spikes, and one in which populations of neurons with rate-based internal dynamics communicate via rates — and make the dynamics as similar as possible in all other respects by having the same type of field carry either spike profiles or spike rates in the respective cases. In other words, one case involves spiking dynamics of neurons coupled by spikes carried by fields, and the second involves rate dynamics of continuous neural matter coupled by rates carried by fields; the fields obey the same propagation equations in both cases.
It is important to understand the relationships between the two limiting approaches, especially because they are complementary, not mutually exclusive. It is thus essential to understand when each is appropriate to be used, whether there are phenomena to which both can be applied, and which is the more convenient and tractable in given cases. Moreover, there can be situations where a fuller understanding requires an application of both approaches. This is analogous to situations arising in many other branches of science. For example, the properties of materials can be studied from a molecular viewpoint but, when dealing with large numbers of molecules, statistical approaches or continuum approximations are more convenient and appropriate starting points for obtaining understanding at the scales of most relevance — hydrodynamics is usually studied in terms of fluids, not molecules, for example. Likewise, statistical mechanics of particles passes over into thermodynamics for many applications as the number of particles becomes large, and there are intermediate regimes that can be addressed using either formalism, or variants such as nonequilibrium thermodynamics.
Some work toward understanding the complementarity of spiking and mean-field approaches has been done, in part by developing hybrid models that preserve aspects of both single-neuron and mean-field approaches. For example, Robinson et al. [15] and Wu et al. [16] showed how to write the spike rate of Wilson neurons [2] in terms of the spike rate itself (rather than instantaneous cellular voltages), thereby eliminating the need to track individual spikes if rate is all that is desired. This work put the Wilson model of spiking and bursting neurons [2] in a form suitable for incorporation into NFT and allowed top-down systems-level influences on single neurons to be analyzed tractably. The predictions of this NFT were subsequently investigated for a model system incorporating a simple delayed feedback loop whose resonances could interact with natural neural spiking and bursting frequencies [16]. Robinson and Kim have very recently developed a series of hybrid methods of treating neural interactions that combine various aspects of spike- and rate-based neural dynamics and of the discrete vs. mean-field features of spatial coupling [14]. Bressloff and Coombes have shown how fluctuations in firing rates consistent with a neural field model can be produced by a network of integrate-and-fire neurons particularly when slow interactions are present [17].
There are other approaches to neuronal modeling which we mention for completeness. The population density approach (e.g. [18]) moves beyond a model based on mean firing rates alone by considering the changes in the distribution of neuron properties. In the population density approach, individual neurons or groups of neurons are not modeled explicitly, rather the change in the probability density function of the state of the neurons is modeled. This approach can be many times faster than a direct Monte Carlo simulation of neurons or groups of neurons. One can also focus on the correlations and higher-order moments of a distribution. This is of significance since correlations in activity may form a significant part of the mechanism through which neuronal signals carry information. Recently, Touboul and Ermentrout [19] have studied the correlation approaches of Bressloff [20] and Buice et al. [21] and shown them to be equivalent when applied to a system of infinite size. This allows large networks of neurons to be analyzed. Significantly, by considering correlations rather that just mean firing rates, dynamical behaviors can appear that cannot be accounted for with a lowest order mean field approach alone.
However, the specific aim of this paper is to investigate and elucidate the complementarity between spike- and rate-based approaches to neural dynamics by use of an overarching approach that can accommodate both pictures in the analysis of a test system that is suited to exposing the key phenomena. Although other approaches may also be informative, we focus on the complementarity between spike- and rate-based simulations in the current work. We begin by reviewing the theoretical background and developing our model. We then present the numerical methods, and give the results of our analyses. We analyze, compare and contrast the dynamics of the spike-based and rate-based approaches. Finally, we interpret the results and discuss their applicability and significance. For simplicity, homogeneous models are used; however, the methods discussed are generalizable to inhomogeneous situations.
In this section we briefly outline the NFT equations required, specializing the treatment to a specific, idealized test system. The model we use is that of a single cortical population, driven by an external drive and incorporating direct interactions between neurons and indirect ones via a delayed feedback loop.
We consider the system of interconnected neurons which includes synaptic input to a set of neurons (labeled by a suffix ) from an external set of neurons (suffix ). The former set consists of a one-dimensional chain of neurons with periodic boundary conditions, and has a feedback both directly and via a loop , where and are the rates of incoming spikes at each synapse (i.e. have dimensions of inverse time), and is time. The loop features a feedback delay time , and the feedback is assumed to be topographically organized (i.e., each point in space feeds back most strongly to itself). This idealized system is sufficiently general to study complementarity between rate- and spike-based treatments; it is also easily generalized to include more types of neurons and higher dimensionality [22], [23]. Biologically, such topographical feedback is found in the thalamocortical loop. Excitatory neurons in the cortex drive the coupled thalamocortical and thalamic reticular neurons of the thalamus; in turn the thalamocortical neurons project back to the cortex in a manner such that a signal returns very close to where it originated [24]. The spiking model is summarized in graphical form in Fig. 1.
When applied to real brain tissue, NFT averages neural properties over linear scales of a few tenths of a millimeter, sufficient to embrace many neurons [10], [22]. The soma potential , measured relative to its resting potential, responds to spikes via synaptic dynamics, dendritic signal dispersion, and soma capacitance. The resulting response to synaptic input approximately obeys [10], [22], [25], [26](1)where(2) is the mean response rate of to synaptic input, is the mean connectivity strength to neurons of type from those of type , is the corresponding mean number of synaptic connections, and (with dimensions voltage times time) is the mean strength of these connections, defined to be the time integral of the postsynaptic potential change due to a spike afferent on a neuron from one of type .
Action potentials are produced at the axonal hillock when the soma potential exceeds a threshold [10], [22], [25]. When averaged over a local population of neurons, a good approximation to the firing rate is(3)where , is the maximum firing rate, and and are the population mean and standard deviation of the threshold [10], [22], and is the mean soma potential averaged over a local population of neurons. We discuss the origins of this relationship below.
Prior work has shown that the mean fields of axonal signals, , , and , propagate approximately as if governed by damped wave equations [22], [27], [28], one form of which is(4)where(5)(6)where , with the axonal velocity, and the characteristic axonal range [22], [27].
Equations (4)–(6) incorporate spatiotemporal coupling between neurons. This is more easily seen through the corresponding Green-function (i.e., propagator) formulation [22], [26]:(7)(8)for , with the Green function satisfying for to ensure causality, and where translation invariance of the system has been assumed. We choose the form (8), which follows from (5) and (6), to preserve the timing and shape of narrow pulses in one dimension, since we want to compare spike-based coupling with field-theoretic coupling in the present work.
The spatial coupling vs. is found by integrating (8) over , [22] which gives(9)Integrating (9) over yields a normalization of unity, which reflects the fact that each pulse that enters an axon ultimately reaches its end. Equations (4)–(6) thus represent signals that propagate along axons at a uniform velocity , but where the number of axons reaching a distance decays exponentially as a function of , with characteristic range . This is a reasonable first approximation to the coupling of cortical neural populations by axons in one dimension. If we were to replace (6) by , we would recover the form introduced in by Robinson et al. [22], which yields broader temporal pulses in response to a delta input. The latter form is actually more realistic in general, especially in two dimensions, since axons are neither identical in velocity nor exactly straight, thereby making delta-function propagation of a mean pulse field very much an idealization. Here we retain (6)–(8) to obtain (9), which is commonly assumed in spike-based analyses. Moreover, this form provides a more stringent test of complementarity with rate-based analyses because it involves no temporal smoothing of the propagated signal, which would tend to make the two cases more similar.
Delayed integrodifferential equations such as these have been well studied, both in general and in the context of neuronal modeling [29]–[31]. The presence of delayed feedback leads to Hopf bifurcations and other dynamic phenomena such as traveling waves [32]. We expect to see such features in the models discussed here.
We now briefly review the Hindmarsh-Rose fast-spiking neuron model [2], [33], [34], and how it can be put in a form compatible and comparable with NFT. Conductance-based equations for the rate of change in membrane potential in a single fast-spiking neuron, appropriate to the mammalian neocortex, can be written [2], [14]–[16], [25], [33], [34](10) is the capacitance per unit area, is an externally imposed input current per unit area (e.g., due to synaptic input from other neurons), is a leakage current per unit area, and and are the and currents per unit area, respectively, and is a transient potassium current that enables these neurons to fire at very low spike rates when is small. Note that use of the script font indicates a voltage measured relative to the extracellular fluid (i.e., a membrane potential) rather than a measurement taken relative to the resting state — there is a constant offset between and equal to the resting potential ; i.e., . Each of the currents is assumed to obey Ohm's law, with(11)where is the conductivity per unit area and is the equilibrium potential of the ion .
Numerous authors have investigated Eqs (10) and (11) for fast-spiking neurons, the main population in the mammalian neocortex, and have found simplified expressions for their dynamics, which can be closely approximated by just two equations that gave an adequate description of spiking dynamics [2], [33]–[36]. There is one equation for the membrane voltage and one for a dimensionless recovery variable that describes the coupled opening of channels and corresponding closure of channels [2], [34]. The equations are(12)(13)where(14)(15)with being the reversal potential, the reversal potential, , , , , , , , , and for fast-spiking neurons.
The dynamics of (12) and (13) have been discussed in detail elsewhere (e.g., [2]), so we summarize very briefly here. At low they have three steady-state solutions: at these are a stable node with and , an unstable saddle point at somewhat higher , and an unstable spiral point at still higher . The first of these represents the resting (non-firing) state. As increases, the two lower fixed points approach one another, then generate a saddle-node bifurcation when they coalesce at the critical current with and . This gives rise to a limit cycle that encircles the resulting spiral point. Each orbit of the limit cycle corresponds to the generation of one spike; hence the picture of spike generation being due to a nonlinear oscillator.
The frequency of the limit cycle (i.e., the firing rate) satisfies [5], [15], [37](16)for and for , which corresponds to a continuous increase from zero firing rate as increases beyond . Simulations show in Eq. (16) [2], [15], [34].
We next show that we can couple individual model spiking neurons together in a way that can be compared directly with NFT of the same system. In NFT, the mean membrane potential of a population of cells is driven by the incoming axonal pulse rate. However, in the spike theory, membrane potential is driven explicitly by current entering the cell body from the dendritic tree. Standard cable equations imply that this current is proportional to minus the spatial derivative of the voltage at the soma boundary. Hence, the functional form of the driving current to a cell induced by a delta function spike at a synapse has the same temporal dependence as , apart from a dimensional constant of proportionality [15], [25]. Thus, it obeys(17)where is the time course of the part of the afferent signal that is above the channel opening threshold and is defined in Eq. (2). This can be approximated as(18)in NFT notation, where the quantities () have units of conductance per unit area and the connection strengths have been introduced. These incorporate the membrane conductance per unit area and will be used in the model to control the relative strengths of the direct and loop feedback, and external drive. Henceforth, we discuss the model in terms of connection strengths rather than the . This formulation allows communication between neurons via the intermediate fields (), which can propagate spike profiles, not just average rates, provided we now replace (4) by(19)where is a constant which we determine shortly. That (19) reproduces (4) can be seen by averaging (19) over timescales much longer than a spike width. Explicitly, averaging Eq. (19) over the inter-spike interval (which varies as a function of space and time) gives(20)where the angle brackets denote the average over . Since changes over time-scales much longer than a spike, we can write , leaving(21)If(22)then it is clear that(23)where is the spike rate. Here we have assumed that the integral over is only significant within the vicinity of the spike, i.e. can be caluclated from Eq. (22) by considering a stereotypical spike profile. We assume (22) henceforth.
In dealing with rates in populations of neurons the idealized square root form (16) of the response curve discussed earlier must be convolved with a distribution (e.g., a Gaussian) of some width that encapsulates fluctuations in the properties of the neurons and their input: e.g., variations in number and strength of synaptic connections, and in the various channel conductances, especially from neuron to neuron. Such convolutions smear (16) over a width [15]. A further source of broadening is fluctuation in arrival rate of spikes and associated changes in membrane voltage [38]. A good approximation that also captures saturation effects is the sigmoidal function(24)which is equivalent to the rate-voltage relationship (3) via(25)where [15] is a conductance per unit area and .
We now in a position to write explicitly a set of coupled differential equations for our 1D chain of identical neurons, in a form that is consistent with NFT in the relevant limit. For each neuron at a point in space, we use the Wilson neuron model to describe its membrane potential and recovery variable . We emphasize here that the spikes are carried through a field rather than through pairwise interactions, which corresponds to the neuron-in-cell approach recently introduced by Robinson and Kim [14].
The input to the neuron comes from both synaptic input from other neurons (through a current term , which is explicitly modeled below), and the input from the external drive term, labeled . The set of coupled differential equations is now obtained from (12) and (13) for the neural dynamics, (17) and (18) for the synaptic dynamics, and (19) for the propagation of fields along axons. To model a level of random external inputs, a white noise current density term is added to the neural dynamics on a grid, where and with a constant. The resulting equations are(26)(27)(28)(29)(30)where is a constant external drive. The variables , , , , and describe the state of the system, with distinguishing the locations of the neurons. Values of constants used in this paper are mostly taken from previous work [2], [23], [39] or, in the cases of , , and , numerical analysis of the Wilson model neuron [2], and are listed in Table 1. The level of noise, through the parameter , is chosen so that fluctuations are small and linear approximations are valid when used. Here we have separated the direct and loop feedbacks (29) and (30), respectively, to enable the use of , and in general.
Before we discuss the numerical implementation of the equations we emphasize that we have not proved that there are well-behaved solutions to these. However, wave equations are well understood physically and numerically; e.g. [22], [27], [28], [40]. Moreover, numerical simulations as discussed below produce results that do not diverge with time. In numerical implementation of the model, the Eqs (23)–(27) are discretized on a 1D spatial grid. When spatially discretizing, several issues must be considered: (i) we must ask whether the numbers of neurons and system size are sufficiently large to ensure results adequately represent real brain dynamics and are not numerical artifacts. The suitability of the number of neurons can be estimated by asking the question of how many input spikes are needed to generate an output spike. In the human brain this is large, with each neuron receiving input of order 10 spikes per second at each of thousands of synapses. Overall, if the effective soma integration time leading to a spike is s, several hundred presynaptic input spikes contribute to each postsynaptic spike [41]. In our simulations with , each neuron is locally coupled to neighboring neurons in approximately 8 cm of tissue. By placing neurons approximately 0.25 cm apart, this gives us locally coupled neurons. Typically in the simulations neurons fire at a rate of spikes per second, giving spikes arriving at each neuron per second, implying that each spike is generated as a result of a neuron receiving input spikes during the relevant integration time. This is much lower than in the cortex; however, computational demands, which scale linearly with the number of neurons, necessitate the use of relatively few neurons. However, we have also carried out some larger runs with considerably more sampled spikes, to begin to explore the effects of relaxing this limitation. No qualitative difference is observed, suggesting that our levels of temporal and spatial discretization are sufficient. (ii) We also anticipate that a system that is too small would introduce artifacts: e.g., with periodic boundary conditions if the system is too small, long-wavelength modes of activity are not captured. Moreover, a model that is of order or smaller in size would be affected by wrap-around of connections through the periodic boundary conditions. However, biologically, it should be remembered that the cortex is not of infinite size; the ratio of to cortical radius is approximately 0.61. A system size of 20 cm is used for most runs; this is adequate in terms of removing numerical artifacts and computer resources and does not represent an implausible size biologically. Some simulations have been carried out with a larger system size and results are not significantly different.
Initially, the variables , and are assigned the value zero for all spatial points. The membrane potential and recovery variable are assigned the values they would have at equilibrium when no external current is applied, namely and 0.279 respectively. The equations are integrated forward in time with a second-order stochastic predictor-corrector method [40]. In order to generate initial activity a high driving current is applied for the first second of simulation and then removed. The Courant condition requires that the time step must be smaller than the grid spacing divided by the velocity of a pulse to ensure numerical stability [42]. The typical step size of s is comfortably within this limit.
We also treat the system of Fig. 1 using the complementary neural field approach of coupling neurons using rate of firing, rather than individual spikes. These rates are propagated using the same Green functions (and same wave equations) as for spikes in the spike-based approach, but individual spikes are not tracked.
In the NFT approach, each grid point is taken to represent the average dynamics of a local population of neurons. To do this we replace the equations for the membrane potential (26) and recovery variable (27) with a single equation that relates the firing rate to the input current, via the square-root function (16), whose parameters were calibrated to reproduce the dynamics of fast-spiking neurons in previous work [14]–[16]. This rate is used to provide input to the wave equation, rather than using the potential term explicitly. A small amount of white noise is added to the current, where , , where is a constant. The noise provides a small perturbation to the system to allow it to quickly explore phase space and ensure that no two simulations are identical. Therefore we obtain the following nonlinear set of four coupled equations for the variables , , and .(31)(32)(33)(34)Numerically, this set of neural field equations can be integrated forward in time using the same approach as for the spike-based case. In this case, the time step can be made larger than for the spike based model, although subject to the Courant condition for numerical stability, because the spike profiles are not modeled explicitly. This is a major advantage of field-based approaches over spike-based approaches.
Equation (31) is appropriate rather than Eq. (24) since for simplicity we will consider homogeneous parameters. Equation (31) allows us to compare explicitly the firing rates predicted by the neural-field approach to those of the spike-based approach. If inhomogeneous parameters were used, Eq. (24), with values of , and specific to the parameter distribution used, would be appropriate.
Equations (31)–(34) can be used to compute the firing rate at various points in space and time; i.e., the mean firing rate of all neurons in the vicinity of each grid point vs. time. Hence, for this to give a good representation of average dynamics, each grid point should correspond to multiple neurons. In the present case, this means the separation of grid points in the neural field model must be much larger than the length scale between neurons, which is satisfied in the present work. Therefore, in carrying out detailed comparisons between the spiking model and the field model, the results of the spiking model need to be coarse-grained (i.e., averaged over the appropriate length scale). We again emphasize that in this work we have carried out simulations at various length scales and neuron densities, and results are qualitatively unaltered by changing the scale (i.e. our discretization is fine enough for the purposes of this work).
It is found that the system (31)–(34) has at least one spatially uniform equilibrium state, which is obtained by setting all the temporal and spatial derivatives to zero. In general there may be one or three solutions (plus a special case of two solutions); however if as in this work, there is only a single solution. Equations (33) and (34) can then be solved to obtain equilibrium values , , and so via Eq. (32):(35)Equation (31) gives(36)for , and otherwise. Squaring Eq. (36) and substituting Eq. (35) for gives a quadratic equation for that is easily solved for the positive firing rate solution.
By writing the deviations from their equilibrium values of , , , , and the noise input in terms of their Fourier components in both space and time, we can establish the power spectrum of fluctuations in both temporal frequency and spatial frequency . To calculate these quantities, we linearize Eqs (31)–(34) in small deviations from equilibrium, and write the Fourier form(37)with similar expressions for , , and . We also note that the noise has an equilibrium value of zero and omit the from henceforth. This gives us the linearized equations(38)(39)(40)(41)These equations can be solved for (or any other of the variables) in terms of the noise input to give us , where the transfer function is given by(42)
In general, Eq. (42) is difficult to analyze further analytically, especially because of the term . However, a useful limiting case can be seen when the system has only loop feedback whose time delay is much longer than the timing of the synaptic current pulses and wave events; i.e., , and . In this case we can make the approximations and to give us the transfer function at :(43)The response (43) will have a resonance when the phase of the complex exponential is a multiple of , which gives resonances at angular frequencies of and its harmonics [16]. When both direct and loop feedback are present, these resonances modulate the combined spectrum to produce peaks, as found originally by Robinson et al. [43], [44].
It might appear that use of a neural field model, where only spike rates are calculated, might remove all information about individual spike times; however, this is not the case [14]. Neural field theory yields instantaneous spike rates as functions of position and time, so the integral of the local rate over some time period is the expected number of spikes that occur at that location in this time period; i.e.,(44)Moreover, when the integral increments by one, we know that there must be exactly one spike during this interval, so(45)defines the expected time at which the next spike occurs, given that one occurred previously at . To construct a membrane potential time series from the spike timings , where is an integer, one can write the potential as a function of the noninteger part of the integral (44). Therefore,(46)where is the integer part of and is a function that describes the spike profile. If required, this profile can be quickly computed from a look-up table [14]. Since this approach requires only the tracking of rather than spike profiles, it is computationally less intense than a spike-based approach, since larger time steps can be used. Implementing this approach requires initial phases to be specified for the neurons, or for enough time to pass that the system loses memory of its initial conditions.
Before presenting the results, we emphasize that in all cases fields are used to describe propagation of signals between neurons, either carrying spikes from neuron to neuron through Eqs. (29) and (30) or conveying rates fields between spatial locations through Eqs. (33) and (34).
Simulation of the spike-based equations (26)–(30) generates output of each of the state variables as a function of position and time. Particularly useful is the membrane potential from which the times of firing of the neurons can be readily extracted. A plot of membrane potential vs. space and time gives an immediate representation of the system dynamics (e.g., synchronous firing, bursting, traveling waves of activity). In the NFT case, simulation of Eqs. (31)–(34) generates output for the state variables; the time series of the membrane potential can then be reconstructed by the method described above.
Also useful is a Fourier space representation of the results, which enables robust identification of wave modes and, in particular, firing rates. One can in principle apply a Fourier transform in space and time to any one of the five state variables , , , and (for the spike-based case) or the four state variables , , and and the reconstructed potential (for the NFT case). In this work we concentrate on the variables and . The former is most directly related to an experimentally measureable quantity, namely the membrane potential. The disadvantage of using is that the highly nonlinear spike features lead to high frequencies in the spectrum that can mask the subtleties of subthreshold fluctuations. The latter is chosen since it is temporally the smoothest of the state variables and so its Fourier transform contains fewer features due to the nonlinearities and thus is most suitable for comparison with a linearized calculation.
The utility of comparing rate- and spike-based approaches via analysis of or depends on the primary mode of behavior of the system. Where the spike-based model shows a spike-dominated behavior (e.g. spiking at a constant frequency) a Fourier analysis based upon provides a meaningful comparison with the predictions from the NFT; where a rate-based oscillation dominates (e.g. spike rate fluctuates or depends upon time delay) a more appropriate comparison would be with the NFT predictions for .
Typically, simulations are run for a total time of 20 seconds. For the first second, a high external drive current is used, as this is sometimes required to initiate spiking in the system; after this time the drive current is removed. Typically, the first four seconds of each time series are discarded to exclude initial transients, the remaining time is split into short periods (typically 4 seconds). Each period is windowed by applying a Hamming window, then the spectrum of or is calculated, as appropriate. The spectra are averaged over all the windows to produce a final power spectrum or . For the case of ,this can be compared with the power spectra predicted by the mean field result Eq. (42). In order to show the effect of individual model parameters on the results, we also show plots of the breathing-mode power spectrum [i.e., or ] for various values of each such parameter of interest. A further analysis is the evaluation of the spatial correlation function , which is given by the inverse Fourier transform of or for the cases of current density and voltage, respectively.
Before exploring the parameter dependences of the model in detail, we first show a typical case, by way of illustration. In later subsections, the results of the model are illustrated with a variety of different cases. In particular, we compare the predictions of the spike-based analysis with the neural field approach to highlight similiarities and differences in behavior. We illustrate the change in behavior of the model system as a function of the key parameters (time delay , external drive current , and direct and loop connection strengths and ) by keeping all but one parameter constant, and varying the others. We also present a comparison of the spiking events from the spike based model and a reconstruction from the neural field model.
To start, we demonstrate typical behavior of the spike-based state variables , , and . For this illustration we use a small positive loop feedback; i.e., , . The external drive current is chosen to be equal to the critical current . This external current would put an individual neuron at the point of spiking, so that the positive feedback between neurons ensures that they obtain a modest spike rate; this allows us to explore the interaction between spike-based and collective oscillations. Biologically, it is reasonable that a neural system can organize to be near a critical point [45], [46]. A very short time delay is used, s so that delays between the direct and delayed feedbacks are negligible compared to the timescales of the dominant neural activity in this case (i.e., the interspike interval). Fig. 2 shows the membrane potential of each neuron over a typical 1 second period. The spike events are clearly shown, indicating a spike rate of about . There is clearly evidence of spatial structure in the firing pattern, which we elucidate through the spectrum below.
The current is plotted in Fig. 3. It is much more smoothly varying than the spiking voltage. However, firing events can still be discerned via their associated rapid increases in versus time, meaning that a firing rate, as opposed to modulation in rate, will be the most obvious feature on any spectrum.
The variables and are shown in Fig. 4. In order to show the spatio-temporal structure of these fields, only a small part of the spatio-temporal domain is shown here. The scales are different for Figs. 3 and 4. These variables denote the propagation of signals between neurons. The plots show signals emanating from each firing event as ‘’-shaped features. The apex corresponds to the firing event, while the two arms of the ‘’ show the propagation of the signal forward in time at constant speed in both spatial directions. The gradient of the arms of the ‘’ for the direct feedback term is , corresponding to the signal speed given by ; likewise, the gradient of the features for the loop feedback is , which equals . It is also clear from the length of the arms that the direct feedback events in Fig. 4(a) have a longer spatial range than the loop-mediated ones in Fig. 4(b), in accord with cm and cm here.
We now examine the power spectra for and . To complete comparisons, we have carried out simulations for the spike-based model, Eqs. (26)–(30), the NFT, Eqs. (31)–(34), and evaluated the theoretical field prediction through the transfer function of Eq. (42). To consider the effect of the remnants of spike features on , we also have constructed the power spectrum of a series of stereotypical spike features in , which can be added to the NFT predictions of . To illustrate these spectra, we have carried out simulations for the case of , , and = 0.06 s. In Figs. 5 and 6 we show results for analyses of the current density term and membrane potential , respectively.
In Fig. 5, the four rows, in order, represent analyses of the spike-based model, the simulations of the NFT equations, the theoretical analysis of the spectrum of the NFT through Eq. (42), and a theoretical analysis of NFT as for the third row but augmented with spike features. The three columns represent the breathing mode power , the spatial correlation function [from the inverse Fourier transform of ] and the full spatio-temporal power spectrum . Panel A shows for the spike-based model. The power is large at zero frequency and falls with increasing frequency; however, it is dominated in this case by features related to the spike rate; namely peaks at about 22 Hz and its harmonics. Panel B shows the correlation function, showing that there is some significant spatial order in the system; with decaying to in about 2 cm. Panel C shows ; here we see that there are large features at 22 Hz and 44 Hz associated with spiking behavior superposed on a more smoothly varying background with a maximum at (0,0). The spatial frequency extent of these features, about , is equivalent to the correlation length seen in panel B. Panels D–F show the equivalent for the NFT simulation. The obvious difference is the lack of spike-features, since the NFT simulation does not contain spiking events. Otherwise, the shape (but not the magnitude) of the behavior is very similar to that of panels A–C. Panel G shows a theoretical calculation of from Eq. (42); it is evident that it is very similar to that of the simulation of panel D. Panel H shows the spatial correlation function; it has a similar correlation length to those of B and E; however, it does not have the same minimum at approximately 0.07 m that is the case for panels B and E. This negative correlation in panels B and E may be attributable to the toroidal boundary conditions in space. Panel I shows , which agrees with the simulation of panel F and the background of the panel C for the spike-based model. Panel J depicts calculated from Eq. (42), with the addition of spike features arising from the spectrum of spikes. This compares well qualitatively to Panel A. The major discrepancy is the magnitude of the power. This is due to the interplay between the spike-based mode and the rate-based mode. The rate-based oscillation influences the synchrony of the spike-based mode, thus magnifying the power when resonances occur, such as for this set of parameters. The size of the major resonance can therefore vary tremendously as a function of . Panel K shows the NFT spatial correlation function, and panel L the NFT prediction of to which has been added the power spectrum due to a series of spike remnants in . Panel L compares moderately well with panel C; the major discrepancy is the greater extent of the resonant features in , corresponding to less synchronization of neurons than is seen in the spike-based simulations in panel C. Overall, for Fig. 5, when features attributable to spikes are taken into account, we note that the NFT theory and simulation generally predict well the underlying shape of the power spectra (though not its magnitude).
In Fig. 6 we show a similar analysis for the membrane potential. The first row represents the simulation of the spike-based model; the second the reconstruction of a spike sequence from the simulations of the NFT model. Note that there is no NFT linearized prediction in this case since the NFT theory does not consider explicitly. The three columns represent the breathing mode power , the spatial correlation function (from the inverse Fourier transform of ) and the full spatio-temporal power spectrum . Panel A shows for the spike-based model. It is dominated by features related to the spike rate; namely peaks at about 22 Hz and its harmonics. Fluctuations due to non-spike (e.g. subthreshold) processes are much lower in magnitude. Panel B shows the spatial correlation function, showing, as in Fig. 5, decaying to in about 2 cm. Panel C shows ; here we see that there are large features at 22 Hz and 44 Hz associated with spiking behavior; the spatial frequency extent of these (around ) approximately equals the inverse of the spatial correlation length. Panels D–F show the equivalent for the NFT simulation, in which spikes have been generated through the process described earlier. One notes that Panel D shows a similar (but not exactly identical) spectrum to Panel A; for example, the spike rates are slightly different and the spikes are less broad, consistent with less variation in inter-spike interval. Panel E shows that there is no discernable correlation between neighboring neurons in this method. Panels F shows the spectrum of the reconstructed spike sequence; the swaths at the spike rates represent very distinct firing frequencies that are uncorrelated in space. There is no equivalent NFT theoretical prediction since the NFT does not contain spiking events explicitly. Overall, for Fig. 6, we note that the NFT theory and simulation predict temporal structure of spiking well, but are not as accurate spatially. This is attributed to the problem of defining initial conditions from the reconstruction of spikes. One could in principle, knowing the result of the spike-based approach, define initial phases to produce a similar correlation. We have not done this. Spike rates also show more fluctuation in the spike-based model than the NFT model.
To summarize, we observe with Figs. 5 and 6 that the analysis of shown in Fig. 5 is more appropriate for analyzing the correspondence between the spike-based model and NFT where there is specific interest in the behavior of the NFT model (e.g. where collective modes dominate behavior). However, the latter analysis of is likely to be appropriate when spike-rates are the dominant issue to consider.
We now consider how the behavior of the models change as key parameters are varied. To do this, we carry out simulations of both sets of equations (26)–(30) and (31)–(34) for the spike-based and NFT models, respectively; and use the methods of [14] to reconstruct spike-sequences from the NFT prediction. The power spectra or as appropriate for both situations are then compared. Plots of the power spectrum against temporal frequency are then stacked to represent graphically the changes in resonances and power fluctuations in response to a variation in a parameter.
Fig. 7 demonstrates the effect of a change in the external driving current . Higher naturally leads to a higher firing rate. Part A shows the breathing mode power for the spike based model as a function of drive current. There is an abrupt change in the spectrum at ; the maximum of shifts from 5.5 Hz to 11 Hz. From this point, the major frequency feature increases in frequency as drive current increases, corresponding to a mean firing rate of the neurons that is in agreement with neural field theory. Part B shows the predictions of the neural field model in terms of the power spectrum of the reconstructed voltage trace. The two graphs show that the resonances occur at similar frequencies, with a trend of increasing frequency with increasing . However, these frequencies are not exactly the same. For example, at , the spike-based approach in Panel A gives a simulated rate of 11 Hz, whereas in the NFT model shown in Panel B a rate of 9 Hz is observed. At , the spike-based and field-based frequencies are 17 Hz and 14 Hz respectively. In this case the spike-based model is dominated by the regular spiking behavior predicted by NFT. Note that the lumpy structure is a result of the resolution limit of the plot; it is not a chain of discrete peaks.
In Fig. 8 we show the effect of varying the time delay for the case of a high external drive current and fairly high loop feedback . Since is well above there is a high firing rate and the feedback is not required to maintain activity. Direct feedback is set to zero. The temporal frequency spectrum at is plotted against time delay. Also shown on the plot are the predictions of resonances from the NFT, through Eq. (42). These are shown by the solid lines. The half-integer multiples of these are shown by the dashed lines. The plot clearly indicates a firing rate of around 22 Hz. A harmonic at around 44 Hz is also present on the plot. However, the firing rate is not completely independent of , and varies between approximately 21 Hz and 22 Hz. There are two clear regions, at about and when the power is very large, indicating a sharp resonance in activity of the breathing mode, a property of the network.
A feature of this plot is the dependence on of the magnitude of the resonance at the spike-rate of 21–22 Hz. A large response occurs when the firing frequency is that of the fundamental predicted by Eq. (43). A low response occurs when the firing rate is exactly double that of the prediction of Eq. (43). We emphasize that in this case synaptic coupling between a neuron and its neighbors is weak compared with the driving current, implying that each neuron has a well established limit cycles for its firing, dominated by . The reason for the discrepancy between the predicted resonances of Eq. (42) and the resonances seen in simulation appears to be the loss of spatial synchrony of the neurons at the time delays predicted by Eq. (42) to be resonances (i.e. the solid lines of Fig. 8). Instead of near synchronous firing, more intricate spatial patterns, e.g. traveling waves [47], [48], are formed causing a reduction in . This phenomenon is not seen when spike rates are significantly reduced by lowering , as discussed below.
Next, the effect of a wide range of time delays is demonstrated for a low firing case. In order to elucidate the interaction between the loop resonances captured by the mean field approach and the effects of spike firings, a small loop connection strength has been chosen, with no direct feedback, and a drive current equal to the critical current. This ensures that any positive feedback will result in significant activity. This time, the appropriate analysis is with the power in the current fluctuations, . Part A of Fig. 9 shows the breathing mode power as a function of time delay , for a simulation of the spike based equations. Part B shows the same plot, but as predicted by the rate-based theory through Eq. (42).
One can see three clear regimes in Fig. 9A. For , the major feature on the power spectrum is a resonance at about 10 Hz, which is double the neuron spike rate. However, there is a hint of power at 5 Hz at . An examination of the neural firing patterns shows that there are strong correlations between neighboring neurons (e.g., a neuron firing at 5 Hz out of phase with its neighbor) leading to the 10 Hz feature being more prominent than the 5 Hz one. For , the major feature is the resonance induced by the delay loop. The neurons adopt a firing rate that is equivalent to the resonance frequency predicted by the NFT (i.e., the resonances of Fig. 9B). Harmonics of this frequency are also clearly visible. The key result in this graph is that the firing rates seen on Panel A for the spike-based model when are in the positions predicted by NFT, while for the NFT resonance is not strong enough to capture this mode and the firing rate reverts to that of the spike-based resonance.
A close-up of part of Fig. 9A is shown in Fig. 10, where the solid lines show the loop frequency prediction of Eq. (43); the two are very closely related. At there is an abrupt change in the power spectrum and for the lowest frequency peak falls in magnitude as increases until it vanishes at . Here, the neurons are no longer able to fire slowly enough to follow the loop resonance frequency which is low for large , and instead the firing rate switches to (nearly) double the loop resonance frequency. However, this transition is subtle and Fig. 10 shows a very slight downward shift in the frequency compared to double the loop frequency.
The above behavior is similar to that found for a population-based neuron model with loop feedback [16]. In that model the authors found that their system could jump between two regimes of behavior as the time delay was varied. In one regime, the system fired with a rate equal to the reciprocal of the time delay, or an integer multiple of this frequency (i.e., it decreased as the time delay increased); in the other regime, it produced a firing rate independent of the time delay. The system alternated between these regimes as the delay time increased. In our model we also see this break between a firing rate roughly independent of time delay (for ), and one where the rate approximately follows the reciprocal of the delay time (for ).
The other major parameters that can be changed are the connection strengths. We illustrate this case by studying the effect of altering the balance between the direct and loop connection strengths and , respectively.
The mean field solution for firing rate depends upon the sum of and . However, fluctuations in firing rate are expected to be different. By setting , we ensure that the equilibrium firing rate is the same in all cases and we can study how the resonances and stability change as the balance between the terms changes. The other key parameters are selected as and .
Fig. 11 shows the breathing mode power as a function of the direct (non-loop) connection strength . In Panel A power is plotted in the form of contours; Panel B shows a three-dimensional representation of the same information. Note that in the plot, , implying that the loop connection strength is negative. The most distinctive feature in this plot is a bifurcation at , as a result of a strong loop negative feedback through . At , the system oscillates at about 5 Hz between a rapidly firing state and a non-firing state. For , the system fires at about 12 Hz, as predicted by NFT. Close to bifurcation the system experiences large fluctuations in firing rate, as expected by NFT [49]. There is some evidence of an increase in power at about 5 Hz just before the bifurcation, for . This part of the spectrum is indicated explicitly on both parts A and B of Fig. 11. The fluctuations for are shown explicitly in Fig. 12 which shows against time and space for one second of time. The plot shows propagating fronts of activity. The velocity of propagation has a range of approximately 0.4– and a typical value of around , and this variation results in the firing rate of each neuron showing considerable fluctuation with time.
The point of bifurcation is also presented in more detail through Fig. 13. Panel A shows ; we see a decrease in power with increasing frequency, with a hint of a peak at around 4 Hz. Panel B shows the spatial correlation function ; there is long-range order here with dropping to in about 3 cm. Panel C shows the for the spike based simulation. There is a background of activity that peaks at (0,0); on top of this there is a diagonal line of peaks with gradient of approximately ; corresponding to a traveling wave with velocity of about , consistent with the typical velocity of a wavefront in Fig. 12.
This is significantly larger than the loop propagation speed illustrating that the rate of propagation of activity along axons is not the sole determining factor for the wavefront velocity. Indeed, Bressloff has demonstrated that propagation of waves in a one-dimensional network of integrate-and-fire neurons is dispersive and dependent upon synaptic strength and delays in addition to axonal properties [50]. Panels D, E and F show the equivalent calculations from the NFT through Eq. (42). In panel D we see a power spectrum with a strong peak at 3 Hz (similar to the peak of panel A) and then clear resonances at higher frequencies. Panel E shows the correlation function ; there is large long range order predicted, consistent with being on the edge of an instability. Panel F shows the predicted . One can see a peak at about 3 Hz and zero spatial frequency, similar to Panel C for the spike-based simulation. However, the major feature of this plot are the deep dips lying on a line of gradient . It is interesting to remark that this gradient is about half that seen for the resonances in Panel C. In these plots it is clear that the spike-based simulations and linearized NFT predictions are considerably different. This is not surprising given that this system lies very close to an instability where linear predictions are expected to break down. It is also possible that the critical point in NFT might be at a slightly different parameter value than for the spiking model.
We have explored the relationships between spiking-based and rate-based neural models by using both approaches to model the same test system — a one dimensional array of neurons coupled both directly and via a delayed feedback loop. The dynamics predicted by both approaches has been compared predominantly through the power spectra of the membrane potential and current density . We have focused on the relationships between resonances associated with the firing of single cells and populations of cells, particularly the overlap and transitions between these two regimes. We have shown how the dynamics, especially prominent resonant effects, depend on the key parameters of the model (specifically delay-loop time, loop connection strengths and drive current density).
The spike-based approach of Eqs (26)–(30) supports two modes of oscillation. First, there is the highly nonlinear spiking mode in which each neuron spikes according to its input. This mode, for a single Wilson neuron, has been well-studied [2], [49], [51]. Also, a collective mode can exist, in which the firing rate undergoes small or large oscillations. The spectrum of these oscillations can be determined through neural field methods [as in Eq. (42)]. Both modes can be obtained through an analysis of an equivalent neural field model: the spiking modes from a reconstruction of the voltage, and demonstrated most directly through ; the collective modes through an analysis of the current fluctuations and demonstrated most directly using .
At this point we again stress the distinction between the firing rate and fluctuations in the firing rate. Neural field theory predicts both, namely the firing rate itself, and how the firing rate fluctuates with time and space. In an extreme case, this could take the form of bursting — a neuron fires rapidly for a short period of time, and then is silent for a period of time. Thus there are two different time scales here, the inverse of the firing rate, and the period for the bursting–silent oscillation. Generally, however, the collective modes are not of this extremely nonlinear bursting form, but can be modeled by the linear analysis of Eqs. (35)–(42).
Results of spike-based and NFT simulations and predictions can be compared through plots of the power spectrum in current density, and membrane potential, . In the latter case, since the membrane potential does not feature in the NFT equations explicitly, a sequence must be reconstructed from knowledge of other variables, either the mean firing rate through time integration or the current density through a neuron-in-cell method [14]. Analysis of , and particularly the breathing mode power proves useful when there is significant power in the collective modes of oscillation; however, when spike-based behavior dominates, an analysis of gives more direct insight. A disadvantage of analyzing the membrane potential is that the reconstruction of spike sequences from the NFT solution does not produce the spatial correlations that are predicted through or seen in the spike-based model. This is because in the reconstruction of a voltage from a firing rate, [e.g., Eq. (46)] spatial effects manifest themselves differently; the reconstructed spike series depends upon the initial conditions which are not known a priori as a function of space.
The collective and spike-base modes are not entirely independent of each other, particularly when the two time scales are the same (or integer or half-integer multiples) of each other. Indeed, Wu et al. [16] found for a rate-based model of a Wilson neuron that receives feedback from itself, the behavior can be dominated by either type of resonance. At certain points a small change in parameters is sufficient to cause the behavior to switch between one resonance and the other. In this spatial model, we see similar behavior. However, the spatial dimension adds a complexity to the behavior that is not present in simpler models. This manifests itself for example in the intricate traveling-wave firing patterns that are demonstrated in Figs. 2 and 12 for example, that have been described by Osan et al. [47].
In some cases, the collective and spike-based modes are both present in a system, without significant signs of interaction. For example, Fig. 5 shows that the NFT predicts the underlying power spectrum, on which features due to spiking events sit. For example, in this case there is a peak at zero temporal and spatial frequency, as is frequent for neural field models away from a resonance condition (e.g., [39]). However, there are situations where one of the two modes can dominate In Fig. 7 the spiking mode dominates; neurons fire at a rate that is close to that predicted by the NFT but show little modulation in this rate — i.e. there is little collective oscillation present.
In Fig. 9A, and in more detail in Fig. 10, for , the collective oscillation dominates to the extent that it captures the spiking oscillation. This can be seen from the correspondence between the solid lines in Fig. 10 and the regions of high power. In this case the firing rate is no longer that predicted by the equilibrium value of in Eq. (31), specifically 5.5 Hz, but is equal to the position of the predicted resonance in the rate. Specifically, the larger , the lower the resonant frequency. We emphasize that this is not a trivial result. Resonances in spike rate as predicted by NFT (i.e. the peaks in the spectrum ) are not in general the same as the mean firing rate itself. Analysis of Fig. 9 requires the use of both the spike-based and rate-based paradigms in order to fully elucidate the results. Analysis of the spiking pattern (not shown) shows that the system is also highly synchronized in space — in other words the collective oscillation has entrained the spiking oscillation. This is consistent with the prior result that a particular firing-rate based model agreed with a corresponding integrate-and-fire spiking model when interactions were sufficiently slow [17]. In Fig. 9B, the predicted power for the NFT case shows clearly the NFT resonances moving to lower frequency as increases. It is also interesting to note the transition between different dominant modes of behavior in this system. For , the strength of the collective oscillation is not sufficient to entrain the firing rate, which reverts to its equilibrium predicted value of 5.5 Hz. The major feature on Fig. 9A is at twice this, 11 Hz; analysis of the firing patterns (not shown) show that neurons are approximately paired; each fires in approximate anti-phase with its neighbor. Such behavior is common in neural simulations and its prevalence depends upon the strength of coupling between neurons, randomness in the couplings, noise and time delays. An anti-phase mode would be most likely when coupling strength and randomness are low [17], [52], noise is low [48], but for a small range of time delays [52], [53]. In our simulations we have not used random connections and have kept connection strengths low in order to ensure firing rates are of similar magnitude to resonances in the neural field simulations. Both of these favor the existence of an anti-phase state.
Fig. 9A is similar to the results seen by Wu et al. [16] in which a simpler spatially uniform model was shown to exhibit similar transitions between spike-based modes and collective modes of behavior as the loop delay was changed. In Ref. [16], spiking rates for one parameter set entrained alternately to either one of the two modes as was increased. Several switches between the modes were observed as delay time was increased from 0 to 0.7 s. In the current study, delay times were limited to what is physical within a thalamocortical system, and only a single switch between a spike-based mode and a collective mode is observed. For a different parameter set, Wu et al. [16] also demonstrated a doubling of the primary frequency of oscillation with as a result of a small change in time delay, similar to the doubling observed in this study in Fig. 9A.
There can also be more complicated interplay between the two forms of oscillation. For example Fig. 8 demonstrates that the power in fluctuations in at a particular frequency can be strongly influenced by the relationship between this frequency (in this case 22 Hz) and the frequency of a collective resonance predicted by NFT shown in the figure by the solid and dashed lines. This variation in power requires particular comment. Naively, one might expect that where an NFT resonance, with positive gain, corresponds to the mean spike rate, there would be an enhancement of power. However, the opposite is the case here; the power is much reduced when the resonance predicted by Eq. (43) corresponds to the spike rate. There are two points to discuss by way of explanation. First, resonances predicted by Eq. (43) are weak when , so interaction between the two might be expected to be small for this parameter range. Second, the spatial nature of the model is important. Consider the propagation of activity in space. When a neuron fires, there is a time frame over which an effect is generated on the axon, through Eq. (30). There is then a delay time for the signal to traverse the thalamocortical loop, followed by a time for an impact to be felt on the receiving neuron through Eq. (28). A signal therefore takes a time to return to the same neuron in the cortex; longer times allow for spatial propagation via the delay loop. At a spike rate of each neuron receives strong input that arose from itself one spike-interval in the past, and consequently spatial communication between neurons is relatively weak. The weak communication encourages the formation of a variety of spatially patterned states [17], [47], [52] which are not synchronized in space and therefore lead to a reduction in . Such a pattern of alternating synchronous and asynchronous behavior has been found in previous studies [17].
It is possible that the dynamics of the spike based system is unduly influenced by the homogeneity of the parameters used [52]. For realistic systems, one would expect a wide range of values for the axonal lengths, synaptic decay times, etc. A system consisting of identical neurons may be particularly sensitive to modes of oscillation (e.g., synchronous in-phase or antiphase firing of all neurons) that are less likely to be seen in practice. Using homogeneous values in a model has the advantage of increased analytic tractability; however, implications of such homogeneity require further study. The methods discussed are easily generalizable to two dimensions with appropriate choice of wave equation (6). Results would be less easy to present, since two spatial dimensions and one temporal dimension would be present. Other neuron models (e.g. the bursting model used by Robinson et al. [15]) could be used by changing the forms of Eqs. (12)–(15), and finding the equivalent rate equation (16). An inhibitory population could be added with another set of variables.
To conclude, we remark that we have demonstrated considerable overlap between the spike-based and the neural-field approaches. Where neural-field resonances are strong, spiking rates can be entrained to these resonances. A system that allows both modes to feature can show interactions between the two. Both spike-based and rate-based paradigms must be used to fully analyze the system. A spatial dimension adds complexity to the situation discussed previously by Wu et al. [16]. The theories can be considered as complementary methods of approaching the neural modeling problem, each offering a different physical emphasis.
|
10.1371/journal.pcbi.1002653 | Model-Free Reconstruction of Excitatory Neuronal Connectivity from Calcium Imaging Signals | A systematic assessment of global neural network connectivity through direct electrophysiological assays has remained technically infeasible, even in simpler systems like dissociated neuronal cultures. We introduce an improved algorithmic approach based on Transfer Entropy to reconstruct structural connectivity from network activity monitored through calcium imaging. We focus in this study on the inference of excitatory synaptic links. Based on information theory, our method requires no prior assumptions on the statistics of neuronal firing and neuronal connections. The performance of our algorithm is benchmarked on surrogate time series of calcium fluorescence generated by the simulated dynamics of a network with known ground-truth topology. We find that the functional network topology revealed by Transfer Entropy depends qualitatively on the time-dependent dynamic state of the network (bursting or non-bursting). Thus by conditioning with respect to the global mean activity, we improve the performance of our method. This allows us to focus the analysis to specific dynamical regimes of the network in which the inferred functional connectivity is shaped by monosynaptic excitatory connections, rather than by collective synchrony. Our method can discriminate between actual causal influences between neurons and spurious non-causal correlations due to light scattering artifacts, which inherently affect the quality of fluorescence imaging. Compared to other reconstruction strategies such as cross-correlation or Granger Causality methods, our method based on improved Transfer Entropy is remarkably more accurate. In particular, it provides a good estimation of the excitatory network clustering coefficient, allowing for discrimination between weakly and strongly clustered topologies. Finally, we demonstrate the applicability of our method to analyses of real recordings of in vitro disinhibited cortical cultures where we suggest that excitatory connections are characterized by an elevated level of clustering compared to a random graph (although not extreme) and can be markedly non-local.
| Unraveling the general organizing principles of connectivity in neural circuits is a crucial step towards understanding brain function. However, even the simpler task of assessing the global excitatory connectivity of a culture in vitro, where neurons form self-organized networks in absence of external stimuli, remains challenging. Neuronal cultures undergo spontaneous switching between episodes of synchronous bursting and quieter inter-burst periods. We introduce here a novel algorithm which aims at inferring the connectivity of neuronal cultures from calcium fluorescence recordings of their network dynamics. To achieve this goal, we develop a suitable generalization of Transfer Entropy, an information-theoretic measure of causal influences between time series. Unlike previous algorithmic approaches to reconstruction, Transfer Entropy is data-driven and does not rely on specific assumptions about neuronal firing statistics or network topology. We generate simulated calcium signals from networks with controlled ground-truth topology and purely excitatory interactions and show that, by restricting the analysis to inter-bursts periods, Transfer Entropy robustly achieves a good reconstruction performance for disparate network connectivities. Finally, we apply our method to real data and find evidence of non-random features in cultured networks, such as the existence of highly connected hub excitatory neurons and of an elevated (but not extreme) level of clustering.
| The identification of the topological features of neuronal circuits is an essential step towards understanding neuronal computation and function. Despite considerable progress in neuroanatomy, electrophysiology and imaging [1]–[8], the detailed mapping of neuronal circuits is already a difficult task for a small population of neurons, and becomes impractical when accessing large neuronal ensembles. Even in the case of cultures of dissociated neurons, in which neuronal connections develop de novo during the formation and maturation of the network, very few details are known about the statistical features of this connectivity, which might reflect signatures of self-organized critical activity [9]–[11].
Neuronal cultures have emerged in recent years as simple, yet versatile model systems [12], [13] in the quest for uncovering neuronal connectivity [14], [15] and dynamics [16]–[19]. The fact that relatively simple cultures already exhibit a rich repertoire of spontaneous activity [18], [20] make them particularly appealing for studying the interplay between activity and connectivity.
The activity of hundreds to thousands of cells in in vitro cultured neuronal networks can be simultaneously monitored using calcium fluorescence imaging techniques [14], [21], [22]. Calcium imaging can be applied both in vitro and in vivo and has the potential to be combined with stimulation techniques like optogenetics [23]. A major drawback of this technique, however, is that the typical frame rate during acquisition is slower than the cell's firing dynamics by an order of magnitude. Furthermore the poor signal-to-noise ratio makes the detection of elementary firing events difficult.
Neuronal cultures are unique platforms to investigate and quantify the accuracy of network reconstruction from activity data, extending analysis tools initially devised for the characterization of macro-scale functional networks [24], [25] to the micro-scale of a developing local circuit.
Here we report a new technique based on information theory to reconstruct the connectivity of a neuronal network from calcium imaging data. We use an extension of Transfer Entropy (TE) [26]–[28] to extract a directed functional connectivity network in which the presence of a directed edge between two nodes reflects a direct causal influence by the source to the target node [29]–[31]. Note that “causal influence” is defined operationally as “improved predictability” [32], [33] reflecting the fact that knowledge of the activity of one node (putatively pre-synaptic) is helpful in predicting the future behavior of another node (putatively post-synaptic). TE has previously been used to study gene regulatory networks [34], the flow of information between auditory neurons [35], to infer directed interactions between brain areas based on EEG recordings [36] or between different LFP frequency bands [37], as well as for the reconstruction of the connectivity based on spike times [38], [39]. Importantly, our data-driven TE approach is model-independent. This is in contrast with previous approaches to network reconstruction, which were most often based on the knowledge of precise spike times [40]–[45], or explicitly assumed a specific model of neuronal activity [43], [44].
A problem inherent to the indirect algorithmic inference of network connectivity from real data is that the true target topology of the network is not known and that, therefore, it is difficult to assess the quality of the reconstruction. In order to characterize the behavior of our algorithm and to benchmark its potential performance, we resort therefore to synthetic calcium fluorescence time series generated by a simulated cultured neural network that exhibits realistic dynamics. Since the “ground truth” topology of cultures in silico is known and arbitrarily selectable, the quality of our reconstruction can be evaluated by systematically comparing the inferred with the real network connectivities.
We use a simplified network simulation to generate surrogate imaging data, improving their realism with the reproduction of light scattering artifacts [6] which ordinarily affect the quality of the recording. Our surrogate data also reproduce another general feature of the activity of neuronal cultures, namely the occurrence of temporally irregular switching between states of asynchronous activity, with relatively weak average firing rates, and states of highly synchronous activity, commonly denoted as “network bursts” [20], [46], [47].
This switching dynamics poses potentially a major obstacle to reconstruction, since directed functional connectivity can be very different during bursting and inter-burst phases and can bear a resemblance to the underlying structural (i.e. synaptic) connectivity only in selected dynamical regimes in which causal influences reflect dominantly mono-synaptic interactions. To restrict our analysis to such “good” regimes, we resort to conditioning with respect to the averaged fluorescence level, as an indirect but reliable indicator of the network collective dynamics. Appropriate conditioning —combined with a simple correction coping with the poor time-resolution of imaging data— allows the method to achieve a good topology reconstruction performance (assessed from synthetic data), out-performing other standard approaches, without the need to infer exact spike times through sophisticated techniques (as is required, on the contrary, in [43], [45]).
Finally, we apply our algorithm —optimized through model-based validation— to the analysis of real calcium imaging recordings. For this purpose, we study spontaneously developing networks of dissociated cortical neurons in vitro and we address, as a first step toward a full topology reconstruction, the simpler problem of extracting only their excitatory connectivity. Early mature cultures display a bursting dynamics very similar to our simulated networks, with which they also share an analogous state-dependency of directed functional connectivity.
Our generalized TE approach thus identifies network topologies with characteristic and non-trivial features, like the existence of non-local connections, a broadened and strongly right-skewed distribution of degrees (although not “scale free”) and a moderate but significant level of clustering.
The Results section is organized as follows. After a brief presentation of the qualitative similarity between real calcium fluorescence data from neuronal cultures and simulated data (see Figure 1), we introduce numerical simulations showing that networks with very different clustering levels can lead to matching bursting dynamics (see Figure 2). We then develop our reconstruction strategy, based on a novel generalization of TE, and examine the different elements composing our strategy, namely “same-bin interactions” and conditioning with respect to the average fluorescence level. We show that only signals recorded during inter-burst periods convey elevated information about the underlying structural topology (see Figure 3). After a discussion of criteria guiding the choice of the number of links to include in the reconstructed network, we illustrate specific examples of reconstruction (see Figure 4 and 5), contrasting systematically TE with other standard linear and nonlinear competitor methods (see Figure 6) and analyzing factors affecting its performance (see Figure 7). Finally, we apply our reconstruction algorithm to biological recordings and infer topological features of actual neuronal cultures (see Figure 8).
In this study, we consider recordings from in vitro cultures of dissociated cortical neurons (see Materials and Methods). To illustrate the quality of our recordings, in Figure 1A we provide a bright field image of a region of a culture together with its associated calcium fluorescence. As previously anticipated, to simplify the network reconstruction problem, experiments are carried out with blocked inhibitory GABA-ergic transmission, so that the network activity is driven solely by excitatory connections. We record activity of early mature cultures at day in vitro (DIV) 9–12. Such young but sufficiently mature cultures display rich bursting events, combined with sparse irregular firing activity during inter-burst periods (cfr. Discussion).
In Figure 1B (left panel) we show actual recordings of the fluorescence traces associated to five different neurons. The corresponding population average for the same time window is shown in Figure 1C (left panel). In these recordings, a stable baseline is broken by intermittent activity peaks that correspond to synchronized network bursts recruiting many neurons. The bursts display a fast rise of fluorescence at their onset followed by a slow decay. In addition, during inter-burst periods, smaller modulations above the baseline are sometimes visible despite the poor time-resolution of a frame every few tens of milliseconds.
In order to benchmark and optimize different reconstruction methods, we also generate surrogate calcium fluorescence data (shown in the right panels of Figures 1B and 1C), based on the activity of simulated networks whose ground truth topology is known. We simulate the spontaneous spiking dynamics of networks formed by excitatory integrate-and-fire neurons, along a duration of 60 minutes of real time, matching typical lengths of actual recordings. Calcium fluorescence time series are then produced based on this spiking dynamics, resorting to a model introduced in [43] and described in the Materials and Methods section.
Although over 1000 cells are accessible in our experiments, we observed in the simulations that neurons suffice to reproduce the same dynamical behavior observed for larger network sizes, while still allowing for an exhaustive exploration of the entire algorithmic parameter space. Furthermore, despite their reduced density, we maintain in our simulated cultures the same average probability of connection as in actual cultures, where this probability (, see Materials and Methods) is an estimate based on independent studies [15], [48].
The fluorescence signal of a particular simulation run or experiment can be conveniently studied in terms of the distribution of fluorescence amplitudes. As shown in Figure 1D for both simulations and experiments, the amplitude distributions display a characteristic right-skewed shape that emerge from the switching between two distinct dynamical regimes, namely the presence or absence of bursts. The distribution in the low fluorescence region assumes a Gaussian-like shape, corresponding to noise-dominated baseline activity, while the high fluorescence region displays a long tail with a cut-off at the level of calcium fluorescence of the highest network spikes. As we will show later, qualitative similarity between the shapes of the simulated and experimental fluorescence distributions will play an important guiding role for an appropriate network reconstruction.
Neurons grown in vitro develop on a two-dimensional substrate and, hence, both connectivity and clustering may be strongly sensitive to the physical distance between neurons. At the same time, due to long axonal projections [13], [49], excitatory synaptic connections might be formed at any distance within the whole culture and both activity and signaling-dependent mechanisms might shape non-trivially long-range connectivity [50], [51].
To test the reconstruction performance of our algorithm, we consider two general families of network topologies that cover a wide range of clustering coefficients. In a first one, clustering occurs between randomly positioned nodes (non-local clustering). In a second one, the connection probability between two nodes decays with their Euclidean distance according to a Gaussian distribution and, therefore, connected nodes are also likely to be spatially close. In particular, in this latter case, the overall level of clustering is determined by how fast the connection probability decays with distance (local clustering). Cortical slice studies revealed the existence of both local [52], [53] and non-local [7], [54] types of clustering.
We will later benchmark reconstruction performance for both kinds of topologies and for a wide range of clustering levels, because very similar patterns of neuronal activity can be generated by very different networks, as we now show.
Figure 2 illustrates the dynamic behavior of three networks (in this case from the non-local clustering ensemble). The networks are designed to have different clustering coefficients but the same total number of links (see the insets of Figure 2B for an illustration). The synaptic coupling between neurons was adjusted in each network using an automated procedure to obtain bursting activities with comparable bursting rates (see Materials and Methods for details and Table 1 for the actual values of the synaptic weight). As a net effect of this procedure, the synaptic coupling between neurons is slightly reduced for larger clustering coefficients. The simulated spiking dynamics is shown in the raster plots of Figure 2A. These three networks display indeed very similar bursting dynamics, not only in terms of the mean bursting rate, but also in terms of the entire inter-burst interval (IBIs) distribution, shown in Figure 2B. In the same manner, we constructed and simulated local networks —with a small length scale corresponding to high clustering coefficients and vice versa— and obtained the same result, i.e. very similar dynamics for very different decay lengths (not shown).
We stress that our procedure for the automatic generation of networks with similar bursting dynamics was not guaranteed to converge for such a wide range of clustering coefficients. Thus, the illustrative simulations of Figure 2 provide genuine evidence that the relation between network dynamics and network structural clustering is not trivially “one-to-one”, despite the fact that more clustered networks have been shown to have different cascading dynamics at the onset of a burst [42].
We focus, first, on the reconstruction of simulated networks, taken from the local and non-local ensembles described above. We compute their directed functional connectivity based on simulated calcium signals. Synthetic fluorescence time series are pre-processed only by simple discrete differentiation, such as to extract baseline modulations associated to potential firing. These differentiated signals are then used as input to any further analyses.
Immediately prior to the onset of a burst the network is very excitable. In such a situation it is intuitive to consider that the directed functional connectivity can depart radically from the structural excitatory connectivity, because local events can potentially induce changes at very long ranges due to collective synchronization rather than to direct synaptic coupling. Conversely, in the relatively quiet inter-burst phases, a post-synaptic spike is likely to be influenced solely by the presynaptic firing history. Hence, the directed functional connectivity between neurons is intrinsically state dependent (cfr. also [55]), a property that must be taken into account when reconstructing the connectivity.
We illustrate here the state dependency of directed functional connectivity by generating a random network from the local clustering ensemble and by simulating its dynamics, including light scattering artifacts to obtain more realistic fluorescence signals. The resulting distribution of fluorescence amplitudes is divided into seven non-overlapping ranges of equal width, each of them identified with a Roman numeral (Figure 3A). Finally, TE is computed separately for each of these ranges, based on different corresponding subsets of data from the simulated recordings.
For simulated data, the inferred connectivity can be directly compared to the ground truth, and a standard Receiver-Operator Characteristic (ROC) analysis can be used to quantify the quality of reconstruction. ROC curves are generated by gradually moving a threshold level from the lowest to the highest TE value, and by plotting at each point the fraction of true positives as a function of the fraction of false positives. The quality of reconstruction is then summarized in a single number by the performance level, which, following an arbitrary convention, is measured as the fraction of true positives at 10% of false positives read out of a complete ROC curve.
We plot the performance level as a function of the average fluorescence amplitude in each interval, as shown by the blue line of Figure 3B. The highest accuracy is achieved in the lowest fluorescence range, denoted by I, and reaches a remarkably elevated value of approximately 70% of true positives. The performance in the higher ranges II to IV decreases to a value around 45%, to abruptly drop at range V and above to a final plateau that corresponds to the 10% performance of a random reconstruction (ranges VI and VII).
Note that fluorescence values are not distributed homogeneously across ranges I–VII, as evidenced by the overall shape of the fluorescence distribution in Figure 3A. For example, the lowest and highest ranges (I and VII) differ by two orders of magnitude in the number of data points. To discriminate unequal-sampling effects from actual state-dependent phenomena, we studied the performance level using an equal number of data points in all ranges. Effectively, we restrict the number of data points available in each range to be equal to the number of samples in the highest range, VII. The quality of such a reconstruction is shown as the red curve in Figure 3B. The performance level is now generally lower, reflecting the reduced number of time points which are included in the analysis.
Interestingly, the “true” peak of reconstruction quality is shifted to range II, corresponding to fluorescence levels just above the Gaussian in the histogram of Figure 3A. This range is therefore the most effective in terms of reconstruction performance for a given data sampling.
For the ranges higher than II, the reconstruction quality gradually decreases again to the 10% performance of purely random choices in ranges VI and VII. The effect of adopting a (shorter) equal sample size is particularly striking for range I, which drops from the best performance level almost down to the baseline for random reconstruction. As a matter of fact, range I is the one for which the shrinkage of sample length due to the constraint for uniform data sampling is most extreme (see later section on dependence of performance from sample size).
The above analysis leads to a different functional network for each dynamical range studied. For the analysis with an equal number of data point per interval, the seven effective networks are drawn in Figure 3C (for clarity only the top 10% of links are shown). Each functional network is accompanied with the corresponding ROC curve.
The lowest range I corresponds to a regime in which spiking-related signals are buried in noise. Correspondingly, the associated functional connectivity is practically random, as indicated by a ROC curve close to the diagonal. Nevertheless, information about structural topology is still conveyed in the activity associated to this regime and can be extracted through extensive sampling.
At the other extreme, corresponding to the upper ranges V to VII —associated to fully developed synchronous bursts— the functional connectivity has also a poor overlap with the underlying structural network. As addressed later in the Discussion section, functional connectivity in regimes associated to bursting is characterized by the existence of hub nodes with an elevated degree of connection. The spatio-temporal organization of bursting can be described in terms of these functional connectivity hubs, since nodes within the neighborhood of the same functional hub provide the strongest mutual synchronization experienced by an arbitrary pair of nodes across the network (see Discussion and also Figure S2).
The best agreement between functional and excitatory structural connectivity is clearly obtained for the inter-bursts regime associated with the middle range II, and to a lesser degree in ranges III and IV, corresponding to the early building-up of synchronous bursts.
Overall, this study of state-dependent functional connectivity provides arguments to define the optimal dynamical regime for network reconstruction: The regime should include all data points whose average fluorescence across the population is below a “conditioning level” , located just on the right side of the Gaussian part of the histogram of the average fluorescence (see Materials and Methods). This selection excludes the regimes of highly synchronized activity (ranges III to VII) and keeps most of the data points for the analysis in order to achieve a good signal-to-noise ratio. Thus, the inclusion of both ranges I and II combines the positive effects of correct state selection and of extensive sampling.
The state-dependency of functional connectivity is not limited to synthetic data. Very similar patterns of state-dependency are observed also in real data from neuronal cultures. In particular, in both simulated and real cultures, the functional connectivity associated to the development of bursts displays a stronger clustering level than in the inter-burst periods. An analysis of the topological properties of functional networks obtained from real data in different states (compared with synthetic data) is provided in Figure S3. In this same figure, sections of fluorescence time-series associated to different dynamical states are represented in different colors, for a better visualization of the correspondence between states and fluorescence values (for simplicity, only four fluorescence ranges are distinguished).
Our generalized TE, conditioned to the proper dynamic range, enables the reconstruction of network topologies even in the presence of light scattering artifacts. For non-locally clustered topologies we obtain a remarkably high accuracy of up to 75% of true positives at a cost of 10% of false positives. An example of the reconstruction for such a network, with , is shown in Figure 4A. For locally-clustered topologies, accuracy typically reaches 60% of true positives at a cost of 10% of false positives, and an example for is shown in Figure 5A.
In both topologies, we observe that for a low fraction of false positives detection (i.e. at high thresholds ) the ROC curve displays a sharp rise, indicating a very reliable detection of the causally most efficient excitatory connections. A decrease in the slope, and therefore a rise in the detection of false positives and a larger confidence interval, is observed only at higher fractions of false positives. The confidence intervals are broader in the case of locally-clustered topologies because of the additional network-to-network variability that results from the placement of neurons (which is irrelevant for the generation of the non-locally clustered ensembles, see Materials and Methods).
The performance level (fraction of true positives for 10% of false positives, denoted by ) provides a measure of the quality of the reconstruction, and allows the comparison of different methods for different network topologies, conditioning levels, and external artifacts (i.e. presence or absence of simulated light scattering). We test linear methods, XC and GC (of order 2; the performance of GC of order 1 is very similar and not shown), and non-linear methods, namely MI and TE (of Markov orders 1 and 2; see Materials and Methods for details). XC and MI are correlation measures, while GC and TE are causality measures. Note that, for each of these methods, we account for state dependency of functional connectivity, performing state separation as described in the Materials and Methods section.
In the case of the non-local clustering ensemble and without light scattering (Figure 6, top row), even a linear method such as XC achieves a good reconstruction. This success indicates an overlap between communities of higher synchrony in the calcium fluorescence, associated to stronger activity correlations, and the underlying structural connectivity, especially for higher full clustering indices.
GC-based reconstructions have an overall worse quality, due to the inadequacy of a linear model for the prediction of our highly nonlinear network dynamics, but they show similarly improved performance for higher .
In a band centered around a shared optimal conditioning level , both MI and generalized TE show a robust performance across all clustering indices. This value is similar to the upper bound of the range II depicted in Figure 3A, i.e. it lies at the interface between the bursting and silent dynamical regimes. In particular for TE and in the case of low clustering indices (which leads to networks closer to random graphs), conditioning greatly improves reconstruction quality. At higher clustering indices the decay in performance is only moderate for conditioning levels above the optimal value, indicating an overlap between the functional connectivities in the bursting and silent regimes. Note, on the contrary, that the performance of MI rapidly decreases if a non-optimal conditioning level is assumed.
The introduction of light scattering causes a dramatic drop in performance of the two linear methods (XC and GC), and even of MI and TE with Markov order . The performance of TE at Markov order 2 also deteriorates, but is still significantly above the random reconstruction baseline in a broad region of parameters. Interestingly, for the optimal conditioning level the performance of the TE for does not fall below for any clustering level or value. It is precisely in this optimal conditioning range that we obtain the linear relations between reconstructed and structural clustering coefficients, for both the non-local and the local clustering ensembles.
A similar trend is obtained when varying the length scale in the local ensembles (see Supplementary Figure S5). For very local clustering and without light scattering, both XC and TE achieve performance levels up to 80%. The introduction of light scattering, however, reduces the performance of all measures except for MI (but only in the narrow optimal conditioning range) and for TE of higher Markov orders (robust against non-optimal selection of conditioning level). Overall, the performance of the reconstruction for the local clustering ensembles is lower than for the non-locally clustered ensembles. This is also true, incidentally, in absence of light scattering since networks sampled from this ensemble tend to be very similar to purely random topologies (of the Erdös-Rényi type, see e.g. [62]) as soon as the length scale is sufficiently long, and for which performance is generally poorer (cfr. top row of Figure 6, for weak clustering levels).
Our new TE method significantly improves the reconstruction performance compared to the original TE formulation [Eq. (9)]. As shown in Figure 7A for both the local and the non-local clustered networks, reconstruction with the original TE formulation (Figure 7A, blue line) yields worse results than a random reconstruction, as indicated by the corresponding ROC curves falling below the diagonal. Such a poor performance is due in large part to “misinterpreted” delayed interactions. Indeed, by taking into account same bin interactions, a boost in performance is observed (red line). Figure 7A also shows that an additional leap in performance is obtained when the analysis is conditioned (i.e. restricted) to a particular dynamical state of the network, increasing reconstruction quality by 20% (yellow line in Figure 7A). The determination of the optimal conditioning level is discussed later and takes into account the considerations introduced above (cfr. Figure 3).
Note that the introduction of same bin interactions alone (red color curves) or conditioning on the dynamical state of network alone (yellow color curves) already brings the performance to a level well superior to random performance. However, at least for our simulated calcium-fluorescence time series, a remarkable boost in performance is obtained only when the inclusion of same-bin interactions and optimal conditioning are combined together (green color curves). Although, in principle, conditioning is enough to indirectly select a proper dynamical regime, the poor time-resolution of the analyzed signals (constrained not only by the frame-rate of acquisition but also intrinsically by the kinetics of the dissociation reaction of the calcium-sensitive dye [63]) also requires the potential consideration of causally-linked events occurring in the same time-bin.
A different way to represent reconstruction performance are “Positive Precision Curves” (as introduced in [56] and described in the Materials and Methods section), obtained by plotting, at a given number of reconstructed links, the “true-false ratio” (TPR), which emphasizes the probability that a reconstructed link is present in the ground truth topology (true positive). For the same networks and reconstruction as above, we plot the PPCs in Figure S6 (for the ROC curves see Figure 7A). Over a wide range of the number of reconstructed links (TFS), the PPC displays positive values of the TFR, indicative of a majority of true positives over false positives. For both the locally and non-locally clustered networks, the PPC reaches a maximum value of the TPR about 0.5 and remains positive up to about 18% of included links for the clustered topology, or up to 12% in the case of the local topology.
In Figure 7B, we analyze the performance of our algorithm against changes of the sample size. Starting from simulated recordings lasting 1 h of real time (corresponding to about 360 bursting events) and with a full sample number of , we trimmed these recordings producing shorter fluorescence time series with samples, with being a divisor of the sample size. For both network topology ensembles, we found that a reduction in the number of samples by a factor of two (corresponding to 30 minutes or about 180 bursts) still yields a performance level of . By further reducing the sample size, we reach a plateau with a quality of for about 40 bursts (corresponding to 6 minutes).
All the experiments analyzed in this work are carried out with a duration between 30 and 60 minutes. Since conditioning, needed to achieve high performance, requires one to ignore a conspicuous fraction of the recorded data, we expect long recordings to be necessary for a good reconstruction, albeit the fact that it is possible to increase the signal-to-noise ratio by increasing the intensity of the fluorescent light. However, the latter manipulation has negative implications for the health of neurons due to photo-damage, limiting our experimental recordings to a maximum of 2 hours.
We apply our analysis to actual recordings from in vitro networks derived from cortical neurons (see Materials and Methods). To simplify the network reconstruction problem, experiments are carried out with blocked inhibitory GABA-ergic transmission, so that the network activity is driven solely by excitatory connections. This is consistent with previously discussed simulations, in which only excitatory neurons were included.
We consider in Figure 8 a network reconstruction based on a 60 minutes recording of the activity of a mature culture, at day in vitro (DIV) 12, in which active neurons were simultaneously imaged. A fully analogous network reconstruction for a second, younger dataset at DIV 9 is presented in Supplementary Figure S7. In general, fluorescence data neither affected by photo-bleaching nor by photo-damage during this time, as proved by the stability of the average fluorescence signal shown in the Supplementary Figure S8A.
The probability distribution of the average fluorescence signal is computed in the same way as for the simulated data. Neuronal dynamics and the calcium fluorescence display the same bursting dynamics that are well captured by the simulations, leading to a similar fluorescence distribution (Figure 1D). Thanks to this similarity we can make use of the intuition developed for synthetic data to estimate an adequate conditioning level. We select therefore a conditioning level such as to exclude the right-tail of high fluorescence associated to to fully-developed bursting transient regimes. We have verified, however, that the main qualitative topological features of the reconstructed network are left unchanged when varying the conditioning level in a range centered on our “optimal” selection. More details on conditioning level selection are given in the Materials and Methods section.
Reconstruction analysis is carried out for the entire population of imaged neurons. We analyze a network defined by the top 5% of TE-ranked links, as discussed in a previous section. Such choice leads to an average in–degree of about 100, compatible with average degrees reported previously for neuronal cultures of corresponding age (DIV) and density [14], [64].
We have introduced a novel extension of Transfer Entropy, an information theoretical measure, and applied it to infer excitatory connectivity of neuronal cultures in vitro. Other studies have previously applied TE (or a generalization of TE) to the reconstruction of the topology of cultured networks [39], [56]. However, our study introduces and discusses important novel aspects, relevant for applications.
As shown in Figure 3 for simulated data and in Supplementary Figure S3 for actual culture data, epochs of synchronous bursting are associated to functional connectivity with a stronger degree of clustering and a weaker overlap with the underlying structural topology. This feature of functional connectivity is tightly related to the spatio-temporal organization of the synchronous bursts.
In Figure S2A, we highlight the position of selected nodes (of the simulated network considered in Figure 3), characterized by an above-average in-degree of functional connectivity (see Materials and Methods for a detailed definition in terms of TE scores). We denote these nodes —different in general for each of the dynamic regimes numbered from I to VII— as (state-dependent) functional connectivity hubs.
Given a specific hub, we can then define the community of its first neighbors in the corresponding functional network. Consistently across all dynamic ranges (but the noise-dominated range I) we find that synchronization within each of these functional communities is significantly stronger than between the communities centered on different hubs ((, Mann-Whitney test, see Materials and Methods). The results of this comparison are reported in Figure S2B, showing particularly marked excess synchronization during burst build-up (ranges II, III and IV) or just prior to burst waning in the largest-size bursts (VII). Therefore, functional connectivity hubs reflect foci of enhanced local bursting synchrony.
Other studies (in brain slices) reported evidence of functional connectivity hubs, whose direct stimulation elicited a strong synchronous activation [69], [70]. In [69], the functional hubs were also structural hubs. In the case of our networks, however only functional hubs associated to ranges II and III have an in-degree (and out-degree) larger than average as in [69]. In the other dynamic ranges, this tight correspondence between structural and functional hubs does not hold anymore. Nevertheless, in all dynamic ranges (but range I), we find that pairs of functional hubs have an approximately three-times larger chance of being structurally connected than pairs of arbitrarily selected nodes (not shown).
The timing of firing of these strong-synchrony communities is analyzed in Figure S2C. There we show that the average time of bursting of functional communities for different dynamic ranges is shifted relatively to the average bursting time over the entire network (details of the estimation are provided in Materials and Methods). This temporal shift is negative for the ranges II and III (indicating that functional hubs and related communities fire on average earlier than the rest of the culture) and positive for the ranges V to VII (indicating that firing of these communities occurs on average later than the rest of the culture). The highest negative time delay is detected in range III, such that the communities organized around its associated functional hubs can be described as local burst initiation cores [71], [72].
In this study we did not consider inhibitory interactions, neither in simulations nor in experiments (GABAergic transmission was blocked), but we attempted uniquely the reconstruction of excitatory connectivity.
We would like to point out that this is not a general limitation of TE, since the applicability of TE does not rely on assumptions as to the specific nature of a given causal relationship – for instance about whether a synapse is excitatory or inhibitory. In this sense, TE can be seen as a measure for the absolute strength of a causal interaction, and is able in principle to capture the effects on dynamics of both inhibitory and excitatory connections. Note indeed that previous studies [39], [56] have used TE to infer as well the presence of inhibitory connections. However, TE alone could not discriminate the sign of the interaction and additional post-hoc considerations had to be made in order to separate the retrieved connections into separate excitatory and inhibitory subgroups (e.g., in [39], based on the supposedly known existence of a difference in relative strength between excitatory and inhibitory conductances).
In summary, we have developed a new generalization of Transfer Entropy for inferring connectivity in neuronal networks based on fluorescence calcium imaging data. Our new formalism goes beyond previous approaches by introducing two key ingredients, namely the inclusion of same bin interactions and the separation of dynamical states through conditioning of the fluorescence signal. We have thoroughly tested our formalism in a number of simulated neuronal architectures, and later applied it to extract topological features of real, cultured cortical neurons.
We expect that, in the future, algorithmic approaches to network reconstruction, and in particular our own method, will play a pivotal role in unravelling not only topological features of neuronal circuits, but also in providing a better understanding of the circuitry underlying neuronal function. These theoretical and numerical tools may well work side by side with new state-of-the-art techniques (such as optogenetics or high-speed two-photon imaging [96]–[100]) that will enable direct large-scale reconstructions of living neuronal networks. Our Transfer Entropy formalism is highly versatile and could be applied to the analysis of in vivo voltage-sensitive dye recordings with virtually no modifications.
On a shorter time-scale, it would be important to extend our analysis to the reconstruction of both excitatory and inhibitory connectivity in in vitro cultures, which is technically feasible, and to compare diverse network characteristics, such as neuronal density or aggregation. Our algorithm could be used to systematically reconstruct the connectivity of cultures at different development stages in the quest for understanding the switch from local to global neuronal dynamics. Another crucial open issue is to design suitable experimental protocols allowing to confirm the existence of at least some of the inferred synaptic links, in order to validate statistically the reconstructed connectivity. For instance, the actual presence of directed links to which our algorithm assigns the largest TE scores might be systematically probed through targeted paired electrophysiological stimulation and recording. Furthermore, GFP transfection or inmunostaining might be used to obtain actual, precise anatomical data on network architecture to be compared with the reconstructed one. Finally, it might be interesting to reconstruct connectivity of cultured networks before and after physical disconnection of different areas of the culture (e.g. by mechanical etching of the substrate or by chemical silencing). These manipulations would provide a scenario to verify whether TE-based reconstructions correctly capture the absence of direct connections between areas of the neuronal network which are known to be artificially segregated.
We generated synthetic networks with neurons, distributed randomly over a squared area of 0.5 mm lateral size. We chose as the connection probability between neurons [48], leading to sparse connectivities similar to those observed in local cortical circuits [53]. We used non-periodic boundary conditions to reproduce eventual “edge” effects that arise from the anisotropic cell density at the boundaries of the culture.
We considered two general types of networks: (i) a locally-clustered ensemble, where the probability of connection depended on the spatial distance between two neurons; and (ii) a non-locally clustered ensemble, with the connections engineered to display a certain degree of clustering.
For the case of a non-local clustering ensemble, we first created a sparse connectivity matrix, randomly generating links with a homogeneous probability of connection across pairs of neurons. We next selected a random pairs of links and “crossed” them (links and became and ). We accepted only those changes that updated the clustering index in the direction of a desired target value, thereby maintaining the number of incoming as well as outgoing connections of each neuron. The crossing process was iterated until a clustering index higher or equal to the target value was reached. The overall procedure led to a full clustering index of the reference random network of (mean and standard deviation, respectively, across 6 networks). After the rewiring iterations, we then achieved standard deviations from the desired target clustering value smaller than 0.1% for all higher clustering indices.
We measured the full clustering index of our directed networks according to a common definition introduced by [65]:(1)The binary adjacency matrix is denoted by , with for a link , and zero otherwise. The adjacent matrix provides a complete description of the network topological properties. For instance, the in-degree of a node can be computed as , and the out-degree as . The total number of links of a node is given by the sum of its in-degree and its out-degree (). The number of bidirectional links of a given node (i.e. links between and so that and are reciprocally connected by directed connections) is given by .
The adjacency matrix did not contain diagonal entries. Such entries would correspond “toautaptic” links that connect a neuron with itself. Note that our directed functional connectivity analysis is based on bivariate time series, and therefore it would be structurally unfit to detect this type of links.
For the case of the local clustering ensemble, two neurons separated a Euclidean distance were randomly connected with a distance dependent probability described by a Gaussian distribution, of the form , with a characteristic length scale. To guarantee that a constant average number of links was present in the network, this Gaussian distribution was rescaled by a constant pre-factor, obtained as follows. We first generated a network based on the unscaled kernel and computed the resulting number of links . With this value we then generated a final network based on the rescaled kernel .
The dynamics of the generated neuronal networks was studied using the NEST simulator [101], [102]. We modeled the neurons as leaky integrate-and-fire neurons, with the membrane potential of a neuron described by [103], [104]:(2)where is the leak conductance and is the membrane time-constant. The term accounts for a time-dependent input current that arises from recurrent synaptic connections. In the absence of synaptic inputs, the membrane potential relaxes exponentially to a resting level set arbitrarily to zero. Stimulation in the form of inputs from other neurons increase the membrane potential, and above the threshold an action potential is elicited (neuronal firing). The membrane voltage is then reset to zero for a refractory period of .
The generated action potential excites post-synaptic target neurons. The total synaptic currents are then described by(3)where is the adjacency matrix, and is a synaptic time constant. The resulting excitatory post-synaptic potentials (EPSPs) have a standard difference-of-exponentials time-course [105].
Neurons in culture show a rich spontaneous activity that originates from both fluctuations in the membrane potential and small currents in the pre-synaptic terminals (minis). The latter is the most important source of noise and plays a pivotal role in the generation and maintenance of spontaneous activity [106]. To introduce the spontaneous firing of neurons in Eq. (3), each neuron was driven, through a static coupling conductance with strength , by independent Poisson spike trains (with a stationary firing rate of , spikes fired at stochastic times ).
Neurons were connected via synapses with short-term depression, due to the finite amount of synaptic resources [104]. We considered only purely excitatory networks to mimic the experimental conditions in which inhibitory transmission is fully blocked. Concerning the recurrent input to neuron , the set represents times of spikes emitted by a presynaptic neuron , is a conduction delay of , while sets a homogeneous scale for the synaptic weights of recurrent connections, whose time-dependent strength depends on network firing history through the equations(4)(5)In these equations, represents the fraction of neurotransmitters in the “effective state”, in the “recovered state” and in the “inactive state” [103], [104]. Once a pre-synaptic action potential is elicited, a fraction of the neurotransmitters in the recovered state enters the effective state, which is proportional to the synaptic current. This fraction decays exponentially towards the inactive state with a time scale , from which it recovers with a time scale . Hence, repeated firing of the presynaptic cell in an interval shorter than gradually reduces the amplitude of the evoked EPSPs as the synapse is experiencing fatigue effects (depression).
Random networks of integrate-and-fire neurons coupled by depressing synapses are well-known to naturally generate synchronous events [104], comparable to the all-or-none behavior that is observed in cultured neurons [46], [47]. To obtain in our model a realistic bursting rate [47], the synaptic weight of internal connections was set to result into a network bursting of for all the network realizations that we studied, and in particular for any considered (local or non-local) clustering level. Therefore, after having generated each network topology, we assigned the arbitrary initial value of to internal synaptic weights and simulated 200 seconds of network dynamics, evaluating the resulting average bursting rate. If it was larger (smaller) than the target bursting rate, then the synaptic weight was reduced (increased) by 10%. We then iteratively adjusted by (linearly) extrapolating the last two simulation results towards the target bursting rate, until the result was closer than 0.01 Hz to the target value. The resulting used values of are provided in Table 1.
Note that we defined a network burst to occur when more than 40% of the neurons in the network were active within a time window of 50 ms. Such a criterion does not play any role in the reconstruction algorithm itself, where state selection is achieved through conditioning, but is only used for the automated generation of random networks with a prescribed bursting rate. Typically, for a fully developed burst, more than 90% of the neurons fire within a 50 ms bin, while, during inter-burst intervals, less than 10% do. Due to the clear separation between these two regimes, our burst detection procedure does not depend significantly on the precise choice of threshold within a broad interval.
To reproduce the fluorescence signal measured experimentally, we treated the simulated spiking dynamics to generate surrogate calcium fluorescence signals. We used a common model introduced in [43] that gives rise to an initial fast increase of fluorescence after activation, followed by a slow decay (). Such a model describes the intra-cellular concentration of calcium that is bound to the fluorescent probe. The concentration changes rapidly by a step amount of for each action potential that the cell is eliciting in a time step , of the form(6)where is the total number of action potentials.
The net fluorescence level associated to the activity of a neuron is finally obtained by further feeding the Calcium concentration into a saturating static non-linearity, and by adding a Gaussian distributed noise with zero mean:(7)For the simulations, we used a saturation concentration of and noise with a standard deviation of 0.03.
We considered the light scattered in a simulated region of interest (ROI) from surrounding cells. Denoting as the distance between two neurons and and by the scattering length scale (determined by the typical light deflection in the medium and the optical apparatus), the resulting fluorescence amplitude of a given neuron is given by(8)
A sketch illustrating the radius of influence of the light scattering phenomenon is given in Figure S9. The scaling factor sets the overall strength of the simulated scattering artifact. Note that light scattered, according to the equation shown above, could be completely corrected using a standard deconvolution algorithm, at least for very large fields of view and a scattering length known with sufficient accuracy. In a real setup however, the relatively small fields of view (on the order of ), the inaccuracies in inferring the scattering radius , as well as the inhomogeneities in the medium and on the optical system, make perfect deconvolution not possible. Therefore, artifacts due to light scattering cannot be completely eliminated [107], [108]. The scaling factor , that we arbitrarily assumed to be small and with value , can be seen as a measure of this residual artifact component.
In its original formulation [26], for two discrete Markov processes and (here shown for equal Markov order ), the Transfer Entropy (TE) from to was defined as:(9)where is a discrete time index and is a vector of length whose entries are the samples of at the time steps , , …, . The sum goes over all possible values of , and .
TE can be seen as the distance in probability space (known as the Kullback-Leibler divergence [109]) between the “single node” transition matrix and the “two nodes” transition matrix . As expected from a distance measure, TE is zero if and only if the two transition matrices are identical, i.e. if transitions of do not depend statistically on past values of , and is greater than zero otherwise, signaling dependence of the transition dynamics of on .
We use TE to evaluate the directed functional connectivity between different network nodes. In a pre-processing step, we apply a basic discrete differentiation operator to calcium fluorescence time series , as a rather crude way to isolate potential spike events. Thus, given a network node , we define . This pre-processing step also improves the signal-to-noise ratio, thus allowing for a better sampling of probability distributions with a limited number of data points. To adapt TE to our particular problem we need to take into account the general characteristics of the system. We therefore modified TE in two crucial aspects:
(10)
We then include all data points at time instants in which this average fluorescence is below a predefined threshold parameter , i.e. we consider only the time points that fulfill . We only make an exception that corresponds to the simulations of Figure 3 and Figure S2, where we considered time points that fall within an interval bounded by a higher and a lower thresholds, i.e. .
Using these two novel aspects, we have extended the original description of Transfer Entropy [Eq. (9)] to the following form(11)
Probability distributions have to be evaluated as discrete histograms. Hence, the continuous range of fluorescence values (see e.g. the bottom panels of Figure 1) is quantized into a finite number of discrete levels. We typically used a small , a value that we justify based on the observation that the resulting bin width is close to twice the standard deviation of the signal. The presence of large fluctuations, most likely associated to spiking events, is then still captured by such a coarse, almost non-parametric description of fluorescence levels.
Generalized TE values are obtained for every possible directed pair of network nodes, and using a fixed threshold level . The set of TE scores are then ranked in ascending order and scaled to fall in the unit range. A threshold is then applied to the rescaled data, so that only those links with scores above are retained in the reconstructed network.
A standard Receiver-Operator Characteristic (ROC) analysis is used to assess the quality of the reconstruction by evaluating the number of true positives (reconstructed links that are present in the actual network) or false positives (not present), and for different threshold values [110]. The highest threshold value leads to zero reconstructed links and therefore zero true positives and false positives. At the other extreme, the lowest threshold provides both 100% of true positives and false positives. Intermediate thresholds give rise to a smooth curve of true/false positives as a function of the threshold. The performance of the reconstruction is then measured as the degree of deviation of this curve from the diagonal, and that corresponds to a random choice of connections between neurons.
To provide a simple method to compare different reconstructions, we arbitrarily use the quantity , defined as the fraction of true positives for a 10% of false positives, as indicator for the quality of the reconstruction.
An alternative to the ROC is the Positive Prediction Curve [56], plotting the “true-false ratio” (TFR) against the number of reconstructed links, called “true-false sum” (TFS). The TFR represents the fraction of true positives relative to the false positives. Denoting by #TP the absolute number of true positives and by #FP the number of false positives, TFR is therefore defined in [56] in the following way:(12)The case corresponds to the case that, for any given reconstructed link, it is on average equally likely that it is in fact a true positive rather than a false positive.
To gain further insight into the quality of our reconstruction method, we compare reconstructions based on TE with three other reconstruction strategies, namely cross-correlation, mutual information, and Granger causality.
Cross-correlation (XC) reconstructions are based on standard Pearson cross-correlation. The score assigned to each potential link is given by the largest cross-correlogram peak for lags between and , of the form(13)
In a similar way, the scores for Mutual Information (MI) reconstructions are evaluated as(14)Analogously to TE, the sum goes over all entries of the joint probability matrix.
For the reconstruction based on Granger causality (GC) [33] we first model the signal by least-squares fitting of a univariate autoregressive model, obtaining the coefficients and the residual ,(15)In a second step, we fit a second bivariate autoregressive model that includes the potential source signal , and determine the residual ,(16)Note that in the latter bivariate regression scheme we take into account “same bin” interactions as for Transfer Entropy (index of the second sum starts at ). Given , the covariance matrix of the univariate fit in Eq. (15), and , the covariance matrix of the bivariate fit in Eq. (16), GC is then given by the logarithm of the ratio between their traces:(17)GC analyses were performed at an order . Analyses at yielded however fully analogous performance (not shown).
We note that the same pre-processing used for TE is also adopted for all the other analyses. The same holds for conditioning on the value of the average fluorescence , which can be applied simply by only including the subset of samples in which .
Connectivity in reconstructed networks is often inhomogeneous, and groups of nodes with tighter internal connectivity are sometimes visually apparent (see e.g. reconstructed topologies in Figure 3C). We do not attempt a systematic reconstruction of network communities [111], but we limit ourselves to the detection of “causal sink” nodes [112], which have a larger than average in-degree. We define this property in terms of the sum of TE from all other nodes to one particular node (), choosing the top 20 nodes for each particular network as selected “hub nodes”.
We then analyze the dynamics of these selected hub nodes and of their neighbors. Specifically we define as the subgraph spanned by a given hub node and by its first neighbors. We analyze then the cross-correlogram of the average fluorescence of a given group with the average fluorescence of the whole culture:(18)The -notation indicates that we correlate discretely differentiated average fluorescence time series, rather than the average time series themselves. Indeed, cross-correlograms for these differentiated time series are well modeled by a Gaussian functional form, due to the slow change of the averaged fluorescence compared to the sampling rate.
Therefore, we fit a Gaussian to the cross-correlogram :(19)determining thus a cross-correlation amplitude , a cross-correlation peak lag and the standard deviation .
The cross-correlation peak lag indicates therefore whether nodes in a given local hub neighborhood fire on average earlier or later than other neurons in the network.
Relative strength of synchrony within a local hub neighborhood can be analogously evaluated by computing XCs, as defined in Eq. (13), for all the links within and comparing it with peak XCs over the entire network.
Primary cultures of cortical neurons were prepared following standard procedures [14], [113]. Cortices were dissected from Sprague-Dawley embryonic rat brains at 19 days of development, and neurons dissociated by mechanical trituration. Neurons were plated onto 13 mm glass cover slips (Marienfeld, Germany) previously coated overnight with 0.01% Poly-l-lysine (Sigma) to facilitate cell adhesion. Neuronal cultures were incubated at , 95% humidity, and for 5 days in plating medium, consisting of 90% Eagle's MEM —supplemented with 0.6% glucose, 1% 100× glutamax (Gibco), and gentamicin (Sigma) — with 5% heat-inactivated horse serum (Invitrogen), 5% heat-inactivated fetal calf serum (Invitrogen), and B27 (Invitrogen). The medium was next switched to changing medium, consisting of of 90% supplemented MEM, 9.5% heat-inactivated horse serum, and 0.5% FUDR (5-fluoro-deoxy-uridine) for 3 days to limit glia growth, and thereafter to final medium, consisting of 90% supplemented MEM and 10% heat-inactivated horse serum. The final medium was refreshed every 3 days by replacing the entire culture well volume. Typical neuronal densities (measured at the end of the experiments) ranged between 500 and . Cultures prepared in these conditions develop connections within 24 hours and show spontaneous activity by day in vitro (DIV) 3–4 [14], [18], [20]. GABA switch, the change of GABAergic response from excitatory to inhibitory, occurs at DIV 6–7 [14], [20].
Neuronal activity was studied at day in vitro (DIV) 9–12. Prior to imaging, cultures were incubated for 60 min in pH-stable recording medium in the presence of 0.4% of the cell-permeant calcium sensitive dye Fluo-4-AM (Invitrogen).
Recording solution includes (in mM): , , , , 10 glucose, and 10 HEPES; pH titrated to 7.4 and osmolarity to with sucrose.
The culture was washed off Fluo-4 after incubation and finally placed in a chamber filled with fresh recording medium. The chamber was mounted on a Zeiss inverted microscope equipped with a 5× objective and a 0.4× optical zoom.
Neuronal activity was monitored through high-speed fluorescence imaging using a Hamamatsu Orca Flash 2.8 CMOS camera attached to the microscope. Images were acquired at a speed of 100 frames/s (i.e. 10 ms between two consecutive frames), which were later converted to a 20 ms resolution using a sliding window average to match the typical temporal resolution of such recordings [13], [43], [45], [90], [114]. The recorded images had a size of pixels with 256 grey-scale levels, and a final spatial resolution of . This settings provided a final field of view of that contained on the order of 2000 neurons.
Activity was finally recorded as a long image sequence of 60 minutes in duration. We verified that the fluorescence signal remained stable during the recording, as shown in Supplementary Figure S8A, and we did not observe neither photo-bleaching of the calcium probe nor photo-damage of the neurons.
Before the beginning of the experiment, inhibitory synapses were fully blocked with bicuculline, a antagonist, so that activity was solely driven by excitatory neurons.
Since cultures were studied after GABA switch, the blockade of inhibition resulted in an increase of the fluorescence amplitude, which facilitated the detection of neuronal firing, as illustrated in Figure S8B.
The image sequence was analyzed at the end of the experiment to identify all active neurons, which were marked as regions of interest (ROIs) on the images. The average grey-level on each ROI along the complete sequence finally provided, for each neuron, the fluorescence intensity as a function of time. Each sequence typically contained on the order of a hundred bursts.
Examples of recorded fluorescence signal for individual neurons are shown in Figure 1B.
The fluorescence data obtained from recordings of neuronal cultures was analyzed following exactly the same procedures used for simulated data (e.g. processed in a pipeline including discrete differentiation, TE or other metrics evaluation, ranking, and final thresholding to maintain the top 10% of connections).
Due to the lack of knowledge of the ground-truth topology, optimal conditioning level cannot be known. However, based on the similarity between experimental and simulated distributions of calcium fluorescence, we select a conditioning level such to exclude the high fluorescence transients associated to fully-developed bursting transients while keeping as many data points as possible. Concretely, this is achieved by taking a conditioning level equal to approximately two standard deviations above the mean of a Gaussian fit to the left peak of the fluorescence histogram. Such a level coincides with the point where, when gradually increasing the conditioning level, the reconstructed clustering index reaches a plateau, i.e. matches indicatively the upper limit of range II in Figure 3.
To check for robustness of our reconstruction, we generated alternative reconstructions based on different conditioning levels. For the selected conditioning value, and for both the experimental datasets analyzed (Figures 8 and S7), we verified that the inferred topological features, including notably the average clustering coefficient and connection distance, were stable in a range centered on the selected conditioning value and wide as much as approximately two standard deviations of the fluorescence distribution.
To identify statistically significant non-random features of the real cultured networks in exam, we compared the reconstructed topology to two randomizations.
A first one consisted in a complete randomization that preserved only the total number of connections in the network, but scrambled completely the source and target nodes. The resulting random ensemble of graphs was an Erdös-Rényi ensemble (see, e.g. [62]) in which each possible link exists with a uniform probability of connection , where is the total number of connections in the reference reconstructed network.
A second partial randomization preserved the in-degree distributions only, and was implemented by shuffling the entries of each row of the reconstructed adjacency matrix, internally row-by-row. In this way, the out-degrees of each node were preserved. In both randomization processes, we disallowed diagonal entries.
For both randomizations we calculated the in-degree, the distance of connections and the full clustering index for each node, leading to distributions of network topology features that could be compared between the reconstructed network and the randomized ensembles, to identify significant deviations from random expectancy.
|
10.1371/journal.pcbi.1004400 | Controlled Measurement and Comparative Analysis of Cellular Components in E. coli Reveals Broad Regulatory Changes in Response to Glucose Starvation | How do bacteria regulate their cellular physiology in response to starvation? Here, we present a detailed characterization of Escherichia coli growth and starvation over a time-course lasting two weeks. We have measured multiple cellular components, including RNA and proteins at deep genomic coverage, as well as lipid modifications and flux through central metabolism. Our study focuses on the physiological response of E. coli in stationary phase as a result of being starved for glucose, not on the genetic adaptation of E. coli to utilize alternative nutrients. In our analysis, we have taken advantage of the temporal correlations within and among RNA and protein abundances to identify systematic trends in gene regulation. Specifically, we have developed a general computational strategy for classifying expression-profile time courses into distinct categories in an unbiased manner. We have also developed, from dynamic models of gene expression, a framework to characterize protein degradation patterns based on the observed temporal relationships between mRNA and protein abundances. By comparing and contrasting our transcriptomic and proteomic data, we have identified several broad physiological trends in the E. coli starvation response. Strikingly, mRNAs are widely down-regulated in response to glucose starvation, presumably as a strategy for reducing new protein synthesis. By contrast, protein abundances display more varied responses. The abundances of many proteins involved in energy-intensive processes mirror the corresponding mRNA profiles while proteins involved in nutrient metabolism remain abundant even though their corresponding mRNAs are down-regulated.
| Bacteria frequently experience starvation conditions in their natural environments. Yet how they modify their physiology in response to these conditions remains poorly understood. Here, we performed a detailed, two-week starvation experiment in E. coli. We exhaustively monitored changes in cellular components, such as RNA and protein abundances, over time. We subsequently compared and contrasted these measurements using novel computational approaches we developed specifically for analyzing gene-expression time-course data. Using these approaches, we could identify systematic trends in the E. coli starvation response. In particular, we found that cells systematically limit mRNA and protein production, degrade proteins involved in energy-intensive processes, and maintain or increase the amount of proteins involved in energy production. Thus, the bacteria assume a cellular state in which their ongoing energy use is limited while they are poised to take advantage of any nutrients that may become available.
| Many global changes in cellular physiology occur during the growth of a typical laboratory culture of a microorganism, such as Escherichia coli, as it transitions from exponential growth to starvation where it eventually ceases dividing as nutrients become exhausted [1]. However, how these changes affect specific cellular components and processes is not fully known. Existing surveys, even if conducted at the genome scale, tend to have limited completeness, in at least two ways. First, most studies collect only one type of genome-scale data. For example, they either measure changes in gene expression, through RNA or protein levels, or they measure changes in metabolites. Second, technological limitations often prevent the detection of some subset of molecules in a category of interest. For example, small bacterial RNAs with key roles in regulation may be lost from a sample when using typical methods to purify “total” RNA from cells [2]. Furthermore, DNA microarray-based methods for profiling gene expression can only detect specific RNA sequences depending on the design of their probes, whereas RNA-seq transcriptomic methods theoretically recover all RNA species in a sample [3]. Similarly, in proteomics, 2-D gel electrophoresis approaches typically detect many fewer proteins than newer mass spectrometry based shotgun methods [4,5].
Moreover, while the short-term changes in cellular physiology that occur in a laboratory culture of E. coli have been the subject of intensive study, considerably less is known about the changes in cellular composition that occur during the long-term survival of E. coli and other non-spore-forming microbes under starvation, despite the likely prevalence of this condition in nature [6]. Most studies of this metabolic state have concentrated on the long-term survival of cells in rich medium [7]. Under these conditions, E. coli experience an ecological catastrophe in which 90–99% of the cells die within a few days due to pH and nutrient changes in the medium, and mutants emerge that continue to divide on the resources released from dead cells [8–10] Thus, these are studies of genetic adaptation to changed conditions rather than purely of changes in cellular physiology in stressed and starving, but genetically wild-type, cells.
Finally, most genome-wide analyses of gene regulation focus on comparing differential changes across only two or three distinct environmental conditions or between two different time points. These studies reveal a snapshot of global physiological regulation but they do not provide insight into the underlying dynamics of regulation. By studying the dynamics of gene regulation over time, we can develop an understanding of how a cell’s physiology is regulated in the face of a natural environment that may undergo frequent changes.
Here we performed a time course experiment of E. coli B REL606 growth and starvation up to two weeks. We used a chemically defined glucose-limited medium in which cells entered a starvation state but did not lose viability for at least one week. We collected genome-wide RNA and protein levels at multiple time points, as well as lipid-modification and central metabolic-flux data, all under identical, controlled experimental conditions. The resultant data set serves as a rich resource for computational models that span and integrate cellular sub-systems and for cataloguing and correlating the responses of specific genes and/or molecules across cellular subsystems during growth and long-term starvation. We analyzed these data using a novel, general approach for unbiased classification of expression time courses. We found that the mRNA pool was drastically reduced during starvation, possibly to limit new protein synthesis overall, and that some proteins declined rapidly in abundance, in proportion to their mRNAs, while others were buffered to rapid changes in their transcripts. Overall, we observed a pattern where starving E. coli cells employ transcriptional and translational/post-translational regulation to limit energy requirements while remaining capable of nutrient uptake and metabolism.
We grew multiple cultures of E. coli REL606, from the same stock, under identical growth conditions of long-term glucose starvation, in the same medium. The samples were subsequently distributed to different laboratories that measured RNA, protein, lipids, and central metabolic flux ratios. Freezer stocks of the REL606 strain were revived for 24 h, diluted and preconditioned for another 24 h, and diluted again to initiate the experimental time course (Fig 1A). Each biological replicate was performed on separate days. In a pilot experiment a growth curve was measured to determine informative time points for analysis (Fig 1B). Time points spanning three hours to two weeks were collected and used to measure RNA via RNA-seq, proteins via LC/MS, lipids via MALDI-TOF MS and ESI MS, and central metabolic fluxes via 13C labeled glucose and GC-MS (Fig 1C). In our conditions, the optical density at 600 nm (OD600) changed little once cells entered stationary phase (Fig 1B). Additionally, cell viability remained constant after entry to stationary phase at 24 h for up to one week. From one to two weeks, the number of viable cells per culture count decreased by 38% (Fig 1B).
We first assessed reproducibility of protein and RNA measurements. For both, we found that measurements from separate biological replicates correlated highly with each other. We saw Spearman correlations of 0.92, 0.92, and 0.95 between biological repeats of raw proteomics counts and correlations of 0.93, 0.93, and 0.94 for raw RNA-seq counts between the 3 h biological replicates (S1 Fig). Furthermore, we also compared the overlap in protein IDs between the first three time points (3, 4, and 5 hrs), when the cells were exponentially dividing and the protein concentrations were more-or-less at steady state, and we found a high overlap among these time points. Each single time point yielded just over 2600 protein IDs, any pair yielded just over 2300 common protein IDs, and all three time points yield over 2100 overlapping protein IDs (S2 Fig). Thus, our measurements were highly reproducible.
We next compared how many different RNA and protein species we detected compared to previous 'multi-omic' studies (Table 1). Yoon et al. used 2D gels and microarrays to measure 60 significantly changing proteins and 4,144 mRNAs in E. coli REL606, the same strain used in this study [11]. By comparison, at 3–4 h, we observed over 2,600 proteins, with ~1,200 that changed significantly at some point in the time course, along with 4,116 mRNAs, 85 tRNAs, and 89 other noncoding RNAs (ncRNAs), a category that is largely made up of small RNAs. Even though the total number of proteins Yoon et al. observed at early exponential phase was not reported [11], it was likely an order of magnitude less than our observations, if it followed the same pattern as the proteins found to have significant changes in expression. Taniguchi et al. measured protein and mRNA content of single cells using YFP fusions and FISH, resulting in the measurement of 1,018 proteins and 137 transcripts in an E. coli K12 strain [12]. Lewis et al. also measured ~1,000 proteins and RNA expression of 4,428 genes. Although these data sets were published separately, they were performed in the same lab and under similar conditions and thus were also comparable to a degree [13,14]. In summary, our proteomics measurements were far more complete than comparable studies, providing more than 1,000 additional protein observations than the most comprehensive other study, as many mRNAs as other studies, and additional data on tRNAs and ncRNAs.
Our experiments also provided coverage comparable to or better than other experiments that focus on proteomics or RNA measurements alone. Using stable isotope labeling of amino acids (SILAC), Soares et al. observed 2,053 proteins in at least 1 of 2 biological repeats, at a false discovery rate (FDR) of <1% [4]. We measured 2,658 proteins in at least 1 of 3 biological repeats with around 2,200 protein IDs per sample using the same FDR cutoff. A more recent study, using the filter-aided sample preparation (FASP) method, also observed around 2,200 proteins per sample, comparable to our recovery [5]. Additionally, using RNA-seq, we recovered as many mRNAs as microarray approaches do, with the added benefit of measuring 89 ncRNAs and 85 tRNAs from the same sample. As a point of reference, previous RNA-seq experiments on the E. coli K-12 strain identified 133 putative ncRNAs and 4,161 mRNAs [15]. Thus our recovery of both proteins and RNA represents the state of the art of the field, far outperforming recent comparative studies. As an added benefit of our study, we also simultaneously characterized lipid A and phospholipid composition in cell membranes and measured flux ratios in central metabolism, covering a wider range of cellular components than previous comparison studies.
We next investigated changes in relative mRNA and protein abundance over time. Due to translational and post-translational regulation we expected differences in the response of mRNA transcripts and proteins after entry to stationary phase. mRNA counts at each time point were normalized via DESeq [16], relative to the total pool of mRNA, tRNA, and ncRNA. Protein counts at each time point were normalized relative to the total protein count.
To visualize changing mRNA and protein levels we compared and contrasted the general trends in the response of mRNA and proteins by way of K-means clustering. To simplify the analysis we focused on only those mRNAs and proteins that were changing significantly (as measured by false discovery rate and fold-change cutoff, respectively) throughout the time course, yielding a total of ~1900 significantly changing transcripts/proteins. To perform K-means clustering, an arbitrary choice for the number of clusters must be made such that the profiles are well separated into groups with unique and distinct behaviors. (We also developed an alternative classification approach that does not depend on such an arbitrary choice, see below.) We varied the number of clusters for both mRNA and protein profiles, and we found the best clustering performance, assessed by visual inspection, to be around 15 clusters for the mRNA profiles and 25 clusters for the protein profiles. Thus, the mRNAs appeared to respond in a more uniform manner than the proteins did. This finding is illustrated by the heat map of the cluster centroids of mRNA and protein (Fig 2A and 2B, respectively). The vast majority of the differentially regulated mRNAs were down-regulated, while the protein response was much less uniform. Additionally, the mRNA profiles showed a clear separation between early and late time points with a transition period around 6–8 h. After this transitional period of entry to a starved state, the transcription profiles remained relatively constant, with only minor changes in expression. At two weeks some of the transcripts began changing again, perhaps signaling a further shift in cell state.
As the cells ran out of glucose, overall demand for new protein synthesis was significantly decreased, demand for certain stress response proteins increased, and resources became limiting. New protein synthesis could be globally limited in at least three ways: by reducing the amounts of rRNAs, charged tRNAs, or mRNAs. To understand how these different RNA pools changed relative to each other, we calculated the relative amount of mRNA, tRNA, ncRNA, and rRNA present in both ribosome depleted and non-ribosome depleted samples (S3 Fig). In the non-ribosomal depleted case the fraction of rRNA changed very little throughout the course of the experiment while the tRNA fraction increased and the mRNA fraction decreased. In the ribosome depleted samples (in which we removed residual rRNA counts due to incomplete depletion before analysis), the tRNA fraction also increased as the mRNA fraction decreased, confirming that this effect was not due to sensitivity or sampling-bias issues resulting from rRNA dominating the RNA pool in the non-ribosome depleted sample.
We would like to emphasize that the above clustering of the RNA and protein abundances were performed independently of each other. Therefore, we could not directly compare individual clusters between Fig 2A and 2B. The next section addresses the correlation between absolute and relative changes in abundance of individual proteins and their transcripts.
While it has been observed that absolute levels of proteins do not necessarily correlate strongly with their corresponding transcripts, we expected at least a moderate correlation between absolute mRNA and protein levels at a given time point. We also expected a correlation within individual time courses between the relative levels of a protein and its transcript. To relate the relative levels of a protein to its transcript we had to account for the underlying dynamics of the time courses. We considered two limiting cases: At one extreme we assumed each protein had a degradation rate slower than the time scale of the experiment. At the other extreme we assumed each protein was degraded on a time scale that was fast compared to the time scale of the experiment. In the first limiting case proteins integrate their transcript levels over time. In the second limiting case (relative) protein levels track with their (relative) transcript level. Obviously, we expect that some proteins do not match either of these extreme cases but fall into an intermediate regime between the two.
Plotted in Fig 2C and 2D are histograms of the Spearman correlation coefficients (ρ) calculated for proteins vs. the integrals of their transcripts (integral regulation) and proteins vs. their transcripts (proportional regulation), respectively. Approximately 15% of the proteins correlated highly (ρ>0.70) with the integrals of their transcripts whereas approximately 20% correlated highly with their transcript levels. There was little overlap between the two sets, as can be seen by the strong anti-correlation in the 2D histogram in Fig 2E of protein versus the integral and proportional levels of mRNA. Genes that were proportionally regulated were enriched for, among other things, locomotion and cell division. Genes that were integrally regulated were enriched for glycerol, alditol, and polyol metabolism. For a full list of proteins that were either proportionally or integrally related to their transcripts see S1 and S2 Tables, respectively. Approximately 65% of proteins did not fit one of these limiting models of how transcript and protein abundance were correlated; they may experience intermediate protein degradation rates or their expression and activity may be controlled by more complex post-translational modifications.
To put proteins and RNA within a given sample on comparable absolute scales, we normalized protein counts using the APEX method [17] for absolute quantification, and we normalized mRNA counts to the length of each transcript. Both protein and mRNA levels were then averaged across all three biological replicates. Additionally, all proteins and mRNAs were scaled by the average of all proteins and mRNA.
The strongest absolute correlation, across the time course, between mRNA and protein occurred at three hours (Fig 2F, Spearman ρ = 0.71, P = 10−224). Absolute correlation between proteins and their corresponding transcripts were relatively strong for time points ≤8 h, with a correlation coefficient of ~0.71. After 8 h, when cells had entered a starved state, the correlation was much weaker, with correlations around 0.3–0.4 (S4 Fig). The correlation at three hours was somewhat higher than is usually observed for correlations between RNA and protein for other measured prokaryotes and eukaryotes, which typically have Spearman correlations around 0.5 between proteins and their transcripts [18–23].
Genes within an operon are co-transcribed as a single RNA and thus are likely to be under the same transcriptional control. Differences in translational efficiency between genes often lead to larger differences in protein expression in the same operon, as regulation via changes in subcellular localization, post-translational modifications, or control of degradation rates may differently impact the activities of each of these proteins [24–27]. We expected to see a high correlation between counts of RNAs for each gene within an operon, as the genes within an operon are under the same transcriptional control; however, we expected there to be less correlation between proteins within an operon, as the proteins are not guaranteed to be subject to the same translational/post-translational regulation.
As a measure of correlation of gene expression within an operon we took the average of the pairwise Spearman correlation coefficient for all possible pairs of transcripts and proteins within an operon. Approximately eighty percent of transcripts had a mean pairwise correlation coefficient greater than 0.8 within an operon (Fig 3A). On the other hand, less than fourteen percent of proteins had a mean pairwise correlation coefficient greater than 0.8 within an operon (Fig 3B). Genes closer together within an operon were more likely to have correlated protein profiles (see Fig 3C), which we took as evidence that distance between genes was a strong indicator of translational regulation. Also shown are a few examples of highly correlated transcripts and proteins for individual operons (Fig 3D, 3E and 3F, respectively).
Typical analysis of RNA expression data often involves performing a hierarchical clustering of profiles followed by a term enrichment of subsets of genes found in the emerging patterns. In this approach the patterning that comes from hierarchical clustering can be arbitrary, depending on the level of the hierarchy one chooses to focus on. Here, instead, we sought to sort the time courses into general behaviors in an unbiased manner. To accomplish this goal we fit each individual mRNA and protein to a piecewise continuous curve (S5A Fig). This curve was defined by four free time parameters and three free amplitude parameters. To fit the curve we used a population-based differential evolution (DE) algorithm with the fitness function used in minimization scaled to the experimental error (see Methods). Thus, our algorithm provided confidence intervals for our fit based upon the variability in biological replicates.
To demonstrate the effectiveness of our fitting strategy we randomly selected five mRNA profiles and their respective fits (S5B–S5E Fig). Green circles show the average of three biological replicates with their standard deviations (green bars) and the blue line and bar show the average and standard deviation of the population of fits, respectively. Both the data and fit were normalized to the average of the time course. We also plotted histograms of the time scale parameters we found by fitting the piecewise continuous curve to our data (S6 Fig). The most informative time scales were t1, the time to first inflection, and t2+t3+t4, the time it takes for the profile to stop changing. The majority of proteins and their transcripts began changing before the 10 h mark (or just after the cells enter a starved state). Once the profiles began to change it took >10 h before they stopped changing again. However, in this case the apparent long time scale of proteins and transcripts changing could be due to the low time resolution of our experiment after the cells had entered a starved state.
As can be seen in S5B–S5E Fig, there was generally good agreement between the data and model for mRNAs. Thus, the fits gave us reasonable estimates of the distribution of time scales involved in the response. S5F Fig shows the distribution of t1, the time to first inflection. Most of the mRNAs responded between 3–8 h, with a strong peak at around 6 h (when cells began entry to a starved state). To better understand the regulation of cellular processes (and mRNAs) in our dataset, we sorted the mRNA profiles into five general categories, defined on the basis of our fitted parameters: up-regulated, down-regulated, transiently up-regulated, transiently down-regulated, or ambiguous. The confidence intervals for our fits allowed sorting individual mRNAs into these five categories with high confidence. The mRNAs in the categories down-regulated and up-regulated showed significant enrichment for GO terms. The average of the mRNAs in each of these terms is shown in Fig 4A and 4B. Terms enriched in the set of down-regulated transcripts represented translation, carboxylic acid biosynthetic process, and nitrogen compound biosynthetic process. These processes were likely down-regulated for energy conservation purposes in the face of limiting resources. Terms enriched in the set of up-regulated transcripts represented carbohydrate catabolic processes.
To characterize the protein response we followed the same general strategy of fitting, classification, and GO enrichment as we had done for the RNA profiles. The distribution of the time to first inflection for the proteins was a little broader than for the mRNAs. However, the first-inflection times still mostly fell into the range of 3–8 h, and very few proteins had not responded by the time the cells entered a starved state. There were many proteins that were present for the duration of the time course, compared to the mRNAs where very few remained present for the entire duration of the experiment. Fig 4C shows the average abundance of the proteins in a given GO term that were enriched in the set of proteins that were being up-regulated. As in the case of down-regulated RNAs these proteins were likely down-regulated to conserve energy, and they included proteins involved in translation and locomotion. Up-regulated proteins were, like the up-regulated transcripts, involved in carbohydrate catabolism but also included terms involved in stress response and metabolism of glycerol. The average protein abundances for GO terms being down-regulated had a much wider distribution of decay times compared to the RNAs being down-regulated, likely due to differing protein degradation rates (and/or thermodynamic stability) (Fig 4D).
As a complementary approach we also averaged all proteins in a given KEGG pathway regardless of their behavior. Many pathways showed little to no differential regulation, on average, in their protein levels. Pathways that changed cohesively are plotted in Fig 4E and 4F, depending on whether they were down- or up-regulated, respectively. As in the previous term-enrichment analysis, we saw motility to be down-regulated, as well as other energy consuming processes involved in metabolism and biosynthesis. Interestingly, biosynthesis of siderophores was up-regulated, likely due to do increased demands for or reduced supply of iron.
We used flux ratio analysis to measure the relative metabolic fluxes passing through different branches of central metabolism [28,29]. To measure flux ratios we used the FiatFlux software that fits a metabolic model to the amino acid labeling pattern [30]. Importantly, this analysis represents the integral of metabolism until the time at which the measurement was taken. As there was little ab-initio protein synthesis after the cells stopped growing (after ~8 h), we did not include the flux ratios after this point, except for the two-week time point. Our major observation was that there was little change in flux ratios throughout growth, and for most of the experiment this initial labeling remained (S7A–S7I Fig). Interestingly, we observed changes at two weeks in the flux ratio in P5P from G6P lower branch (S7G Fig). Given that there is not expected to be any net synthesis of amino acids after growth ceased, we cannot use the steady-state approach to interpret these data. They do suggest, however, that either internal amino acid recycling or some de novo amino acid synthesis from recycling nutrients released by dead cells occurred after one week.
Using negative-ion MALDI-TOF and ESI mass spectrometry (MS), we analyzed lipid A and phospholipid profiles, respectively, of cells at each time point. Beginning before one week, we observed an appearance of an MS peak associated with the acylation of lipid A with a C16 chain (Fig 5A and 5C). In the phospholipid analysis, a notable increase began around 6 h in the cyclopropanation of one unsaturated double bond within molecules of the major phospholipids, phosphatidylethanolamine (PE) and phosphatidylglycerol (PG). This change was identified by the gradual relative increase of peaks at ~702.5 m/z and ~733.5 m/z, respectively. (Representative data for PE is shown in Fig 5D.) Both the modifications to lipid A and phospholipids continued to increase up to the two-week time point. In fact, the 702.5 m/z peak corresponding to cyclopropanation of phospholipid was barely detectable before six hours but became the predominant peak by the end of the time course.
The enzymes relevant to the above lipid A and phospholipid modifications are lipid A palmitoyl transferase (PagP) and cycloproponated fatty acid synthase (CFA), respectively [31,32]. PagP is known to be constitutively transcribed at low levels and remain latent in the outer membrane until enzyme activation [33]. It is also up-regulated by the transcriptional regulator, PhoP, under various stressful conditions encountered by a cell [34]. However, during our time course, transcript levels of PagP and PhoP did not change significantly. Furthermore, neither PagP nor PhoP was observed at the protein level. In the case of PagP, this could be due to the difficulty in detecting outer membrane beta-barrel proteins by our mass-spec proteomics method. With respect to phospholipid modification, CFA synthase protein levels increased between 3–6 h before decreasing again. This observation agreed with prior data showing that CFA synthase was important during the transition to stationary phase [32]. CFA synthase RNA levels increased again around one week, which was consistent with the activity observed in phospholipid modification, although it is not clear why we did not observe a corresponding increase in protein levels at this point (Fig 5B).
We have collected a comprehensive E. coli time course and have developed computational techniques to analyze such data. Our computational techniques are general and can be applied to other time-course data collected in future studies. In particular, fitting piecewise continuous curves to expression profiles allowed us to reliably sort individual profiles into four basic groups, up-regulated, down-regulated, transiently up-regulated, or transiently down-regulated. Additionally, we have developed an unbiased approach to compare mRNA and protein profiles and to identify those proteins whose abundances followed their mRNA levels and those that were buffered against rapid mRNA changes.
Our results provide a coherent picture of E. coli stationary phase, as summarized in Fig 6. E. coli could survive for over a week when starved for glucose in a well-buffered minimal medium, with little change in cell viability (Fig 6A). The fraction of mRNA relative to all RNA was down regulated after cells entered stationary phase (Fig 6B). As cells ceased to divide, the demand for new protein synthesis declined. Reducing the overall pool of mRNA could contribute to limiting new protein synthesis. Upon entry to stationary phase, lipid A and phospholipids were modified by PagP and CFA synthetase, respectively (Fig 6C). Modification of lipids continued gradually until eventually the lipid species that were rare during growth dominated at two weeks.
All genes started to change in expression by 10 h, and mRNA expression clustered temporally into two regimes, before and after 10 h (cells entered a starved state at around 8 h) with some late changes in expression beginning around two weeks (Fig 6D). We found that 20% of observed proteins were regulated in proportion to their transcripts (Fig 6E), allowing for rapid down-regulation of the processes they were involved in. On the other hand, 15% of the observed proteins were integrally related to their transcripts (Fig 6E) and likely served to buffer against environmental changes. In addition to measuring and characterizing RNA and protein changes upon entry to stationary phase, we also demonstrated how a piecewise curve-fitting strategy allowed us to classify expression profiles into different categories. The enriched terms in the resulting classification were reasonably aligned with what was known about, or at least consistent with, cells coping with starvation (Fig 6D). Importantly, this classification was accomplished in an unbiased manner, without any ad hoc assumptions about the number of clusters that should exist in the data.
We found that, as cells entered a starved state, the total pool of mRNA was depleted compared to all other RNAs and many individual transcripts were down-regulated, possibly as part of a broader strategy to reduce the production of new protein. Reducing overall protein production could also be achieved by limiting the available ribosomes or by limiting the pool of available tRNA. The stringent response, activated in starving cells through the ppGpp alarmone, down-regulates new rRNA synthesis [35]. However, in our data, the relative fraction of rRNA within a cell changed little over time, and the tRNA fraction actually increased with time. Thus, new protein synthesis in starving cells may be limited more by the reduced mRNA pool than by reduced translational efficiency due to decreases in rRNA or tRNA abundance. Even if the total rRNA decreased over the time course, the total mRNA would have decreased more by a proportional amount. Said another way, the down-regulation of new rRNA synthesis by the stringent response may be most important for shutting down the production of ribosomes needed by new cells in an actively dividing culture, rather than for reducing the level of ribosomes in already existing cells.
It has been suggested that the degradation rate of many proteins in E. coli is much slower than the doubling time during growth [36,37]. As a consequence, when cells cease to divide, such as in the case of glucose starvation, not all proteins can respond immediately to possible changes in transcript levels. In effect, the amounts of some proteins may be buffered against relatively fast changes in nutrient availability. At the same time certain proteins may need to be rapidly regulated to ensure survival upon starvation. We found that a subset of the proteome, ~20% of proteins, fell into the rapidly regulated category that may be degraded quickly—they maintained an abundance that was proportional to their transcripts. Another subset, ~15% of proteins, tended to be much more stable—they were proportional to the integrated abundance of their transcripts over the time scale of our experiment. For example, the abundance of several flagellar proteins was proportional to their transcript levels, whereas proteins involved in metabolism and energy production integrated their transcript levels over time. Turning off proteins involved in cell division and the flagellar machinery, both energy-intensive processes, needs to happen relatively quickly. By contrast, the proteins that were relatively stable were enriched for energy production terms. Thus, these proteins presumably persist so that if nutrients were to become available again the cell will be capable of using them to re-initiate growth. For proteins to track dynamically with their transcripts they must have a short half-life. For this reason, we can compare those terms enriched for proteins that dynamically correlate with their transcripts to the COG terms reported by Maier et al. [23] that have shorter than average half lives in M. pneumoniae. We found that those COG terms with shorter than average half-lives were generally consistent with terms that were enriched in highly dynamically correlated proteins and mRNAs. In particular, Maier et al. found that terms involved with energy production (COG term C), metabolism (COG terms H, I, G), protein turnover (COG term O), and signaling (COG term T) had protein turnover rates significantly faster than the overall average.
Among the terms that were significantly regulated in stationary phase, we saw that motility was down-regulated, likely because it places a high energy burden on cells [1]. Additionally, it has been shown that flagella in E. coli are down-regulated by the stringent response [38]. Other observed differential regulation is related to energy conservation (shutting down expensive or unneeded pathways), catabolism (breaking down non-essential components for food), stopping translation of new protein (as there is no longer demand for protein from new cells), or a general stress response (increasing nutrient influx or bolstering membrane integrity).
We also found many uncharacterized genes (both among the protein and the RNA profiles) that were significantly up- or down-regulated upon entry to stationary phase. A subset of these proteins have computationally predicted functions [39] that were consistent with our findings for annotated genes. For instance, several uncharacterized proteins that were up-regulated are predicted to be involved in stress response and cell-wall biogenesis. Other predictions seem to be inconsistent with our observations for annotated genes or indicate that these genes regulate rather than take part in these processes. For example, some uncharacterized proteins that were up-regulated are predicted to be involved in translation, even though translation was heavily enriched in down-regulated genes. These uncharacterized genes might down-regulate the activity of ribosomes, for example. Lists of proteins and transcripts that were significantly regulated in our time course are provided in the Supplemental materials (S3 and S4 Tables).
Even though mRNA abundances within an operon were highly correlated (as expected), in many cases their protein profiles were only weakly correlated. This finding could be due to different translation efficiencies between proteins [40] as well as differing degradation rates. In support of the former, we saw a tendency for proteins separated by a larger distance within a transcript to be less correlated than those located closer to one another. However, it was likely that different protein degradation rates also played a role in the low correlation between proteins within an operon. Indeed, many proteins coded by proximal regions of a transcript showed poor correlation in their profiles (Fig 3C). Other explanations for this tendency of proteins nearby on the genome to be more highly correlated could be due to distance from the transcript start site or transcript length. Yet, our data did not show evidence for either of these scenarios. Distance from the transcript start site was not correlated with protein expression (ρ = –0.02, P = 0.65) and transcript length was only very weakly correlated with protein expression (ρ = 0.12, P = 0.003). However, we cannot necessarily rule out other explanations for the observed intra-operon protein correlation vs. distance between genes.
In addition to the expected disparities between RNA and protein levels, we also observed surprising changes in enzyme activity that did not correspond to the respective RNA-seq and proteomics analysis. For example, we saw striking levels of lipid modification late during the time course. These modifications were easily explained by their association with adaptation to stressful environments such as depleted nutrients and cations as well as increased acid resistance during starvation [32,34]. However, the stark differences in RNA, protein, and activity trends of the enzymes responsible for the lipid modifications, PagP and CFA synthase, highlight the fact that activation does not necessarily follow abundance measurements. In support of this idea, it has been shown that cylopropination by CFA synthase depends upon the concentration of bicarbonate, which could lead to a decoupling between protein levels and activation [41].
Metabolic fluxes were quite constant throughout the growth phase of the experiment, and these labeling patterns remained in place once growth ceased. At the two-week time point, however, the labeling patterns in histidine changed substantially, which during steady-state growth on glucose would have been interpreted as a change in the flux ratio corresponding to P5P from the G6P lower branch declining. This change so late in the experiment was unexpected, since we did not anticipate substantial turnover in cellular composition that late in stationary phase. The observation suggests that either internal amino acid recycling or some de novo amino-acid synthesis, possibly related to the moderate decline in the number of viable cells, occurs past the one week time point.
A goal of systems biology has been to understand how phenotype originates from genotype. The phenotype of a cell is determined by complex regulation of processes including cell signaling, gene regulation, metabolism, and lipid biochemistry. Understanding the connection between phenotype and genotype is crucial to understanding disease and for synthetic engineering of biology. Even though computational models of individual component subsystems, such as flux models of metabolism [42–44], have enjoyed a long history of success, they remain limited in their application. Much effort is currently being spent on understanding how to best integrate data from multiple subsystems. For example, there are many proposed approaches to combining gene expression with metabolic flux networks [20,45–52] while other studies have focused on integrative, whole-cell models [53,54]. Given the growing interest in integrative modeling approaches, there is a pressing need for studies that collect high quality genome-wide data across multiple cellular subsystems from the same biological samples. Our data set is a rich resource for comparing and contrasting the response of multiple cellular subsystems. Additionally, in the future we plan to use the techniques developed in this paper to measure the response of E. coli to several other environmental conditions, which will allow for more detailed models of regulation.
Despite the completeness and quality of our data set, however, there were a few key limitations concerning our approach. Our analysis via RNA-seq and shotgun MS allowed for high confidence when comparing the relative levels of a particular transcript or protein over time. However, due to potential differences in detection efficiency between individual RNAs or peptides, care should be taken when comparing absolute abundances. In our analysis we used the APEX method to account for differences in protein detection efficiency. We normalized RNA by the length of a transcript as an estimate of RNA detection efficiency, for a particular experiment. This approach resulted in a correlation coefficient of ~0.71 between proteins and their transcripts, a finding on the high end for such correlation measurements. Previous reports on the correlation between mRNA and protein levels in E. coli and M. pneumoniae have yielded correlation coefficients of ~0.5 [18–23]. Thus, even straightforward means of correcting our experimental bias led to reasonable comparisons of levels between individual RNAs or proteins. Additional detection biases we did not account for, such as GC content bias in RNA-seq [55], were likely responsible for some of the remaining unexplained variation [56].
The stepwise linear function we used for modeling works for a majority of our expression profiles. However, in some cases it over-fits the data and in other cases the function is unable to capture the underlying behavior. An example of a profile that may be under-constrained is a gene that is up- or down-regulated without further changes to expression. In this case the free time parameters t2 through t4, along with amplitude parameter A2, may be under-constrained. Even in this case, however, the parameters t1, t2+t3+t4, A1, and A3 are still well constrained, providing enough information to reliably sort the profiles based upon behavior. Thus, since we used our model for classification and not for prediction purposes, any potential parameter over-fitting did not substantially affect our final results. More complicated temporal profiles, such as multiple peaks separated in time, could not be captured by our function. The presence of these more complicated behaviors was rare enough as to not warrant special consideration.
In summary, our study provides both the most complete measurement, to our knowledge, of multiple cellular components in a changing environment, and novel computational approaches to analyze such data. Thus, this work represents an important step toward understanding how regulation of a cell’s physiology is coordinated, on a global, systems level, by interactions between multiple cellular subsystems.
E. coli B REL606 was inoculated from a freezer stock into 10 ml of Davis Minimal medium supplemented with 2 μg/l thiamine (DM) [57] and limiting glucose at 500 mg/l (DM500) in a 50 ml Erlenmeyer flask. This culture was incubated at 37°C with 120 r.p.m. orbital shaking over a diameter of 1". After overnight growth, 500 μl of the culture was diluted into 50 ml of prewarmed DM500 in a 250 ml flask and grown for an additional 24 h under the same conditions. On the day of the experiment, 500 μl of this preconditioned culture was added to ten 250 ml flasks, each containing 50 ml DM500, to initiate the experiment. At each time point, aliquots of these cultures were removed as necessary to harvest a constant number of cells given the changes in cell density over the growth curve. Each sample was pelleted by centrifugation, washed with sterile saline (0.85% (w/v) NaCl), and then spun down again. After removing the supernatant, the resulting cell pellet was flash frozen using liquid nitrogen and stored at –80°C. Each of the three biological replicates was performed on a separate day. Samples for each type of cell composition measurement were taken from the same batch of flasks, except for those used for flux analysis, for which an additional batch was grown in 20% [U-13C] glucose.
For graphs of OD600 and colony-forming units (CFU), cultures were grown separately from the main batches used for harvesting cells but under identical conditions. The OD600 (absorbance at 600 nm) of a sample removed from the culture at each time point was measured relative to a sterile DM500 glucose blank. These samples were also diluted in sterile saline and plated on DM agar supplemented with 0.2 g/l glucose. After incubation at 37°C for 24 h colonies on these plates were counted to determine CFUs.
Total RNA was isolated from cell pellets using the RNAsnap method [2]. After extraction, RNA was ethanol precipitated and resuspended in 100 μl H2O. Each sample was then DNase treated and purified using the on-column method for the Zymo Clean & Concentrator-25 (Zymo Research). RNA concentrations were determined throughout the purification using a Qubit 2.0 fluorometer (Life Technologies). DNase-treated total RNA (≤5 μg) was then processed with the Gram-negative bacteria RiboZero rRNA removal kit (Epicentre). After rRNA depletion, each sample was ethanol precipitated and resuspended in H2O again. A fraction of the RNA was then fragmented to ~250 bp using NEBNext Magnesium RNA Fragmentation Module (New England Biolabs). After fragmentation, RNA was ethanol precipitated, resuspended in 20 μl ultra-pure water, and phosphorylated using T4 PNK (New England Biolabs). After another ethanol precipitation cleanup step, sequencing library preparation was performed using the NEBNext Small RNA Library Pre Set for Illumina, Multiplex Compatible (New England Biolabs). Samples were ethanol precipitated again after library preparation and separated on a 4% agarose gel. All DNA fragments greater than 100 bp were excised from the gel and isolated using the Zymoclean Gel DNA Recovery kit (Zymo Research). Libraries were sequenced using an Illumina HiSeq 2500 at the Genomic Sequencing and Analysis Facility (GSAF) at the University of Texas at Austin to generate 2×101-base paired-end reads.
For RNA-seq analysis, we implemented a custom analysis pipeline using the REL606 Escherichia coli B genome (GenBank:NC_012967.1) as the reference sequence [58]. We updated annotations of sRNAs in this genome sequence using the Rfam 11.0 database [59]. Prior to mapping, we trimmed adapter sequences from Illumina reads using Flexbar 2.31 [60]. Mapping was carried out in single-end mode using Bowtie2 2.1.0 with the –k 1 option to achieve one unique mapping location per read [61]. The raw number of reads mapping to each gene was counted using HTSeq 0.6.0 [62]. The full computational pipeline is available at https://github.com/wilkelab/AG3C_starvation_tc_RNAseq.
E. coli cell pellets were resuspended in 50 mM Tris-HCl pH 8.0, 10 mM DTT. 2,2,2-trifluoroethanol (Sigma) was added to 50% (v/v) final concentration and samples were incubated at 56°C for 45 min. Following incubation, iodoacetamide was added to a concentration of 25 mM and samples were incubated at room temperature in the dark for 30 min. Samples were diluted 10-fold with 2 mM CaCl2, 50 mM Tris-HCl, pH 8.0. Samples were digested with trypsin (Pierce) at 37°C for 5 h. Digestion was quenched by adding formic acid to 1% (v/v). Tryptic peptides were filtered through Amicon Ultra 30 kD spin filtration columns and bound, washed, and eluted from HyperSep C18 SpinTips (Thermo Scientific). Eluted peptides were dried by speed-vac and resuspended in Buffer C (5% acetonitrile, 0.1% formic acid) for analysis by LC-MS/MS.
For LC-MS/MS analysis, peptides were subjected to separation by C18 reverse phase chromatography on a Dionex Ultimate 3000 RSLCnano UHPLC system (Thermo Scientific). Peptides were loaded onto an Acclaim C18 PepMap RSLC column (Dionex; Thermo Scientific) and eluted using a 5–40% acetonitrile gradient over 250 min at 300 nl/min flow rate. Eluted peptides were directly injected into an Orbitrap Elite mass spectrometer (Thermo Scientific) by nano-electrospray and subject to data-dependent tandem mass spectrometry, with full precursor ion scans (MS1) collected at 60,0000 resolution. Monoisotopic precursor selection and charge-state screening were enabled, with ions of charge >+1 selected for collision-induced dissociation (CID). Up to 20 fragmentation scans (MS2) were collected per MS1. Dynamic exclusion was active with 45 s exclusion for ions selected twice within a 30 s window.
Spectra were searched against an E. coli strain REL606 protein sequence database and common contaminant proteins (MaxQuant using SEQUEST (Proteome Discoverer 1.4; Thermo Scientific)). Fully-tryptic peptides were considered, with up to two missed cleavages. Tolerances of 10 ppm (MS1) and 0.5 Da (MS2), carbamidomethylation of cysteine as static modification, and oxidized methionine as dynamic modification were used. High-confidence peptide-spectral matches (PSMs) were filtered at <1% false discovery rate determined by Percolator (Proteome Discoverer 1.4; Thermo Scientific).
Flux ratios were obtained from the samples grown with 13C labeled glucose, using methods previously described [28,29]. Cell pellets were resuspended in 200 ml of 6 N HCl, hydrolyzed at 105°C overnight, and dried at 95°C for up to 24 h. To the hydrolyzed cell material we added 40 ml of dimethylformamide (DMF) and gently mixed until a “light straw” color was obtained. The DMF resuspension was transferred to a GC-MS vial with plastic insert and 40 ml of N-tert-butyldimethylsilyl-N-methyltrifluoroacetamide with 1% tert-butyldimethyl-chlorosilane (v/v); vials were capped and baked at 85°C for 2 h, and samples were analyzed within 2 days of derivitization.
Analysis of derivitized samples was performed on a Shimadzu QP2010 Plus GC-MS (Columbia, MD) with autosampler. The GC-MS protocol included: 1 mL of sample injected with 1:10 split mode at 230°C; an oven gradient of 160°C for 1 min, ramp to 310°C at 20°C/min, and hold at 310°C for 0.5 min; and flow rate was 1 mL/min in helium. A total of five runs were performed for each sample: a blank injection of DMF to waste, a blank injection of DMF to the column, and three technical replicates of each vial.
Flux inference was performed using the FiatFlux software as described [29,30]. Each time point was analyzed separately, and the reported fluxes represent the integral of growth up until that point. For this reason, we do not display fluxes beyond 8 h, when growth ceased and there should not be any more net amino acid synthesis. We did, however, monitor the labeling patterns in all amino acids for the later time points. Although most patterns were unchanged, we did note that histidine labeling changed substantially at the final time point of two weeks.
Lipid A and phospholipids were isolated from bacterial pellets containing 3–9×109 cells. Pellets were resuspended in 5ml 1:2:08 chloroform:methanol:water for 20 min and spun at 10,000×g for 10 minutes. Pellets containing lipid A were further purified by the Bligh/Dyer method as previously described [63]. Phospholipids in the supernatant were further purified by extractions as previously described [64]. Mass analysis of purified lipid A fractions was performed using a MALDI-TOF/TOF (ABI 4700 Proteomics Analyzer) mass spectrometer in the negative ion linear mode as previously described [63]. Phospholipid analysis was performed by liquid chromatography/ESI-mass spectrometry as previously described [64]. One of the three replicates used for lipid analysis was an additional independent biological replicate, prepared identically to all other replicates but not used for RNA-seq or proteomics analysis.
We analyzed raw counts from the proteomics and RNA-seq experiments as follows. Initially, proteins with low counts (<10) over the entire duration of the time course were filtered out. Each time point was then normalized to the read depth (e.g. the sum of all counts for that particular time point). Only proteins with a fold change of ≥1.5 were considered for further analysis. Protein profiles were then normalized to the maximum value for a given protein time course. To estimate the absolute protein abundance we made use of the APEX normalization method [65].
To analyze relative RNA levels, raw RNA read counts per gene (ignoring rRNAs) were normalized within each sample using DESeq [16]. To identify RNAs that had changed significantly, we carried out a differential expression analysis between the 3 h time point and each subsequent time point, using DESeq, and we kept RNAs with a significant difference (p < 0.05) at least one time point for further analysis. To compare absolute RNA abundances within a single time point, raw RNA counts were normalized by gene length. Finally, normalized RNA and protein profiles, both relative and absolute, were averaged across all three biological replicates.
Clustering of protein profiles was performed using the python library scipy [66]. We used the k-means clustering algorithm with the number of protein clusters set to 25 and RNA clusters to 15. To compare relative protein profiles with the integral of their relative transcript levels we integrated each of the transcript profiles, from the initial time to each additional time point, using the trapezoidal method implemented by the python library numpy [67].
We used a piecewise continuous curve to fit both RNA and protein profiles. This curve was defined by seven free parameters, four free time parameters, and three free amplitude parameters. To fit the profiles we used a custom implementation of a differential evolution (DE) algorithm [68]. Briefly, the DE algorithm initially generates an ensemble of random parameter guesses within a predefined range; subsequently, vectors of individual parameter sets (sometimes called agents) are mixed together at a predefined crossover rate, only those crossover events that yield a smaller error (defined by a predefined cost function) are kept, and the process is iterated until a convergence criterion is met. In our fits we used an ensemble of 15 agents with a crossover frequency of 0.75 and a mixing strength of 0.6. The crossover frequency determines the probability that an agent will be changed at any given iteration and the mixing strength determines how large a change an agent undergoes if it was chosen to be altered. The crossover frequency and mixing strength were picked based upon an empirical study of the dependence of convergence efficiency on these parameters [69] for some standard optimization problems. The cost function is given by
Fi=∑j(di(tj)−si(tj))2σi2(tj)
where di(tj), σi(tj), and si(tj) are the average of all experimental repeats of protein (or mRNA) i at time tj, the standard deviation of the experiments of the protein (or mRNA) i at time tj, and the average of the ensemble simulations i at time tj, respectively. Scaling by the standard deviation places a relatively lower weight on data points with relatively larger errors for a given protein or mRNA.
Some of the profiles may be slightly over-fit by our curve (e.g. profiles that are up-regulated or down-regulated once during the time course without further modulation of expression). Thus care needs to be exercised in the interpretation of some of the parameters. However, we found t1 to reliably represent the time to first inflection, the sum of t2, t3, and t4 was a decent proxy to how long it took an RNA/protein to reach a steady state after entering a starved state, and we could reliably sort the behavior into four categories based upon the amplitude parameters. The four categories we used were that of up-regulated, down-regulated, transiently up-regulated, and transiently down-regulated. Genes that were up (or down) regulated were those genes that increased (or decreased) at some point during the time course and did not decrease (or increase) at some later time. Genes that were transiently up (or down) regulated were those genes that increased (or decreased) at some point during the time course but decreased (or increased) at some later time. The sorting into categories was aided by our estimate of the distribution of parameters that allowed for a good fit within the population of fits. A fit was considered good if it was on average (across the time course) one standard deviation or less away from the experimental average.
We used the DAVID database (david.abcc.ncifcrf.gov) to perform Gene Ontology term enrichment on each subset of sorted genes: up-regulated, down-regulated, transiently up-regulated, and transiently down-regulated. Specifically we made use of DAVID's API, instead of the web interface, to generate the GO-enrichment through a python script. GO terms were clustered based upon genes in a given term to reduce redundancy in the returned results.
As a complementary approach we also enriched for KEGG pathway terms in the entire set of significantly changing proteins (without presorting) using the DAVID database API. The protein levels within each returned KEGG pathway were then averaged to see if there was any consistent response across the entire pathway. Those KEGG terms that gave inconsistent responses across proteins in that pathway returned a relatively flat average and were filtered out.
All of the scripts used to perform the above analysis can be downloaded at https://github.com/marcottelab/AG3C_starvation_tc.
Raw data are available from the Dryad Digital Repository: http://dx.doi.org/10.5061/dryad.hj6mr. Raw Illumina read data and processed files of read counts per gene and normalized expression levels per gene have been deposited in the NCBI GEO database (accession GSE67402) [70]. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository (accession PXD002140) [71].
|
10.1371/journal.pcbi.1005002 | How to Estimate Epidemic Risk from Incomplete Contact Diaries Data? | Social interactions shape the patterns of spreading processes in a population. Techniques such as diaries or proximity sensors allow to collect data about encounters and to build networks of contacts between individuals. The contact networks obtained from these different techniques are however quantitatively different. Here, we first show how these discrepancies affect the prediction of the epidemic risk when these data are fed to numerical models of epidemic spread: low participation rate, under-reporting of contacts and overestimation of contact durations in contact diaries with respect to sensor data determine indeed important differences in the outcomes of the corresponding simulations with for instance an enhanced sensitivity to initial conditions. Most importantly, we investigate if and how information gathered from contact diaries can be used in such simulations in order to yield an accurate description of the epidemic risk, assuming that data from sensors represent the ground truth. The contact networks built from contact sensors and diaries present indeed several structural similarities: this suggests the possibility to construct, using only the contact diary network information, a surrogate contact network such that simulations using this surrogate network give the same estimation of the epidemic risk as simulations using the contact sensor network. We present and compare several methods to build such surrogate data, and show that it is indeed possible to obtain a good agreement between the outcomes of simulations using surrogate and sensor data, as long as the contact diary information is complemented by publicly available data describing the heterogeneity of the durations of human contacts.
| Schools, offices, hospitals play an important role in the spreading of epidemics. Information about interactions between individuals in such contexts can help understand the patterns of transmission and design ad hoc immunization strategies. Data about contacts can be collected through various techniques such as diaries or proximity sensors. Here, we first ask if the corresponding datasets give similar predictions of the epidemic risk when they are used to build a network of contacts among individuals. Not surprisingly, the answer is negative: indeed, if we consider data from sensors as the ground truth, diaries are affected by low participation rate, underreporting and overestimation of durations. Is it however possible, despite these biases, to use data from contact diaries to obtain sensible epidemic risk predictions? We show here that, thanks to the structural similarities between the two networks, it is possible to use the contact diaries to build surrogate versions of the contact network obtained from the sensor data, such that both yield the same epidemic risk estimation. We show that the construction of such surrogate networks can be performed using solely the information contained in the contact diaries, complemented by publicly available data on the heterogeneity of cumulative contact durations between individuals.
| Knowledge of the structure of human interactions is crucial for the study of infectious diseases spread and the design and evaluation of adequate containment strategies. The structure of contact networks, [1], the presence of communities [2], bridges or “linkers” between communities [3–5], “super-spreaders” [6–8], the heterogeneity of contact durations [9], are all important characteristics that determine potential transmission patterns. The study of human contacts is particularly relevant in contexts such as schools, working places, hospitals where individuals might spend several hours in close proximity [5, 10–22].
Interactions and contacts between individuals are conveniently seen within the framework of networks in which nodes represent individuals and (weighted) links correspond to the occurrence of contacts (the weight giving the duration of the contacts). Measuring directly such networks represents an important challenge [20]. Many studies have relied on contact diaries or surveys [10, 12, 23–32], while technological advances have led to a strong increase in the use of wearable sensors in the recent years [9, 13–15, 17–20, 33–35]. Quantitative comparisons between datasets obtained from sensors and self-reported diaries, in terms of the numbers and durations of contacts between individuals and of the contact network statistics, are however scarce, mainly because very few studies have combined these two data collection means [21, 36]. These investigations have shown that diaries suffer from small participation rates, under-reporting of contacts, and over-estimation of the contact durations. Under-reporting is particularly strong for short contacts, while long ones are better reported, and some studies have put forward methods to estimate its magnitude and to correct for it [29, 37]. Interestingly, and despite the much lower number of nodes and links in contact networks inferred from contact diaries, the overall structure of these networks is very similar to the one obtained from wearable sensors. Moreover, the links with largest weights (as measured by sensors), which might play a major role in propagation processes, are reported with high probability in the contact diaries.
In this paper, we go beyond the comparison of the contact networks obtained by these methodologies and explore the impact of their differences on the evaluation of the epidemic risk when such datasets are used in numerical simulations of infectious disease propagation. Our goal is to understand to what extent and how the information gathered from contact diaries can be used in such simulations in order to yield an accurate description of the epidemic risk, despite the biases mentioned above. We first compare the outcomes of spreading simulations performed using data coming from wearable sensors and from contact diaries that describe the contacts between students in the same context (a high school) and on the same day. Although the two networks are supposed to describe the same reality, we observe important differences in the simulations, due to the low participation rate in the diaries and to a stronger community structure in the contact diaries network than in the contact sensors network. We then design and evaluate a set of methods to use the information contained in the contact diaries to build surrogate versions of the contacts that yield, when used in the simulations, a better estimation of the real epidemic risk as quantified by the distribution of epidemic sizes (considering as ground truth the dataset from the sensors). We show that good results are obtained when the contact diary information is complemented by known stylized facts characterizing human interactions, in particular the heterogeneity of contact durations.
We use two datasets collected in a French high-school in 2013 and made publicly available in [21]. The data describe face-to-face contacts between students of 9 classes as collected by (i) the SocioPatterns infrastructure [38] based on wearable sensors, during one week and (ii) self-reported contact diaries filled on a specific day of the same week (Dec. 5th, 2013). In the diaries, contact was explicitly defined as close (less than 2 m) face-to-face proximity, in order to match as much as possible this definition to the contacts detected by sensors. Using these data, we build two distinct contact networks for the day in which the diaries were collected: the Contact Sensors Network (CSN) and the Contact Diaries Network (CDN). In each network, nodes represent students and a link is drawn between a pair of nodes (i, j) if at least one contact between students i and j is present in the corresponding dataset during the considered day. We present and compare in the Supporting Information the main networks’ characteristics. Note that, in the diaries, some participants reported contacts with non-participants. One could a priori use this information and build a contact network including both participants and non-participants. However, since by definition the contacts of non-participants are unknown, this would introduce a potentially strong and most importantly completely uncontrolled bias in the measures of the network’s structural properties such as, e.g., its clustering or the node degrees.
A weight can moreover be assigned to each link (i, j): for the CSN, the weight wij is given by the cumulative duration of the contacts registered by the sensors on that day between i and j; for the CDN we can use the duration reported by the students in the diaries, building the network CDND, or use for each link a duration taken at random from the list of durations registered by the sensors, obtaining the network CDNS (see Methods for details). The rationale behind building CDNS comes from the results of [21, 36] that show that durations reported by students tend to be strongly overestimated. Since, on average, contacts reported in the diaries as long tend also to be long according to the sensor data, we will also consider a different assignment of links to the CDN, in which we still take durations at random from the list of durations registered by the sensors, but assign the longer durations to the links of CDN with longer reported durations: we denote the resulting network by CDNS’.
The contact sensor network counts 295 nodes (participation rate 77.8%) and 2162 links, while the contact diaries network has 120 nodes (participation rate 31.6%) and 348 links. Incomplete participation, even in the case of the sensor data, leads to biases in the simulations using the CSN with respect to what would be obtained if the whole population had participated, due to the fact that contacts with and among non-participants are not detected. This point has been discussed in [40], together with methods to build surrogate data and obtain estimate of the epidemic risk in the case of such population sampling. In order not to confuse the issues of population sampling and comparison between diaries and sensors, we consider here as ground truth the CSN, collected by wearable sensors for which the definition of contact does not depend on a possible interpretation of the diary question by the students nor on the fact that they might not recall contact events.
In the following, we perform simulations of the spread of infectious diseases in the considered population, using as substrate for propagation events the contact networks described above. It is important to note here that we consider propagation processes on static networks. Indeed, the CDN does not contain information on the timing of the contacts, so that it is natural to compare the outcome of simulations performed on such a static network with simulations performed on a static version of the CSN. Moreover, when modeling the propagation of infectious diseases with realistic timescales of several days, it has been shown in [39] that a daily weighted contact network contains enough information to obtain a good estimate of the process outcome. When dealing with faster processes, the temporal evolution of the network would become relevant; in that case, it would be possible to use the techniques put forward in [40] to build realistic surrogate timelines of contacts on weighted networks, using the robustness of the distributions of the durations of single contact events and of the intervals between successive contacts measured in different contexts. Note also that, even if the networks do not take into account the timing of the contact events, they still include information on the aggregate contact durations through the weights, which are known to play a crucial role in the outcome of spreading processes [39, 41–43].
For simplicity, we consider the paradigmatic Susceptible-Infected-Recovered (SIR) model of epidemic propagation. In this model, each Susceptible node i can be infected by an Infected neighbour j with probability β*wij*dt for each small time step dt. Infected people recover with rate μ and enter in the Recovered category. Recovered individuals cannot be infected again. The process starts with a single Infected individual chosen at random (the seed) and ends when there are no more Infected nodes. The epidemic risk in the population, which depends on the interplay of the ratio β/μ and the network’s structure and weights, is measured by the distribution of the final size of epidemics (i.e., of the fraction of individuals in the Recovered category at the end of the process), obtained by repeating the simulations with randomly chosen seeds. Note that, since we consider static networks, only the ratio β/μ is relevant, and multiplying both by a certain factor only changes the timescale on which the epidemic unfolds. The shape of the distribution of epidemic sizes depends on the features of the underlying network structure in terms of possible patterns of contagion. The comparison of these distributions gives hints about similarities and discrepancies of various datasets for the evaluation of the epidemic risk.
We first compare in Fig 1 the outcome of simulations of the SIR model performed on the CSN and on the two versions of the CDN described above (CDND with weights reported by students and CDNS with weights registered by sensors assigned randomly to the links), for one specific value of β/μ = 30. The three distributions of epidemic sizes are very different from each other. The outcome of simulations performed using CSN is quite standard, with a fraction of small outbreaks that reach only a small fraction of the population and another peak corresponding to large outbreaks. As shown in the Supporting Information, the outcome does not depend on the class of the initial seed. The shape of the distribution obtained when using the CDND is more peculiar, with a series of peaks, including one at very large epidemic sizes. Such structure is typical of spreading processes on networks with a strong community structure [4], which corresponds to the results of [21]: (i) due to the low participation rate and the under-reporting, the community structure of the CDN is stronger than the one of the CSN, with few links between classes; depending on the seed, the simulated disease can thus remain confined in one class or in a group of few classes, leading to the peaks at intermediate values of the epidemic size; we moreover show in the SI that the outcome depends on the class of the initial seed for the CDN but not for the CSN; (ii) on the other hand, as contact durations are overestimated, the propagation probability on each link is also overestimated and, if the disease manages to spread between classes, almost all individuals are affected, leading to the peak at large epidemic sizes. The CDNS case shows a different result: no more than half of the whole population is affected by the spread. As the weights have in this case the same statistics as the CSN, this is simply due to the low participation rate [40] and the much smaller average degree in the CDN with respect to the CSN. We also note that, since the weights are assigned randomly to the links between students, the structure of the contact matrix giving the average durations of contacts between students of different classes can strongly differ between the CDNS and both the CSN and the CDND, leading to different patterns of propagation between classes (see Supporting Information). We finally note that the simulations on the CDNS’, which keeps the distribution of the weights from CSN and in which larger weights are assigned to links with longer reported durations, yield even smaller outbreaks. This is probably due to the fact that the large weights reported in the diaries tend to be within classes, so that the links bridging classes and favoring the spread tend to have smaller weights in the CDNS’ than in the CDNS. We also show in the SI the temporal evolution of the density of infectious individuals for the various cases considered here.
Data on the contact patterns of individuals collected by different methods lead to different contact network structures, and some studies have started to investigate this issue through detailed quantitative comparisons [21, 36]. In the present paper, we have gone further by comparing the outcome of simulations of spreading processes performed on contact networks gathered either through wearable sensors or through contact diaries. Not surprisingly, we have shown that the results differ strongly, due to the low participation rate to the diaries, the under-reporting of contacts and the overestimation of contact durations in diaries. In particular, the direct use of the links and durations reported in the diaries yields a peculiar distribution of epidemic sizes suggesting a very strong community structure that might lead to the design of inadequate containment strategies. On the other hand, using the links reported in the diaries but more realistic weights yields a strong underestimation of the epidemic risk.
In a second and more important step, we have asked if, despite this low participation rate and these biases, the information contained in the contact network built from the contact diaries can be used to build a surrogate contact network whose properties are close enough to the real contact network (considered here to be the one obtained from the wearable sensors) to yield a correct estimation of the epidemic risk when used in simulations of spreading processes. The rationale comes from the structural similarities found in the contact matrices giving the densities of links between individuals of different classes obtained using both sensors and diaries [21]. These similarities suggest to build a surrogate contact network starting from the contact diary network, adding nodes and links in order to maintain this matrix fixed, and assigning weights to the links. We note that two recent works [37, 40] have considered related but different issues. In [37], only diary data is available, and the authors present a synthetic network model based on data and adjusting for under-reporting. This adjustment for reporting errors leads in this case only to a small difference in epidemic predictions. In [40] on the other hand, only sensor data is considered, and the authors assume to have an incomplete information on the contact network registered by sensors due to an uniform population sampling (hence, all contacts between participating individuals are assumed to be known). Here on the other hand, the available dataset is given by diaries, in which population sampling is not uniform (actually, the students who filled in diaries tend to have more contacts than the others) and in which under-reporting implies that many links between participating individuals are also missing. Moreover, we face two additional issues (i) the low sampling rate yields a contact matrix of link densities which contains zeros, in an unrealistic way, and (ii) various possibilities can be considered when assigning weights to the links of the surrogate networks as weights reported in the diaries are strongly overestimated. Despite these issues, the surrogate network we build yields, when used in simulations, a good agreement with simulations performed on the whole contact sensor network in terms of epidemic risk prediction, under the condition of using the list of weights (cumulative contact durations) gathered by the wearable sensors. In order to get rid of this condition, we argue that this list comes from a distribution that has been shown in previous works to be very robust across contexts [9]. We therefore consider weights taken at random from a pool of publicly available datasets, and show that using these weights gives also satisfactory results. Overall, we thus have presented a procedure that uses only the information contained in the contact diaries and in public data, which allows to obtain a good prediction of the epidemic risk, as measured by the distribution of epidemic sizes, when used in simulations of a spreading process.
In the Supporting Information, we moreover consider the issue of using, instead of contact diaries, data coming from friendship surveys, in order to build the surrogate contact network used in the simulations. We show that the epidemic risk prediction obtained through this procedure is not accurate. This could be expected as daily encounters in the school are not necessarily related to the existence of a relationship between students: contacts occur between non-friends due to daily activities, while friends do not meet necessarily every day. This outcome highlights the importance of taking into account the different nature of social ties [21, 46], which can each be relevant for specific processes.
Some limitations of our work are noteworthy. First, our results rely on an assumption made in replacing the zeros observed in the contact matrix of link densities by random values. In the context under scrutiny, zero values can indeed easily be considered as unrealistic. In other contexts, it might however happen that different groups in the population really do not mix. In such a case, one might expect that this kind of information could be gathered from other sources (schedules, location of classrooms or offices, etc) [47] and thus integrated into the procedure. Second, we have considered as ground truth the contact sensor network. On the one hand, this network in fact suffers from an incomplete participation, so that the outcome of spreading processes is underestimated with respect to hypothetical data containing information on the whole population. However, such underestimation can be compensated through the procedure presented in [40]. On the other hand, it is important to note that it is not yet completely clear whether the contacts measured by wearable sensors are the best proxy for potentially infectious contacts. We work therefore under this hypothesis, which is indeed quite widely used but should be kept in mind. Third, we have considered here static networks. As discussed in [39], the outcome of simulations is then close enough to the one of simulations taking into account the full contact dynamics if we consider slow enough processes. For fast processes, the burstiness of contacts becomes very relevant; in this case, it would be crucial to supplement our procedure by the construction of surrogate timelines of contacts and intervals between contacts at high temporal resolution, as done in [40]. Finally, we cannot at this point investigate the efficiency of our procedure in other contexts, for lack of datasets reporting contacts measured by both wearable sensors and contact diaries in the same context and at the same date. Hopefully such datasets will become more available in the future, yielding new testing grounds for our method. Many populations of interest can indeed be divided into groups or categories that do not mix homogeneously, often with more contacts within groups than between groups, and for which the contact matrix formalism and the procedures we present to construct surrogate networks are therefore relevant [40]. We conclude by mentioning that future work could also investigate other dynamical processes on networks, such as information spreading or opinion formation processes.
The datasets we use have been presented and made publicly available in [21]. They correspond to contacts between students of 9 classes in a high school in France, collected through wearable sensors on the one hand and contact diaries on the other hand. The sensors registered contacts with a temporal resolution of 20s for 327 participating students (out of 379 in the 9 classes, i.e., a 86.3% participation rate) during the week of Dec. 2–6, 2013. Contact diaries contain data reported by students about encounters and their cumulative durations for Dec. 5, 2013. In these diaries, students were asked to report the cumulative durations of their contacts choosing among four intervals: at most 5 minutes, between 5 and 15 minutes, between 15 minutes and 1 hour, more than one hour. The students belong to 9 classes with different specializations: “MP” classes focus more on mathematics and physics, “PC” classes on physics and chemistry, “PSI” classes on engineering studies and “BIO” classes on biology. We collected data among students of nine classes corresponding to the second year of such studies: 3 classes of type “MP” (MP, MP*1, MP*2), two of type “PC” (PC and PC*), one of type “PSI” (PSI*) and 3 of type “BIO” (2BIO1, 2BIO2, 2BIO3).
Using these datasets, we build two networks of contacts among students for the same day (Dec 5, 2013): the Contact Sensors Network (CSN) and the Contact Diaries Network (CDN). In each network, nodes represent students, and a links is drawn between two students if:
The resulting networks have 295 nodes for the CSN (other students were absent or did not wear the sensors on that day) and 120 nodes for the CDN. In particular, no student from classes PC* and PSI* filled a diary, and only one from MP*1. We thus discarded these classes in most of the analysis and in particular in the contact matrices, remaining with 6 classes.
Each link carries a weight. In the CSN it represents the cumulative duration of contacts registered by sensors during the day. For the CDN we consider several possibilities. In the CDND we use weights reported in the diaries: we associate to each time-interval its maximum possible value (5, 15, 60 minutes respectively for the first three intervals and 4 hours for the last one. This choice takes into account data reported and registered and the school schedule) and, if two students reported different durations for their encounter, we use the average of the reported values. In the CDNS on the other hand, we consider weights randomly drawn from the distribution of contact durations registered by sensors. Results are averaged over 1000 such weight assignments. For the CDNS’ finally, we start from CDN and rank the E links in decreasing order of their reported weights (assigned as in CDND, and with random order for equal weights). We extract E weights from the distribution of contact durations registered by sensors, rank them as well in decreasing order, and assign the weights to the links of CDN in such a way to match the two orderings (i.e., assigning the largest weights to the links with largest reported weights).
The CSN, the CDND and the CDNS are matched to retain only nodes appearing in both CSN and CDN (see Table 1 for details about classes size before and after matching). We refer to them as the matched networks: CSNm, CDN D m and CDN S m.
The CMDN is built by using a Contact Matrix Distribution (CMD). Following [44], we consider a CMD where each entry, (X, Y), is the empirical distribution of durations reported by diaries for contacts between all students in class X and class Y, including zero durations (corresponding to an absence of link between two students). We fit each such distribution by a negative binomial functional form. Then, for each pair of nodes, we draw at random a weight using the corresponding negative binomial fit. Note that in this way we do not maintain fixed the link structure of the CDN. We however keep on average the same density of links between different classes.
The basic steps for building a binary surrogate contact network CDNs for the six considered classes, starting from the matched contact diaries network, are:
Results are averaged over 500 realisations of this procedure.
As explained in the main text, we moreover deal in two different ways with the zero values of the link densities between several class-pairs in the CDN. We either keep these densities or replace them with values drawn at random from a uniform distribution of values between the minimum and maximum values (diagonal excluded) of the contact matrix of Fig 3(b). In this way the contact matrix structure is preserved, with more interactions within than between classes.
To assign weights to the links of the surrogate networks, we consider several possibilities. We first assume homogeneous contact durations and assign to each link a weight equal to the average of cumulative durations registered by sensors. This yields two versions of the surrogate contact networks:
We refer to the contact sensors network under the homogeneous duration hypothesis by CSNH.
If instead we assume heterogeneous contact durations, we obtain two possible surrogate contact diaries networks: we assign weights at random to the links of CDNs, drawn at random with replacement from the list of durations either reported by students or registered by sensors. We thus obtain four versions of the surrogate contact networks:
Finally, the surrogate contact network obtained by assigning weights randomly drawn from the negative binomial fit of the distribution of publicly available contact durations registered by sensors is indicated by the acronym CDN nz,NB s.
|
10.1371/journal.pgen.1004720 | Hsp40s Specify Functions of Hsp104 and Hsp90 Protein Chaperone Machines | Hsp100 family chaperones of microorganisms and plants cooperate with the Hsp70/Hsp40/NEF system to resolubilize and reactivate stress-denatured proteins. In yeast this machinery also promotes propagation of prions by fragmenting prion polymers. We previously showed the bacterial Hsp100 machinery cooperates with the yeast Hsp40 Ydj1 to support yeast thermotolerance and with the yeast Hsp40 Sis1 to propagate [PSI+] prions. Here we find these Hsp40s similarly directed specific activities of the yeast Hsp104-based machinery. By assessing the ability of Ydj1-Sis1 hybrid proteins to complement Ydj1 and Sis1 functions we show their C-terminal substrate-binding domains determined distinctions in these and other cellular functions of Ydj1 and Sis1. We find propagation of [URE3] prions was acutely sensitive to alterations in Sis1 activity, while that of [PIN+] prions was less sensitive than [URE3], but more sensitive than [PSI+]. These findings support the ideas that overexpressing Ydj1 cures [URE3] by competing with Sis1 for interaction with the Hsp104-based disaggregation machine, and that different prions rely differently on activity of this machinery, which can explain the various ways they respond to alterations in chaperone function.
| The cellular chaperone machinery helps proteins adopt and maintain native conformations and protects cells from stress. The yeast Hsp40s Ydj1 and Sis1 are co-chaperones that regulate Hsp70s, which are key components of many chaperone complexes. Both of these Hsp40s are crucial for growth and Ydj1 directs disaggregation activity of the Hsp100-based machinery to provide stress protection while Sis1 directs this activity to promote prion replication. Ydj1 also cures yeast of certain prions when overexpressed. We show that C-terminal domains that possess substrate-binding function of Ydj1 and Sis1 can mediate these and other functional distinctions and that the degree that prions depend on Sis1 activities could underlie differences in how they respond to alterations of chaperones. These findings support a view that Hsp40s regulate and specify functions of the chaperone machinery through substrate discrimination and cooperation with Hsp70. The disproportionate evolutionary expansion of Hsp40s (J-proteins) relative to their Hsp70 partners led to a proposal that this amplification allows increased regulation and fine-tuning of chaperone machines for increasingly complex processes. Our findings support this idea and provide insight into fundamental aspects of this cooperation.
| The protein disaggregation machinery of microorganisms and plants is driven by an Hsp100-family chaperone that cooperates with Hsp70 and its nucleotide exchange factor (NEF) and J- protein (Hsp40) co-chaperones [1]. This machinery promotes cell survival after environmental stresses that cause proteins to aggregate by extracting proteins individually from aggregates [2]–[4]. Organisms encode multiple Hsp70s, Hsp40s and NEFs and there is much to learn about how these chaperones cooperate and regulate each other's activity to effect protein remodeling and reactivation, and how different combinations of chaperones and co-chaperones determine efficiency and specificity of the machinery.
In yeast, this Hsp104-based resolubilization process also targets prions, which are cellular proteins that propagate as highly structured fibrous protein aggregates called amyloid [5]–[7]. The widely studied prions [URE3], [PSI+] and [PIN+] (also known as [RNQ1+]) are composed of the proteins Ure2, Sup35, and Rnq1, respectively [8]–[10]. Propagation of these and other amyloid-based yeast prions requires proper functioning of the disaggregation machinery [5], [11]–[15], which promotes prion replication by fragmenting fibers into more numerous pieces, or seeds, that continue to propagate the prion state [4], [16], [17].
Hsp70s act in various cellular chaperone machines by binding and releasing hydrophobic surfaces on partially folded proteins. This activity is necessary for essential cellular processes where proteins are partially folded, such as transport across membranes, and for preventing protein aggregation under conditions of stress [18], [19]. Effective interactions of Hsp70 with substrates rely on its regulation by J-proteins and NEFs (reviewed in [20]). The major yeast cytosolic Hsp40s Sis1 and Ydj1 are structurally related J-proteins that function as dimers [21]. Both have an amino-terminal J domain that mediates physical interaction with Hsp70s and an adjacent glycine-phenylalanine (GF) rich region that confers some functional distinction [22], [23]. Both also have carboxy-terminal regions that bind substrates with a specificity that overlaps Hsp70 [24]–[26]. The class I J-protein Ydj1 has a zinc-finger element within its C-terminal region and a CAAX motif at its extreme C-terminus that directs its farnesylation. This modification localizes much of Ydj1 to membranes and influences cooperation of Ydj1 with Hsp90, another abundant and highly conserved protein chaperone [27], [28]. The class II J-protein Sis1 lacks these elements, but has a glycine-methionine (GM) rich extension of its GF domain.
Altering function or abundance of Sis1 or Ydj1 inhibits propagation of prions, but in different ways. By an undefined mechanism, overexpressing Ydj1 causes cells to lose [URE3] and some variants of [PIN+], but not [PSI+] [12], [29]. Increasing expression of Sis1 does not destabilize these prions [30], [31]. Depleting Sis1 causes [URE3] and [PIN+] to be lost rapidly as cells divide, but causes [PSI+] to be lost gradually and only after a long delay [14]. Additionally, all non-essential functions of Sis1 are dispensable for [PSI+] propagation, but deleting only the GF region of Sis1 causes cells to lose [PIN+] [30], [32]. Thus, the way these Hsp40s influence prion propagation goes beyond their general roles of regulating Hsp70. Additionally, when cooperating with E. coli disaggregation machinery in yeast, Sis1 is specifically required for prion propagation and Ydj1 for protecting cells from exposure to lethal heat (thermotolerance) [15]. Both Hsp40s are critical for survival and while no other J-protein can compensate for loss of Sis1, Sis1 and other J-proteins, as well as J-domains alone, can improve growth of cells lacking Ydj1 [14], [27], [33], [34]. What defines these functional differences of Sis1 and Ydj1 is uncertain.
Here, we used Sis1-Ydj1 hybrid proteins to identify structural elements that determine the distinct functions of Sis1 and Ydj1 in various cellular processes and systematically assessed the importance of Sis1 activities for propagation of [URE3] and [PIN+]. We found that the C-terminal regions of Sis1 and Ydj1 possessed functional distinctions that directed the action of the Hsp104 machinery in prion propagation and thermotolerance, and the Hsp90 machinery in galactose-induced transcription. We also found that [URE3] was highly sensitive to alterations of Sis1 and that [PIN+] was less dependent on Sis1 than [URE3], but more dependent than [PSI+]. Our results support the idea that differences in ways prions respond to various chaperone alterations can be due to differences in their dependence on the disaggregation machinery.
Earlier we showed that the E. coli disaggregation machinery (ClpB, DnaK and GrpE, which are analogous to yeast Hsp104, Hsp70, and NEF, respectively) function in yeast by cooperating with yeast Hsp40s [15]. ClpB, DnaK and GrpE are herein abbreviated B, K and E, respectively. We modified this system to use a DnaK mutant (R167H, designated K*) that can interact only with J-proteins that have the compensatory D36N J-domain mutation (designated Sis1* and Ydj1*). BK*E cannot cooperate with wild type J-proteins, so we can monitor interactions of the BK*E machinery specifically with Sis1* or Ydj1* even in the presence of their wild type counterparts. Using this system we showed BK*E cooperates specifically with Sis1* to propagate [PSI+] prions and with Ydj1* to protect cells from exposure to lethal heat (thermotolerance). The ability of Sis1 and Ydj1 to direct activities of the disaggregation machinery could be due to differences in the ways they recruit or regulate Hsp70 components of the machinery or interact with different types of substrates. To determine the basis of these and other functional differences we used Sis1-Ydj1 hybrid proteins.
In earlier work using hybrids of Ydj1 and Sis1 each of the CTDs was divided into two parts (CTDI and CTDII) and the adjacent GM region of Sis1 was included on the same fragment containing the amino-terminal portion of the Sis1 CTD [35], [36]. Exchanging this fragment splits the contiguous functionally redundant GF-GM regions of Sis1 [22], which complicates interpretations of swapping GF domains. In order to simplify comparisons we exchanged only the three most conserved J, GF(GM), and CTD regions to form six hybrid proteins (see Figure 1A, Materials and methods). Hybrids are named according to their structural components. For example, YYS and SSY have their CTDs swapped. All of our J-protein hybrids for the BK*E experiments contain the D36N mutation, which is indicated in their names by an asterisk (e.g. Y*YS is the D36N mutant of YYS).
To test for ability of these modified Sis1-Ydj1 hybrid proteins to cooperate with the E. coli chaperones to propagate [PSI+], they were first expressed in [PSI+] cells that have ClpB in place of chromosomal Hsp104 and express K*, E and Hsp104 from plasmids. They were then assessed for ability to continue propagating [PSI+] after the cells lose the plasmid encoding Hsp104. The results are presented in Figure 1B. The upper panel shows cells on medium that allows growth of all strains and the lower panel shows medium that allows growth only of cells propagating [PSI+]. The lack of distinction among strains in the upper panel indicates that in the absence of selection [PSI+] propagates too weakly to confer an obvious phenotype, as reported earlier [15]. Nevertheless, the strong confluent growth of cells transferred onto medium selecting for [PSI+] indicates the prions in these cells are mitotically stable. The presence of [PSI+] was confirmed by its dominant phenotype in crosses and by guanidine curability (see Figure S1A). Wild type Sis1 and all hybrids containing the CTD of Sis1 propagated [PSI+]. In contrast, Ydj1 and all proteins with the CTD of Ydj1 were unable to propagate [PSI+]. Thus, the Sis1-specific function that is necessary for these full-length Hsp40s to cooperate with the bacterial ClpB disaggregation machinery to propagate [PSI+] prions resides in the Sis1 CTD.
These results were somewhat unexpected because Sis1 lacking its CTD (Sis1ΔCTD) propagates [PSI+] in cells expressing Hsp104 [32]. We therefore tested if BK*E could cooperate with the truncated Sis1*ΔCTD to propagate [PSI+]. We found it did, but [PSI+] was very unstable and was lost rapidly when selection for the prion was not maintained (Figure 1C). These results indicate that Hsp40 requirements of BK*E for [PSI+] propagation differ when full-length and truncated Hsp40s are used. Inconsistencies between truncated and full-length Hsp40s have been seen before (see below) [22], [30], [33].
Because the dimerization domain is at the C-terminus, it was also possible that Y*YS was able to propagate [PSI+] only by combining with the wild type Sis1 present in the cells to form heterodimers that functioned with BK*E. Since Sis1 lacking its dimerization domain (Sis1ΔD) propagates [PSI+] in cells with Hsp104 [32], we tested if the D36N version of Sis1ΔD was able to support [PSI+] propagation with BK*E. At the same time we tested a version of Y*YS lacking its dimerization domain. Both of these proteins were able to cooperate with BK*E to propagate [PSI+], but with noticeably reduced efficiency compared with their full-length counterparts (Figure 1C, note slower growth and pink color on medium selecting for [PSI+]). Thus, Y*YS did not need to dimerize with wild type Sis1 to support propagation of [PSI+].
Using our hybrids to assess Ydj1 function in thermotolerance we found that, except for Y*SY, proteins containing the CTD of Ydj1 worked better at restoring thermotolerance to hsp104Δ cells expressing BK*E (Figure 1D). Overall, the range of survival conferred by the different hybrids was somewhat broad, which suggests that complex interactions among different parts of the proteins can influence Hsp40 functions in this process. Notably, however, S*SY, in which only the CTD is from Ydj1, functioned most like Ydj1 in this assay, while Y*YS performed only slightly better than Sis1*. Thus, as with prion propagation, a determinant of functional specificity between Sis1 and Ydj1 that directs BK*E activity in thermotolerance resides in the CTD.
Although Sis1 is essential for viability, its J, GF and CTD regions are each dispensable for growth of sis1Δ cells [22], [30]. However, other work shows that a full-length Sis1-Ydj1 hybrid that contains the substrate-binding region of Sis1, but not Ydj1, supports growth of sis1Δ cells [35]. These earlier findings show inconsistencies in the way that Sis1 and Ydj1 sub-regions determine specific Hsp40 functions. The variable dependency of [PSI+] on the CTD of Sis1 in our full-length versus truncated [32] proteins is another example of such differences. An obvious distinction in these experiments is whether the structural regions are deleted or swapped. While deleting regions allows identifying redundant functions, swapping domains of full-length proteins also allows identifying functions that can be influenced by inter-domain interactions, or if the domains specify interactions with other factors or localization of the J-proteins in the cell.
In line with the earlier findings using full-length hybrid proteins, our results with the E. coli chaperones show the CTDs can impart specific functionality to these Hsp40s. To assess which regions are involved in determining specificity of cooperation with the yeast chaperone machinery we tested ability of our hybrids to function in place of Sis1 or Ydj1, which allows investigation of Hsp40 functions other than those needed for prion propagation and thermotolerance.
We compared the relative abundance of the hybrid proteins using western blot analysis (Figure S2). Sis1 and hybrid proteins containing the CTD of Sis1 were less abundant than the others, so we cannot rule out that differences in abundance were contributing to effects in some assays. In addressing this issue by increasing expression using the stronger GPD promoter on both single and high-copy plasmids, we found that hybrid proteins containing the CTD of Sis1 caused growth inhibition of ydj1Δ cells in a dose-dependent manner (see Figure S3). Therefore, in our complementation assays we expressed proteins regulated by the SIS1 promoter on single-copy plasmids. Despite differences in protein levels, in several experiments the less abundant proteins (i.e. those with CTD of Sis1) functioned better than the others (see results above and below). Therefore, differences in ability of the hybrids to complement functions cannot be explained solely by differences in protein levels.
To determine if individual Sis1 sub-regions within full-length proteins are enough to support growth of sis1Δ cells, both [PSI+] and [psi−] versions of strain 930 (sis1Δ carrying wild type SIS1 on a URA3 plasmid) were transformed with plasmids encoding the hybrid proteins and then grown as patches on medium that allows cells to lose the URA3 plasmid encoding wild type Sis1. These were then replica-plated onto medium containing FOA to select for cells having lost that plasmid. Regardless of prion status, all hybrids that contained the Sis1 CTD, and only these hybrids, supported growth (Figure 2A, left panels). Therefore, in the context of our full-length hybrids, the CTD of Sis1 was necessary and sufficient to provide essential Sis1 activity. These same hybrids supported propagation of [PSI+], which is consistent with our earlier findings that propagation of [PSI+] is minimally dependent on Sis1 and that any Sis1 mutant that supports growth also supports [PSI+] [32].
As expected, these same hybrids supported growth of isogenic sis1Δ strain 1385 (used to monitor [URE3]). However, only those containing both the CTD and GF regions of Sis1 (i.e. Sis1 and YSS) supported propagation of [URE3] (Figure 2B). Thus, propagation of [URE3] depended on the GF/GM and CTD regions of Sis1, but not on the Sis1 J-domain. The Ydj1 GF region did not function in place of Sis1 GF/GM (i.e. in SYS) to support prion propagation. These latter observations are reminiscent of earlier work showing that a short stretch of the GF region in Sis1, which is absent in Ydj1, is important for propagation of [PIN+] [23], and they suggest the same function is important for propagation of [URE3].
Aside from its role in protecting cells from exposure to lethal heat, Ydj1 is important for cell growth under all conditions [27]. Cells lacking Ydj1 are viable, but they grow very slowly at 25°C and do not grow at 34°C. Elevating expression of Sis1 and other J-proteins, or even J-domains alone, can improve growth of ydj1Δ cells [14], [34], which suggests that the functions of Ydj1 in its roles important for viability are more general. Nevertheless, we tested if distinct domains of Sis1 and Ydj1 conferred Ydj1-specific functions important for growth by repeating the plasmid shuffle using ydj1Δ strain MR502, which has YDJ1 on a URA3 plasmid.
We found that all hybrid proteins containing the CTD of Ydj1 restored growth noticeably, even at 34°C (Figure 3A). The substantial ability of YSY and SSY to restore growth indicates that the Ydj1 GF region is effectively dispensable for its functions in cell growth and its J-domain has a small contribution. Compared with the empty vector, Sis1 also improved growth weakly at 30°C, which is in line with earlier data showing increased expression of Sis1 compensates for loss of Ydj1 [27], but it did not support growth at 34°C (Figure 3A). Cells with YSS grew slightly better than those with the empty vector at 25°C and 30°C, but the other hybrids containing the Sis1 CTD failed to improve growth, even at 25°C. These results suggest that in the context of our full-length proteins the CTD of Ydj1 possesses Ydj1-specific functions important for growth.
As indicated above, increasing expression of proteins with the CTD of Sis1 inhibited growth of ydj1Δ cells in a dose-dependent manner (Figure S3). One explanation for this effect is that hybrids with the Sis1 CTD were able to form defective heterodimers with wild type Sis1 and a resulting impairment of Sis1 function would exacerbate the growth defect caused by lack of Ydj1. To test this possibility, we repeated the experiments using hybrids lacking the dimerization domain (ΔD). Even though the abundance of these truncated proteins was at the low level of their counterparts (Figure S2C), YSSΔD improved growth considerably at 30°C and allowed weak growth at 34°C (Figure 3B). SYSΔD and YYSΔD also supported growth at 30°C, but only slightly. These results are consistent with the growth inhibition being caused by dominant inactivation of wild type Sis1 and suggest that lack of complementation was not necessarily due to lower protein abundance.
Deleting the dimerization domain of wild type Sis1 also improved its ability to suppress the ydj1Δ defect, although not enough to support growth at 34°C. Thus, monomeric Sis1 is better at performing Ydj1-specific functions than Sis1 dimers. Together these results agree with earlier work showing a general ability of truncated J-proteins to complement Ydj1 function better than intact proteins [34]. They also suggest dimerization might specify or restrict Hsp40 activities.
Ydj1 is also a critical component of the Hsp90 molecular chaperone system necessary for activating galactose-inducible gene promoters [37]. This machinery is thought to remove nucleosomes from the promoter region to allow access to transcription factors. Deleting YDJ1 abolishes galactose induction by disrupting this process. When assessed for ability to function in galactose induction (Figure 4), all of the hybrids that had the CTD of Ydj1, and only these hybrids, restored GAL expression. Thus, the CTD of Ydj1 determined specificity of Ydj1 in a process that involves its cooperation with the Hsp90 machinery.
We next tested whether the ability of the Hsp40s to exhibit functional discrimination in vivo was reflected in discrimination in vitro. To do this we monitored protein reactivation of two different substrates. When heat-inactivated GFP-38, a GFP fusion protein containing a C-terminal 38 amino acid peptide, was used as the substrate, Sis1 in combination with Hsp104 and Ssa1 restored about 30% of the GFP-38 after an hour (Figure 5A). With Ydj1 in place of Sis1 in the reaction, there was little reactivation (<2%). The hybrid protein containing the CTD of Sis1, YYS, was able to reactivate GFP-38 with Hsp104 and Ssa1, but the rate of reactivation was ∼50% that of Sis1 (Figure 5A and B). There was no detectable reactivation of GFP-38 by SSY under the same conditions. These results show that reactivation of GFP-38 by Hsp104 and Ssa1 requires a Sis1-specific function and that the CTD of Sis1, when appended to the J-GF of Ydj1, was sufficient to provide this function.
In contrast, with heat-inactivated luciferase as substrate, Ydj1 in combination with Ssa1 promoted reactivation and Sis1 with Ssa1 was inactive (Figure 5C). SSY was as active as Ydj1 in reactivating luciferase with Ssa1. YYS in combination with Ssa1 was unable to reactivate heat-denatured luciferase. Together these results show that Sis1 and Ydj1 discriminate between protein aggregates and discrimination is a function of the Sis1 and Ydj1 CTDs.
The extent that [URE3] depends on Sis1 has not been evaluated systematically. We monitored [URE3] in sis1Δ strain 1385, which carries a URA3-based plasmid encoding Sis1 to support viability. We expressed previously described versions of Sis1 engineered to contain deletions or point mutations from a TRP1-based plasmid [32]. Deletions remove defined structural domains, the H34Q substitution in a conserved histidine-proline-aspartate (HPD) motif comprising residues 34–36 in the J-domain disrupts a critical interaction with Hsp70 [38], [39], and the K199A substitution in the CTD disrupts substrate binding [40] (see Figure 6A). To assess evolutionary conservation of Hsp40 function we also included the human Sis1 homolog DnaJB1 (also known as Hdj1). When expressed in place of Sis1, DnaJB1 supports cell viability and propagation of certain variants of [PIN+] and [PSI+] [23], [32], [41], [42]. Depleting functional Ure2 into [URE3] prion aggregates makes our strains grow slowly [15], which is evident when comparing sizes of [ure-o] and [URE3] colonies (see Figure 6B).
When Sis1 proteins lacking the GF region or containing the H34Q point mutation were expressed with wild type Sis1 they had obvious dominant inhibitory effects on [URE3], seen as appearance of red [ure-o] colonies (Figure 6B). Because the H34Q mutation disrupts physical interaction of J-proteins with Hsp70, the dominant inhibition of [URE3] propagation by the H34Q mutant might be caused by its forming defective hetero-dimers with wild type Sis1 or by competing with Sis1 for substrate.
To test these possibilities, we combined H34Q with alterations that interfere with ability of Sis1 to dimerize (ΔD) or to bind substrate (K199A). Both mutations reduced the dominant anti-[URE3] effect to a similar extent (from ∼23% to ∼6% [ure-o] colonies), but did not eliminate it (Figure 6B, rightmost images). Thus, inhibition of [URE3] by Sis1-H34Q depended partially on each of these Sis1 functions, suggesting it could be acting by making defective dimers with wild type Sis1 or by competing with Sis1 for binding to substrate, which in this system would be Ure2 amyloid. Although blocking dimerization can affect cooperative interaction of Sis1 with substrates in vitro [24], these results suggest that Sis1-H34Q can interfere with functions of wild type Sis1 in multiple ways.
To determine if the mutant Sis1 proteins could support propagation of [URE3], we counter-selected against the URA3 plasmid encoding wild type Sis1 as described in Figure 2A (see Figure 6C). Because several mutant Sis1 proteins appeared incapable of supporting [URE3] when this plasmid shuffling was done on plates without selecting for the prion, we also replica-plated the same patches of cells onto a series of similar plates lacking adenine to ensure recovery of cells capable of propagating [URE3], but weakly (Figure 6D). Cells will grow on FOA lacking adenine only if the Sis1 mutant supports both growth and [URE3] propagation.
Cells expressing each of the mutant Sis1 proteins, except those containing the lethal H34Q mutation, grew on the FOA plate that contained adenine (Figure 6C), showing the mutant proteins supported growth in place of Sis1. However, only the cells carrying wild type Sis1 had a normal white [URE3] phenotype on this plate, indicating [URE3] was lost rapidly from cells expressing any of the mutant Sis1 proteins as soon as the plasmid encoding wild type Sis1 was lost. Accordingly, when cells from this plate were streaked for isolated colonies on medium containing adenine, only the cells expressing wild type Sis1 gave rise to uniformly white [URE3] colonies (Figure 6E). Therefore, stable propagation of [URE3] depended on all of the Sis1 activities tested.
Although cells expressing Sis1ΔD, Sis1-K199A and the mutant with both of these mutations propagated [URE3] when selection for the prion was maintained (Figure 6D), they all lost [URE3] rapidly when grown on medium containing adenine (Figure 6F, right panels). On medium lacking adenine, [URE3] cells expressing most of the mutant proteins formed colonies at a rate similar to those expressing wild type Sis1p (Figure 6F, left panels), suggesting that the rapid loss of the prion under non-selective conditions was not due to the prion causing a disproportionate inhibitory effect on growth. However, it was evident that the Sis1ΔCTD [URE3] cells grew much more slowly than the others on medium selecting for [URE3] (Figure 6D, 6F, images on left). Since [ure-o] cells expressing Sis1ΔCTD grew like wild type [ure-o] cells (compare Figure 6E middle right image with upper left image) this slower growth was caused by the presence of [URE3], suggesting that the CTD region of Sis1 protects cells from toxic effects of [URE3]. Similar prion-associated toxicity was seen for [PSI+] cells expressing Sis1ΔCTD in place of Sis1 [32], [41].
We did not recover cells expressing Sis1ΔGF or Sis1ΔGMCTD on FOA plates lacking adenine. Thus, in agreement with results using the hybrid proteins, [URE3] requires the Sis1 GF region to propagate. The inability to recover [URE3] cells expressing Sis1ΔGMCTD might indicate [URE3] is even more toxic in these cells. Alternatively, [URE3] could be unable to propagate in cells expressing only JGF even under conditions selecting for the prion. DnaJB1 propagated [URE3] only under selective conditions, and even then only weakly. Overall, our results indicate that [URE3] depends much more on Sis1 than [PSI+] does.
We showed earlier that BKE (with wild type DnaK) supports [PSI+] propagation [15], so we tested if this system would also be useful for studying [URE3]. BKE did not support [URE3] in hsp104Δ cells (Figure S4). Because wild type DnaK should be able to interact with J-proteins other than Sis1, any or all of the cytosolic J-proteins might be able to compete with Sis1 for interaction with DnaK. Since [URE3] has a stringent requirement for Sis1, a resulting reduction in ability of Sis1 to interact with the BKE machinery probably explains the inability of [URE3] to be propagated. Alternatively, as stable propagation of [URE3] depends critically on which Hsp70 is present [43], [44], the failure might reflect a requirement for a specific Hsp70 activity lacking in DnaK.
Although the underlying mechanism of how overexpressed Ydj1 cures cells of [URE3] is uncertain, interaction with Hsp70 is critical because Ydj1 mutants unable to interact with Hsp70 do not cure and the J-domain alone of Ydj1 or other yeast Hsp40s is enough to cure [14], [45]. Since all parts of Ydj1 except the J-domain can be mutated or deleted without disrupting curing, we expected the CTD of Ydj1 would not have a major influence on the curing of [URE3]. Instead, we anticipated that if a hybrid cannot function in place of Sis1 with the disaggregation machinery that replicates [URE3] prions, then it will interfere with this function if it can compete effectively with Sis1 for interaction with the Hsp70 component of this machinery.
When overexpressed, SYS and YYS cured like Ydj1 (see Figure 7A). SYY and YSY cured somewhat less effectively, and YSS cured inefficiently. Thus, the SYS and YYS hybrids that did not propagate [URE3] cured [URE3] very effectively, while YSS, which supported [URE3], cured only weakly.
Since SYS and YYS possess the dimerization region of Sis1, their ability to cure [URE3] when overexpressed again might be related to an ability to form non-productive dimers with endogenous Sis1, which could contribute to the curing by partially depleting cytosolic Sis1 function. In agreement with this explanation, although disrupting dimerization of Ydj1 does not affect curing considerably [45], hybrids with the CTD of Sis1 that lacked the dimerization domain were significantly reduced in their ability to cure [URE3] (Figure 7B). The residual curing by monomeric SYS and YYS could be explained by their competing with endogenous Sis1 for interaction with Hsp70 or with [URE3] as a substrate.
Our curing data add to much previously published work [14], [15], [23], [45], [46] that support the explanation that overexpressing Ydj1 cures [URE3] by competing with Sis1. If so, then increasing abundance of Sis1 should allow it to compete more effectively for the disaggregation machinery and reduce the curing. In line with this prediction, overexpressing Ydj1 cured [URE3] much less effectively in cells with elevated expression of Sis1 (Figure 7C).
It remained possible that elevating Sis1 reduced this curing through some general stabilizing effect on [URE3]. However, if increasing Sis1 protects [URE3] from Ydj1 curing specifically by improving ability of Sis1 to compete with Ydj1, then increasing Sis1 would not be expected to protect [URE3] from being cured by other ways of impairing disaggregation machinery activity, such as inhibiting Hsp104. Overexpressing the dominant negative Hsp104-2KT mutant [5], which inhibits Hsp104 activity, cured [URE3] very effectively (Figure 7C). Elevating Sis1 expression did not affect this curing. Thus, Sis1 specifically counteracted curing by overexpressed Ydj1, which again is consistent with the idea that Ydj1 cures [URE3] by competing with Sis1 for interaction with the disaggregation machinery.
We repeated the plasmid shuffle in [psi−] [PIN+] strain 930a to assess effects of our panel of Sis1 mutants on propagation of [PIN+] prions (Figure 8). We monitor [PIN+] by the fluorescence status of Rnq1-GFP, which is regulated by the RNQ1 promoter on a single-copy plasmid. Rnq1-GFP is punctate in [PIN+] cells and diffuse in [pin−] cells.
[PIN+] propagated in cells with wild type Sis1 regardless of which of the mutant proteins was also expressed (Figure 8A). However, while wild type cells had single bright foci, those co-expressing the mutant proteins, except for Sis1ΔD and H34Q, had multiple foci (multi-dot). Therefore, several mutant Sis1 proteins dominantly affected propagation of [PIN+]. Cultures expressing Sis1-H34Q had a mixture of cells that possessed either single foci or completely diffuse fluorescence, indicating that H34Q inhibited the wild type Sis1 enough to cause [PIN+] to be lost from some cells. Combining ΔD or K199A with the H34Q mutation led to a multi-dot phenotype and reduced the proportion of [pin−] cells. These effects resemble the way the Sis1 mutants dominantly inhibited [URE3] and again show that the prion-curing effect of Sis1-H34Q depends partially on its dimerization and substrate-binding functions. Unlike [URE3], [PIN+] was not dramatically destabilized by co-expression of Sis1ΔGF, here regulated by the SIS1 promoter on a single copy plasmid. However, overexpressing Sis1ΔGF in [PIN+] cells of another strain background is toxic and causes [PIN+] to be lost [47].
In agreement with earlier work [23], [30], Sis1ΔCTD supported [PIN+] in cells without wild type Sis1, but Sis1ΔGF did not (Figure 8B). However, unlike the earlier work that showed Sis1ΔGMCTD (i.e. Sis1 JGF alone) propagated [PIN+], we did not observe [PIN+] foci in cells expressing Sis1ΔGMCTD. This difference might be due to differences in yeast strain backgrounds or by our variant of [PIN+] being more dependent on Sis1 for its propagation. In an earlier study the single-dot character of [PIN+] aggregates was frequently transformed by Sis1ΔGMCTD to multiple-dots, which were inheritable after transfer to wild type cells [30]. It is possible that the altered Sis1 activity causing this change is related to the loss of [PIN+] in our strains expressing Sis1ΔGMCTD, or that the action of Sis1ΔGMCTD on [PIN+] might be different in the two yeast strain backgrounds due to variation in expression of other chaperones.
[PIN+] also propagated stably enough to be detected among most cells expressing the substrate-binding and dimerization mutants (Figure 8B), showing that while these functions are important for [PIN+] propagation, [PIN+] depended less on these Sis1 activities than [URE3]. Keeping in mind that variations among strains of yeast and prions can influence prion stability, our data showing this intermediate sensitivity of [PIN+] to impairment of Sis1 activity is consistent with it being less sensitive than [URE3], but more sensitive than [PSI+], to curing by overexpressed Ydj1. Finally, we confirm earlier findings that DnaJB1 (Hdj1) supports propagation of [PIN+] [30], [32], [41].
Hsp40s bind misfolded proteins and regulate Hsp70 activity, so the ability of Sis1 and Ydj1 to specify functions of the disaggregation machinery are likely to be mediated through interactions with substrate and Hsp70. Our findings here show that C-terminal domains of Sis1 and Ydj1 can determine their functional differences in prion propagation, thermotolerance, galactose induction and their specific and general roles in supporting cell growth. Because the primary sites in Sis1 and Ydj1 that interact with substrates are contained within the CTDs, our data support the view that functional distinctions among Hsp40s can be due to differences in substrate specificity [34], [35].
The CTD of Sis1 also interacts with Hsp70, however, [47], [48] and although not fully characterized, this interaction likely influences Hsp70 functions. Likewise, functions influenced by the zinc-finger and farnesylation of the CTD of Ydj1 are important for the transfer of substrate to Hsp70 and for protecting cells from a [PIN+] prion-related toxicity [49], [50]. Also, Ydj1 can interact physically with Hsp104 in vitro [2], although the relevance of this interaction in the cell is unclear. While these other activities can be expected to contribute to specificity of these Hsp40s, our in-vivo and in vitro results indicate that the CTDs alone of Sis1 and Ydj1 allow them to discriminate between specific substrates, which is in line with earlier data [35].
A plausible explanation for the functional distinctions we observe would be that the CTD of Ydj1 interacts more readily with amorphous aggregates of stress-denatured proteins while that of Sis1 targets the more structured and homogeneous prion polymers. Sis1 and Ydj1 both bind to prion proteins, although Sis1 seems to bind more avidly, and prion proteins can differ in the number or location of general and distinct binding sites recognized by different Hsp40s [50]–[54]. Additionally, because substrate specificity of Hsp40s can overlap, competition among Hsp40s for substrates could contribute to determining functions of the chaperone machinery.
As seen earlier [47], we found the GF region can confer prion-specific Hsp40 functions. We show that the GF region of Sis1 was needed to propagate [URE3]. [PIN+] also depends on an activity of the Sis1 GF region that cannot be complemented by the Ydj1 GF [22], [23], [30]. However, all testable activities of Sis1 are dispensable for [PSI+] propagation [32], which shows that functions of the Sis1 GF are not necessary for propagation of all prions. Nevertheless, when appended to the JGF of their counterparts, the CTDs of Ydj1 and Sis1 generally were enough to allow the hybrid proteins to perform distinctly and effectively in place of intact Ydj1 and Sis1. Thus, the J and GF regions of Sis1 and Ydj1 possess activities that overlap enough to perform similarly in several distinct tasks. Evidently, more work is needed to learn how the GF region specifies Hsp40 functions in its effects on prions and perhaps other Hsp40-dependent cellular processes.
Much evidence points to Sis1 playing a key role in the replication of yeast prions by acting as a component of the Hsp104 disaggregation machinery that fragments prion fibers [14], [30], [46], [55], [56]. The varying degrees by which the prions depend on Sis1 agree with the supposition that different prions, and even different strains of the same prion (see [46], [57]), require varying degrees of disaggregation machinery activity to be fragmented. Together with the insensitivity of [PSI+] to Sis1 mutation, our finding that [URE3] is acutely sensitive to alteration in any Sis1 activity helps explain why depleting Sis1 causes cells to lose [URE3] much faster than they lose [PSI+] [14]. Our finding that [PIN+] had an intermediate dependency on Sis1 activity is also consistent with the intermediate rate of loss of [PIN+] seen upon Sis1 depletion. Overall our findings are consistent with earlier suggestions of a hierarchical dependency of these and other prions on the disaggregation machinery [57].
Extending this reasoning, our data strongly support an earlier suggestion that curing of [URE3] by Ydj1 or J-domains alone might be a result of competition for interaction with Hsp70 [14]. If Ydj1 cannot cooperate effectively with the disaggregation machinery to propagate [URE3], then by displacing Sis1 from the Hsp70 component of this machinery, less Hsp104 would be directed toward fragmenting prion polymers. This mechanism also explains why J-domains alone are enough to destabilize [URE3] and aligns with the idea that certain intact J-proteins don't cure as effectively because they are normally recruited to defined locations in the cytosol, such as ribosomes, by their other distinct functional domains. Among the three prions tested [URE3] is most sensitive to reductions in Sis1 function, so one might expect that its propagation would be most affected by such competition.
In line with a more stringent requirement of the disaggregation machinery for [URE3] replication, the average number of [URE3] prions per cell is lower than that for [PIN+] and [PSI+] [14], [58], [59]. The different seed numbers among variants of [PIN+] also could reflect differences in susceptibility to fragmentation, which in turn might underlie variation in sensitivity to curing by overexpressed Ydj1 [29]. Differences in susceptibility to fragmentation can be due to subtle differences in the structures of the amyloid that form the prions [60]. Such variation in the amyloid structures that determine differences among variants of [PIN+] might also have a bearing on the distinct pattern of variants of [PSI+] they induce [29], [61]–[63]. The similar intermediate sensitivity of [PIN+] to depletion of Sis1 seen in earlier work and to specific mutations of Sis1 seen here suggests the variants of prions used are similar and that their prion character is largely independent of strain background. Nevertheless, findings might differ if other variants of prions or other strain backgrounds propagating them were compared directly.
It is becoming evident that the differences in ways prions respond to J-proteins and other Hsp70 co-chaperones likely reflect differences in the ways prions depend on Hsp70. Altering activity of Hsp70 directly by mutation or indirectly by altering its co-chaperones can influence prion propagation in the same ways, which supports this idea [64]–[66]. Stability of [SWI+] prions is also highly sensitive to altered expression of Hsp40s and J-domains, which seems to be related to a strict dependency on optimal Hsp70 activity [67]. Yeast has four highly homologous Ssa Hsp70 paralogs and prion phenotypes vary greatly when different Hsp70s are the sole source of Ssa protein [43]. These differences likely reflect differences in the way the Hsp70s interact with or are regulated by the Hsp40s or other factors. Hsp70 also can be a primary factor in recruiting the disaggregation machinery to prion polymers [7]. Notably, however, it is not Hsp70 substrate-binding per se, but the regulation of this binding, presumably by co-chaperones, that specifies distinctions in Hsp70 functions with regard to [URE3] propagation [44]. NEFs can also affect prion propagation through their ability to regulate Hsp70 [13], [68], [69].
Because Hsp70 is a critical component of the Hsp104-based disaggregation machinery, altering Hsp70 or its co-chaperones can be expected to affect propagation of prions by influencing composition and activity of this machinery. The distinct susceptibilities of prions to alterations in various disaggregation machinery components might therefore reveal differences in the ways various chaperones combine to act most effectively on them as specific substrates. Understanding why prions respond differently to the various chaperone machinery components including J-proteins, NEFs and Hsp70s should help us understand both fundamental and subtle ways that these components interact to produce effective protein remodeling machines.
Yeast strains used are isogenic to strain 779-6A (MATa, SUQ5, kar1-1, ade2-1, his3Δ202, leu2Δ1, trp1Δ63, ura3-52) [70], which is used for monitoring [PSI+] and [PIN+]. Knockouts and replacements of chromosomal genes were done using standard transformation procedures [71]. Strain MR386 has E. coli CLPB in place of the chromosomal HSP104 gene [15] and contains plasmids expressing E. coli dnaKR167H (pMR150LG-R167H) and E. coli GrpE (pMR134H) under the control of the GPD (glyceraldehyde-3-phosphate dehydrogenase - TDH3) and FES1 promoters, respectively. [PSI+] is maintained in strain MR386 by pJ312, which is HSP104 on a URA3 plasmid [72]. Strain MR502 has ydj1::KanMX and carries p316YDJ1, which is YDJ1 on a URA3 plasmid. [PSI+] [PIN+] strain 930 has sis1::KanMX and carries plasmid pYW17, a URA3-based plasmid encoding wild type SIS1 [22], [32].
Strain 930a is a [psi−] [PIN+] version of 930 that carries plasmid p313Rnq1-GFP. It was cured of [PSI+] by transient growth on medium containing 3 mM guanidine and then [PIN+] clones among [psi−] isolates were identified by punctate Rnq1-GFP fluorescence. Our [PIN+] variant is uncharacterized, but of the single-dot type, which typically has sturdier fibers and a lower seed number per cell than multi-dot [PIN+] [61]. Isogenic strain 1075, for monitoring [URE3], has ADE2 regulated by the DAL5 promoter (PDAL5::ADE2, see below) in place of ade2-1 [43]. Strain 1385 is strain 1075 with sis1::KanMX and plasmid pYW17, which is SIS1 on a URA3 plasmid. Strains 1408 and 1410 are hsp104Δ versions of strain 779-6A and 1075, respectively [15]. Both carry pJ312. Our parental strains carry only one variant of [PSI+], [URE3] or [PIN+].
SIS1 plasmids used contain wild type and mutant SIS1 alleles on the pRS314 single-copy TRP1 vector [22], [32]. Plasmid pRU4 is LEU2-based single-copy plasmid pRS415 containing the GAL1 promoter and CYC1 terminator flanking the polylinker sites SpeI and XhoI. For Gal-induced expression, YDJ1 and hybrid alleles were inserted into pRU4 as BamHI-SalI fragments. All plasmids used in this study are listed in Table 1. Plasmids encoding E. coli genes or yeast Hsp40 genes with D36N mutations are described [15].
With the exceptions that 1/2YPD plates contain 5 g/L yeast extract, YPAD plates contain 400 mg/L (excess) adenine and solid defined media contain 10 mg/L (limiting) adenine, standard media and growth conditions were used [71].
Sis1 and Ydj1 have clearly defined and characterized J-domains, glycine-phenylalanine (GF) rich middle domains, and a C-terminal region that contains two major elements (CTDI and CTDII) and a dimerization domain. The main structural differences between Sis1 and Ydj1 are a glycine-methionine-rich (GM) extension of the GF domain in Sis1, which has GF-redundant functions, and a Zn-finger domain (ZF) embedded between beta-strands 2 and 3 of the CTDI of Ydj1 that is absent in Sis1 [73]. Because the ZF of Ydj1 is an integral part of CTDI, we designed our Sis1-Ydj1 hybrids using the GF-CTDI junction to restrict the number of domain swaps and avoid using complicated junctions to swap the ZF region. Rather than designating the GM region as a separate domain, we combined the functionally redundant GF and GM regions of Sis1 into a single domain. Thus, hybrid alleles were made by swapping three regions: the J-domain, the glycine-rich region and the C-terminal portion, which includes CTDI, CTDII and the dimerization domains, herein referred to simply as the CTD (see Figure 1A, [30]). Hybrid genes were synthesized by GENEWIZ, Inc (South Plainfield, NJ) and sub-cloned into variants of pRS414 that placed the ORF under the control of the SIS1, YDJ1 or GPD promoters and a downstream CYC1 transcriptional terminator [74]. All constructs contained a c-terminal c-myc tag that had no noticeable affect on functions in vivo.
Depletion of the ribosome release factor Sup35 by its sequestration in [PSI+] prion aggregates causes nonsense suppression. We monitored [PSI+] by suppression of the ade2-1 nonsense allele in our strains. [PSI+] cells are Ade+ and white, while [psi−] cells are Ade− and when grown on limiting adenine are red due to accumulation of a metabolite of adenine biosynthesis. The presence of [URE3] was monitored similarly by use of an ADE2 allele regulated by the DAL5 promoter (PDAL5::ADE2) [75], [76]. Under standard growth conditions Ure2 represses transcription of nitrogen metabolic genes, such as DAL5. [URE3] sequesters Ure2 into prion aggregates, thereby reducing Ure2 function and activating the DAL5 promoter. Thus, [URE3] cells are Ade+ and white, while [ure-o] cells are Ade− and red on limiting adenine. We confirmed that Ade+ phenotypes were due to the presence of prions by their guanidine curability and by crosses with cells lacking prions to produce a dominant, guanidine-curable Ade+ phenotype. We monitored [PIN+] by assessing aggregation status of a plasmid-expressed Rnq1-GFP fusion protein. GFP fluorescence is diffuse in [pin−] cells, but noticeably punctate in [PIN+] cells. In this study we used a typical strong [PSI+] strain and single variants of [URE3] and [PIN+] prions.
Microscopic analysis of Rnq1-GFP fluorescence in live cells was done with a Nikon E-800 microscope with log phase cells grown in medium selecting for the plasmid encoding the Rnq1-GFP fusion protein. Images were captured using IVision software and processed using Adobe Photoshop software.
Log phase cells grown in medium selecting for plasmids were diluted in fresh medium to an OD600 of 0.25 and 100 µL was transferred to 0.5 mL test tubes and placed in a PCR machine for thermocycling as indicated. At various times aliquots were removed and placed on ice. Cooled cells were serially diluted and 5 µL drops were spotted onto YPAD plates.
Strains 930 and 1385 (both sis1Δ) and derivatives of MR502 (ydj1Δ) carrying wild type SIS1 or YDJ1 on URA3-based plasmids were transformed by TRP1-based plasmids carrying wild type, mutant or hybrid alleles. Strain MR502 also carries pMR169 for monitoring GAL induction (see below). Transformants were grown as patches of cells on medium lacking both tryptophan and uracil and then replica-plated onto similar medium containing uracil to allow loss of the URA3 plasmid. These were then replica-plated onto medium containing 5-FOA, which kills cells that did not lose the resident URA3 plasmid.
For Sis1, growth on 5-FOA plates shows complementation of functions essential for growth, and growth without adenine shows complementation of Sis1 functions required for prion propagation. To test complementation of Ydj1 function, 5-FOA resistant cells of strain MR502 were grown overnight in medium selecting for the TRP1 plasmids, normalized to the same cell density (OD600 = 0.25) and five-fold serially diluted. Aliquots of the dilutions (5 µL) were then dropped onto YPAD plates. Scanned images of the plates were taken after they were incubated at the indicated temperatures for 3–4 days.
Aliquots of overnight cultures of MR502 transformants used for growth complementation were transferred to synthetic raffinose medium (SRaf) and grown overnight. These cells carried the TRP1-marked hybrid alleles or empty vector control and a HIS3-marked plasmid encoding firefly luciferase under the control of the GAL10 promoter (pMR169). Cell densities were adjusted to OD600 = 0.3 in fresh medium and the initial reading (t = 0) was taken by mixing 100 µL culture with 50 µL of 1 mM luciferin in 0.1 M sodium citrate, pH 5.0 immediately before reading in a Zylux Femtomaster luminometer, with a 10 s delay and 5 s read time. Galactose from a 20% stock was then added to the cultures to obtain a final concentration of 2% and the cells were incubated on a roller at 30°C. Readings were taken at 30, 60 and 120 minutes after addition of galactose. All samples in three experiments were tested in triplicate. No significant growth occurred during the course of the experiment.
Overnight SD cultures of cells carrying YDJ1 or hybrid alleles on pRU4 for galactose induction were used to inoculate SGal medium to OD600 = 0.05 and incubated with shaking at 30°C. Generations were monitored as doublings of OD600. Cultures were sub-cultured as necessary to keep the OD600 less than 2.0. After 3, 6 and 9 generations of growth in galactose, aliquots were removed and cells plated onto 1/2YPD plates. Loss of [URE3] was assessed by determining the fraction of entirely red (i.e. [ure-o]) colonies after 3 days of incubation at 30°C. To test the ability of overexpressed Sis1 to block Ydj1- or Hsp104- mediated curing of [URE3], strain 1075 was co-transformed with various combinations of CEN plasmids expressing Sis1, Ydj1 or Hsp104 under the control of the GPD promoter.
Cell lysates for western blots were prepared as described [77]. Briefly, cells were suspended in lysis buffer and broken by agitation with glass beads. For each sample 10 µg of protein was separated on 4–20% SDS-PAGE gels, transferred to PVDF membranes and probed using anti-c-myc antibody (AbCam #ab9106) and chemiluminescence. After developing, the blots were stained by amido-black (Sigma #A-8181) as a loading and transfer control.
Hsp104 [78], Ydj1 [79], and GFP-38 [80] were purified as described. SSY was isolated as described for Ydj1 [79]. Sis1 was purified as described [81] with some modifications. Briefly, BL21 (DE3) was transformed with a pET11 plasmid containing the Sis1 gene, cultures were grown at 30°C to OD595 0.8 and cells were induced with 1 mM IPTG for 3 h. Clarified cell lysates were applied to an S-Sepharose-FF column (GE Healthcare) in 20 mM MES buffer, pH 6.0, 0.1 mM EDTA and 1 mM DTT. Sis1 was eluted with a linear gradient from 0–1 M NaCl. Peak fractions were pooled, buffer exchanged into 20 mM MES buffer pH 6.0, 0.1 mM EDTA and 1 mM DTT. The sample was applied to a monoS column (GE Healthcare) and eluted with a linear gradient from 0–1 M NaCl. YYS was purified similarly to Sis1, except that the buffer used was 25 mM HEPES, pH 7.6, 0.1 mM EDTA and 1 mM DTT. For Ssa1, a pET24 plasmid containing the Ssa1 gene was transformed into Rosetta BL21 (DE3) cells. Cultures were grown to 0.8 OD595 at 30°C and induced with 0.2 mM IPTG overnight. The clarified lysate was applied to a Q-sepharose FF column (GE Healthcare) in 20 mM Tris.HCl, pH 7.6, 40 mM KCl, 0.1 mM EDTA and 1 mM DTT. Ssa1was eluted with a linear gradient of 40–400 mM KCl over 20 column volumes. Peak fractions were collected and buffer exchanged into 25 mM HEPES, pH 7.6, 100 mM KCl, 0.1 mM EDTA and 1 mM DTT and further purified over a monoQ column (GE Healthcare) using a linear gradient of 100–400 mM KCl over 20 column volumes. Peak fractions were collected, analyzed, supplemented with 10% glycerol, frozen on dry ice and stored at −80°C.
|
10.1371/journal.pcbi.1003012 | Phosphorylation of the Retinoic Acid Receptor Alpha Induces a Mechanical Allosteric Regulation and Changes in Internal Dynamics | Nuclear receptor proteins constitute a superfamily of proteins that function as ligand dependent transcription factors. They are implicated in the transcriptional cascades underlying many physiological phenomena, such as embryogenesis, cell growth and differentiation, and apoptosis, making them one of the major signal transduction paradigms in metazoans. Regulation of these receptors occurs through the binding of hormones, and in the case of the retinoic acid receptor (RAR), through the binding of retinoic acid (RA). In addition to this canonical scenario of RAR activity, recent discoveries have shown that RAR regulation also occurs as a result of phosphorylation. In fact, RA induces non-genomic effects, such as the activation of kinase signaling pathways, resulting in the phosphorylation of several targets including RARs themselves. In the case of RARα, phosphorylation of Ser369 located in loop L9–10 of the ligand-binding domain leads to an increase in the affinity for the protein cyclin H, which is part of the Cdk-activating kinase complex of the general transcription factor TFIIH. The cyclin H binding site in RARα is situated more than 40 Å from the phosphorylated serine. Using molecular dynamics simulations of the unphosphorylated and phosphorylated forms of the receptor RARα, we analyzed the structural implications of receptor phosphorylation, which led to the identification of a structural mechanism for the allosteric coupling between the two remote sites of interest. The results show that phosphorylation leads to a reorganization of a local salt bridge network, which induces changes in helix extension and orientation that affects the cyclin H binding site. This results in changes in conformation and flexibility of the latter. The high conservation of the residues implicated in this signal transduction suggests a mechanism that could be applied to other nuclear receptor proteins.
| Allosteric regulation of proteins is critically important in many biological processes. Here we focused on the allosteric pathway of communication within a ligand-regulated transcription factor, the Retinoic Acid Receptor (RAR). Recent experimental studies performed with the RARα subtype have shown that phosphorylation of a residue located at one extremity of an α-helix in RAR, leads to a changes in binding affinity at the other extremity of the same helix for cyclin H, a binding partner that is necessary for gene transcription activation. The purpose of our study was to understand the conformational changes occurring within the receptor upon phosphorylation. Molecular dynamics simulations are well suited for this sort of study. Through this approach, we were able to show that although the overall structure of the phosphorylated RAR shows no distinct difference from the unphosphorylated form, evidence is provided for an allosteric regulation pathway that implicates more subtle changes, such as changes in side chain orientations, which affect the internal dynamics of the receptor.
| Nuclear receptors are ligand-dependent transcription factors that participate in many cellular signaling networks involved in various physiological phenomena, such as embryogenesis, cell differentiation, cell growth, reproduction and apoptosis [1]. Disruption or abrogation of these signaling pathways results in a variety of diseases or in malignant cell transformation. The superfamily of nuclear receptors includes the nuclear retinoic acid receptors (RARs), which bind retinoic acid (RA), the active metabolite of vitamin A and which function as heterodimers with a second family of nuclear receptors, the retinoid X receptors (RXRs). There are three RAR and RXR subtypes (RAR-α, -β, -γ and RXR-α, -β, -γ) [2] and they regulate gene expression by binding as RXR/RAR heterodimers to retinoic acid response elements (RAREs) located in the promoter regions of target genes [3]–[5]. RARs and RXRs display a well-defined domain organization, composed mainly of an unstructured N-terminal domain (NTD) and two well-structured domains, a central DNA-binding domain (DBD) and a C-terminal ligand-binding domain (LBD). The DBD is composed of two zinc finger motifs and two α-helices. The LBD is formed by 12 α-helices and one β-sheet, which display the general three layer antiparallel helical sandwich fold found in the NR superfamily (see Figure 1).
The general scenario of RAR activation starts with the binding of the ligand to the LBD and the subsequent departure of co-repressor and recruitment of co-activator proteins [3]. Crystallographic structures and the characterization of co-regulatory complexes, such as Topoisomerase II [6] for repression and DNA-dependent ATPases for activation processes [7] have provided a wealth of information on how RARs regulate transcription. Multiple structures of individual DBDs and LBDs have been determined [8]–[12]. Recent X-ray structures [11], solution state structures by SAXS [13] and cryo-EM structures [14] of full-length nuclear receptor protein complexes continue to increase our understanding. The binding of agonist ligands, such as retinoic acid (RA), to the LBD induces large-scale conformational changes, the most prominent being the repositioning of the C-terminal helix H12. This particular structural rearrangement results in the exposure of a binding site for co-activator proteins that contain the LXXLL consensus motif [15]. Allosteric communication pathways have been identified in silico between the ligand and the co-activator peptide [16] and between functionally relevant protein interfaces [17]. This fine-tuned regulation process leads to the recruitment of several protein complexes with enzymatic activities, such as histone actelytransferases and DNA-dependent ATPases [18], which subsequently induce alterations in the chromatin structure around the target genes promoters.
Besides this classical mode of genomic effects, RA also has non-genomic and non-transcriptional effects exemplified by the activation of p38MAPK/MSK1 pathway [19]–[21]. Activation of this pathway results in the phosphorylation of the RARα ligand-binding domain at serine 369 (S369), located in loop L9–10 within the LBD [22]. This post-translational modification leads to an increase in the binding affinity of the LBD domain for cyclin H [23]. Cyclin H, together with cdk7 and MAT1, form the Cdk-activating kinase (CAK) subcomplex of the general transcription factor TFIIH, which is involved in transcription initiation and DNA repair. Binding of cyclin H to the RARα LBD positions the CAK complex so that the cdk7 kinase can phosphorylate a second serine of RARα (S77) located in the NTD [24], [25]. The phosphorylation of this N-terminal residue is required for the recruitment of RARα to target genes promoters [21].
In a recent study, we showed that phosphorylation of S369 leads to changes in the structural dynamics of the cyclin H binding site, composed of the loop between helices 8 and 9. This change in dynamics was correlated to the increase in the cyclin H/RARα binding affinity [26]. Given that S369 is located almost 40 Å from the binding site of cyclin H, an allosteric mechanism clearly makes an important contribution to this overall process. Interestingly, S369 is essentially absent outside of mammalian RARα indicating that this fine-tuned phosphorylation cascade appears late during vertebrate evolution [25].
In this work, we use molecular dynamics simulations to elucidate the effects of phosphorylation on the conformational flexibility and dynamics of the RARα LBD and to identify the factors that compose the allosteric signal. An extensive analysis of the effects of phosphorylation on atomic fluctuations, salt bridges and ion-pair networks was coupled to quasi-harmonic and correlated motions analyses. The aim was to identify the consequences of phosphorylation on the structural dynamics. From this analysis, we proposed a coherent and complete picture of the allosteric communication between the S369, the phosphorylation site, and the cyclin H binding site.
As a measure of the structural stability of RARα, the overall root mean square deviation (RMSD) of the backbone coordinates was calculated as a function of time and averaged over the three simulations of the phosphorylated and non-phosphorylated protein, respectively. This measure was done relative to the initial energy-minimized structure, which was the same in all simulations. The time series were calculated over the 50 ns of dynamics (see Figure S1) and showed characteristic behavior of stable simulations with no overall drift. Mean values of 1.13 Å and 1.14 Å for phosphorylated RARα (p-RARα,) and unphosphorylated RARα (unp-RARα), respectively, were calculated, suggesting that phosphorylation does not induce major structural changes implicating backbone reorganization.
This stable behavior is coherent with the cluster analysis of the trajectories [27], [28] that was performed in order to evaluate whether the trajectories of the phosphorylated and unphosphorylated forms converge to a representative ensemble of structures. The detailed description is given in Material and Methods. Indeed, the cluster analysis indicated good convergence of the structural ensembles on the timescale of the simulations.
Finally, the local structural deviations were also analyzed from average structures extracted from the last 10 ns of the simulations. The local backbone RMSD from the initial structure was calculated on a per-residue basis (see Figure S2). No major differences were observed between the phosphorylated and unphosphorylated forms, except in the vicinity of S369 (i.e. loop L9–10) where the RMSD is smaller in the case of the phosphorylated protein. This suggests that phosphorylation stabilizes to some extent loop L9–10.
Together, these results suggest that phosphorylation of S369 does not lead to any significant changes in the overall conformation of RARα whereas only small, localized changes in the receptor's backbone conformation occur in the loop L9–10.
Salt bridges play an important role in nuclear receptor structure. A recent structure-based sequence analysis revealed differentially conserved salt-bridges that partition the NR superfamily into two classes related to their oligomeric behavior [29]. Heterodimer-forming receptors, such as RARα belong to class II where, following the nomenclature in [29], conserved salt bridges are formed between, i) E/D42 in H5 and R62 in loop L8–9 and, ii) E50 in H8 and R/K/H90 in H9. Interestingly, these conserved salt bridges involve residues situated in the vicinity of the phosphorylation site (loop L9–10) and of the cyclin H docking site (loop L8–9) [23], [29], suggesting their possible involvement in the allosteric mechanism between the two sites.
Here we analyzed the impact of S369 phosphorylation on the formation of salt bridges networks in the RARα LBD. The analysis was performed by monitoring the distances between carbon atoms of partner residues within a salt bridge (Cζ atom of Arg, Cε of Lys, Cδ and Cγ for the Glu and Asp, respectively). We monitored all salt bridge/ion pairs of the receptor LBD and found that five out of a total of 35 were affected upon S369 phosphorylation (see Figure 2 for a representation of the amino acids involved and Figure 3 for the associated structural changes upon phosphorylation).
A first group of affected interactions is situated in the vicinity of the phosphorylation site. Phosphorylation of S369 introduces a negative local charge that is felt by the nearby R367 (see Figure 2.A for the structure) and as a result, conformations with shorter distances between the Oγ of S369 and Cζ of R367 are observed in the phosphorylated receptor as compared to the non-phosphorylated (see Figure 3.A). This reorientation of R367 affects its interactions with nearby acidic amino acids. The hetero-dimer specific [29] salt bridge E325–R367 (E42–R62 in the unified nomenclature above, see Figure 2.A) is not systematically formed in the phosphorylated receptor, indeed the salt bridge separation distances that fall between 3.5 and 4.5 Å account for 65% of the population in the p-RARα simulations, as opposed to 99.8% in the non-phosphorylated form (see Figure 3.B). In addition, a new maximum located between 6 and 6.5 Å with a population of 22% is observed in p-RARα (see Figure 3.B).
The disruption of the class-specific salt bridge is accompanied by the formation of a new ionic interaction between R367 and E320, which is situated close to E325, but oriented towards the surface of the receptor LBD (see Figures 2.A, 2.B and 3.D). This is reflected in the dominant distance peak moving from 8–9 Å (42% of the conformations) in the unp-RARα simulations to 6–7 Å (35% of the conformations) in the p-RARα simulations (see Figure 3.d). In addition, a new minimum separation distance between 4–5 Å appears with p-RARα (11%, see Figure 3.D). Another affected salt bridge in the phosphorylation region concerns E323 and R192. E323 is situated in H8 while R192 is in H1, in the N-terminal region of the receptor (see Figure 2.B). Phosphorylation leads to a decrease of the separation distance between these two residues. In comparison to the distance distribution in the unphosphorylated form, a new peak containing 9.5% of the conformations appears between 3 to 4 Å, whereas the dominant peak between 7 to 8 Å decreases from 59% in the unp-RARα form to 35% in p-RARα (see Figure 3.C). Overall, distances that are above 7 Å account for 80% of the conformations in unp-RARα and 77% in p-RARα.
Besides these four perturbed ionic interactions in the vicinity of the phosphorylation site, an ionic interaction formed between D256 in the N-terminal part of H4 and R347 in the N-terminal region of H9 (see Figures 2.A and 2.C) is perturbed upon phosphorylation as well. In the unphosphorylated simulations, the salt bridge D256–R347 is formed in 21% of the conformations with short (3 to 4 Å) D256-Cγ to R347-Cζ distances (see Figure 3.E). Upon phosphorylation, the D256-Cγ to R347-Cζ distances increase to values above 4 Å (see Figure 3.E). The separation distances larger than 5 Å represent 82% of all values in the phosphorylated form of RARα, as opposed to 62% in the non-phosphorylated form of RARα (see Figure 3.E). These results suggest that phosphorylation of S369 induces weakening of the R347–D256 salt bridge. In addition, when the distance between D256 and R347 increases, the donor-H-acceptor angle between the two N atoms of R347 and the two O of D256 is disrupted and becomes lower than 120°, further supporting the breaking of the salt bridge in the phosphorylated simulations (data not shown). An important point is that this latter salt bridge is situated in the region of the cyclin H binding site, about 40 Å from the phosphorylated serine.
Together, the modified ionic-pair patterns show that the phosphorylation has a tendency to remodel the network of salt bridge and ion-pair interactions near the phosphorylation site. These changes result in an approaching of the C-terminal end of H8 to H1, as well as of L9–10 to the C-terminal part of H7. By contrast, phosphorylation leads to changes around the cyclin H binding site, where salt bridge distances increase between the N-terminal part of H4 and the N-terminal part of H9, suggesting a loosening of the structure in this region. These changes are illustrated in Figures 2.B and C where the blue lines represent distances that increase and red lines represent distances that decrease upon phosphorylation of RARα.
Sequence analysis of nuclear receptor LBDs revealed that two sets of differentially conserved residues implicated in salt bridge formation partition the NR superfamily into two classes related to their oligomeric behavior [29]. Following a similar line of reasoning, from the alignment performed in Brelivet et al. [29] we extracted information concerning the amino acids implicated in salt bridge changes observed in the molecular dynamics simulations (see Figure 4). Interestingly, residues implicated in the salt bridge and ion-pair changes upon phosphorylation are mostly conserved in heterodimeric receptors, of which RARα is one. The most conserved residues are D256 and E325 (see Figure 2 and Figure 4) with the percentage 91% and 63%, respectively in heterodimers, and correspond to two important residues in the sequences of nuclear receptors [29]. The other residues D323, R347, E320 and R367 (this latter one conserved as a basic residue R, K, or H) also display an important percentage of conservation in heterodimers with values of 53%, 23%, 41% and 48% respectively.
The effect of the S369 phosphorylation on the helix H9 was analyzed by quantifying the helix bend. We estimated the radius of a sphere needed to encircle the Cα atoms of the helix in the non-phosphorylated and phosphorylated forms using the TRAJELIX module [30] in the Simulaid software [31]. A bent helix with a particular sequence requires a smaller sphere than a perfectly extended helix of the same sequence. The average radii of the encircling spheres are 18.20±0.45 Å in the unp-RARα LBD and 18.43±0.34 Å in the p-RARα LBD simulations (see Figure 5). The p-value calculated on these two ensembles using a Student's test is <2.2e−16, supporting the subtle differences between the values of the phosphorylated and the non-phosphorylated forms. In Figure 5, we see that 87% of the calculated radii are above 18 Å for p-RARα as opposed to 64% in unp-RARα. This increase in the radius of the encircling sphere fitted to H9 thus indicates the tendency to decrease the bend of the helix, and, by consequence, the extension of helix H9 upon phosphorylation. The individual distributions from the unphosphorylated and phosphorylated simulations are given in Figure S3.
To further investigate the consequences of phosphorylation, we calculated the relative angles formed between helix H9 and helix H10 and between helix H9 and helix H4. The orientations of the helices were calculated using the Chothia-Levitt-Richardson algorithm [32] as implemented in the CHARMM program. Vectors illustrating the angles calculated are presented in Figure S4. Concerning the angle between helices H9 and H10 (see Figure 6.A), we observe a general decrease upon phosphorylation. Indeed, in p-RARα 32% of the conformations had angle values over 50°, while in unp-RARα, 47% of the conformations had angle values above 50°. The average angles between H9–H10 are 48.7±3.7° and 50.0±3.6° for p-RARα and unp-RARα, respectively. The individual angle distributions for the inter-helix angles for H9–H4 and H9–H10 are given in Figures S5.A and B, respectively.
For the angle between H9 and H4 (see Figure 6.B), phosphorylation induced an increase in the angle, with a displacement of the distribution to higher values. Indeed, 80% of the p-RARα conformations have an H4–H9 angle value greater than 50°, as opposed to 70% of unp-RARα conformations. The average values were 53.5±4.5° for the phosphorylated form and 52.2±4.0° for the unphosphorylated one.
We also calculated the values of the angle formed between H8 and H10 (Figure S6). The results show that the average value of the angle between H8 and H10 is 52.1±4.6° in the case of the unphosphorylated receptor and 49.3±3.9° for the phosphorylated form. The percentage of values greater than 50° is 65% for unp-RARα and 40% for p-RARα. This clearly demonstrates that the orientation of H10 with respect to H9 discussed previously is also applicable with respect to H8, supporting the decrease in the angle formed by vectors H8–H10 and H9–H10.
For all angle averages cited above, the Student's t-test yielded p-values less than 2.2e−16 so they can be considered to be statistically different. In both cases, one can observe a significant amount of overlap of the distributions, with a slight tendency toward larger (or smaller) values, depending of the angle, as a function of the phosphorylation state. Although small, this shift in population is not inconsistent with the published experimental observations [26], which show that, first of all, un-phosphorylated RARα LBD binds cyclin H, and second, an increase in the binding affinity results from phosphorylation. This is embodied in these figures, which show largely overlapping distributions, but with a slight shift in populations upon phosphorylation.
Overall, the results presented here suggest that the changes in the salt bridge networks involving loop L9–10 lead to a decrease of the angle between helices H9 and H10, which induces a detachment of the loop L8–9 from the helix 4 in the core of the receptor. This, in turn, results in an increase of the angle formed between H9 and H4.
To determine if the local changes in the RARα LBD structure described above also affect the structural dynamics and low frequency vibrational modes of the LBD, we analyzed its local and collective structural fluctuations.
In order to characterize the local atomic level flexibility, we calculated the by-residue-averaged atomic root mean squared fluctuations (RMSf) from the individual simulations for the backbone atoms. The results show that for the unp-RARα, the RMSf are in general agreement with the trend determined from the experimental B-factors (See Figure 7.A). This analysis further indicates that upon phosphorylation, there is an increase in the fluctuations in the region of the loop L8–9 and the N-terminal part of H9. On the other hand, loop L9–10, where S369 is located, shows a decrease in flexibility upon phosphorylation. This is consistent with the smaller RMSD value for some residues of this loop (see Figure S2). For the purpose of illustration, the RMSf values for residue D341 of loop L8–9 are 1.28 Å and 1.13 Å for p-RARα and unp-RARα, respectively. For residue R370 of loop L9–10, the fluctuations are of 1.2 Å and 1.54 Å for p-RARα and unp-RARα, respectively. Recalling the analysis of structural changes presented above, we found minimal conformational changes upon phosphorylation, however, there is a more measurable difference in the local dynamics, particularly for loop L8–9 and the N-terminal region of H9, as measured by the RMSf. This observation correlates well with the decompaction of the structure in the region of the cyclin binding site discussed above.
Further analysis indicated that this increased flexibility is associated with small, but significant modification of the low frequency motions of the LBD. Quasi-harmonic analyses on the six molecular dynamics of unp- and p-RARα were done in order to characterize the low-frequency motions. The low-frequency modes describe the collective motions of the protein. Averaging over the results from the three simulations of each phosphorylation state, we found the lowest three frequency modes to be 1.13, 1.51, 1.87 cm−1 for unp-RARα and 1.07, 1.61 and 1.87 cm−1 for p-RARα. Overall, these low frequencies do not differ much between the two forms, which is consistent with our earlier observations that phosphorylation did not alter the dynamics in any major way [26].
From the ten lowest frequency modes of the three phosphorylated and three unphosphorylated simulations, we calculated the RMS fluctuations of the backbone and averaged by residue (Figure 7.B). Interestingly, the fluctuations of loop L8–9 increase measurably when calculated from the low-frequency modes of the p-RARα with respect to the low-frequency modes of unp-RARα. For example, the average fluctuations in the cyclin H docking site reach the value of 0.58 Å in p-RARα in comparison to 0.46 Å in unp-RARα (see Figure 7B). This indicates that phosphorylation affects the low frequency dynamics, which, in turn, modulates the dynamics of the cyclin H binding site.
To further characterize the structural dynamics of the unp- and p-RARα LBDs, cross-correlation (CC) coefficients were calculated for the six trajectories as described in Material and Methods. These measures range from −1 to 1 and provide information on correlated internal motions of the receptor. In Figure 8, we show a spider web diagram that represents the association of amino acids that are correlated in their motions. The detailed correlation matrix is given in the supplementary material (Figure S7). From this data, one can extract detailed information concerning pathways of correlated motions. What we observe is that upon phosphorylation, the correlation network in the LBD is altered and in particular for loops L8–9 and loop L9–10.
The nuclear receptor RARα is an important player in the regulation of gene transcription and many structure-based studies corroborated the role of ligand binding in the regulatory mechanisms [33]–[36]. More recently, phosphorylation at specific residues also proved to be important and attracted significant attention [37]–[39]. In a non-genomic cascade, RA triggers the activation of the MSK1 kinase, which then phosphorylates S369 located in the LBD. Experimental observations have shown that this phosphorylation leads to an increase in the affinity of the LBD for cyclin H, a component of the cdk7/cyclin H/MAT1 subcomplex of TFIIH. The recruitment of the TFIIH complex through the binding of cyclin H leads to the subsequent phosphorylation by cdk7 of a second serine located in the NTD, and finally the recruitment of the receptor to DNA, a necessary step for gene expression. According to the experimental results [22], [23], phosphorylation increases the binding affinity of cyclin H in vivo and in vitro. However, no data was available that sheds light on how this phosphorylation structurally activates the cyclin H binding site in RARα. In this context, the present work elucidates details of the allosteric mechanism that is activated upon phosphorylation of S369 of the LBD and results in the enhanced affinity for cyclin H.
In this study we asked the specific questions (i) how phosphorylation of a serine residue in RARα can induce changes in the binding domain of cyclin H, which is located approximately 40 Å away and (ii) what is the mechanism of this allosteric regulation process arising from the post-translational phosphorylation.
Using molecular dynamics simulations, we studied both the structure and dynamics of the RARα LBD. In order to expand the conformational space sampled, we ran multiple molecular dynamics simulations of the phosphorylated and unphosphorylated forms of the LBD, yielding a total of 150 ns for each form of the receptor. From the molecular dynamics simulations, we analyzed the structural and dynamical changes related to phosphorylation in terms of salt bridge and ion-pair formation, helix bending and angles between helices, correlation networks and internal motion.
Phosphorylation is often used as a trigger of allosteric signaling, with signaling mechanisms associated with conformational [40] as well as dynamical changes [41]. In the case of nuclear receptors, experimental data on the estrogen receptor showed very small structural differences between the non-phosphorylated and phosphorylated forms of the LBD (RMSD of 0.59 Å) [42]. X-ray structural analysis of a mutant S/E of RARγ, an RAR subtype highly homologous to RARα, indicates minor structural changes upon introduction of the glutamic acid in loop 9–10 of the LBD, which emulates the phosphorylation (S. Sirigu et al, ms. in preparation). In coherence with the experimental observations, the molecular dynamics simulations of RARα show that no large-scale conformational changes of the LBD occur upon phosphorylation. This suggests that, in the case of the RARα LBD, we have an example of allostery in the absence of significant conformational change.
Allostery can be viewed as a redistribution of an ensemble of pre-existing states upon effector stimulation, with a shift toward a particular state that is favorable, for example, for signal transduction [43], [44]. This view has modulated over the years as a result of the significant improvements in both theoretical and experimental methods and the role of dynamics in allosteric signal transmission is now fully appreciated [45]–[48]. A theoretical background for allosteric signaling in the absence of conformational changes was put forward by Cooper and Dryden in 1984 [49], and experimental examples have since been reported [50], [51].
In the case of the RARα LBD, although no global changes occur, small structural changes near both the phosphorylation site and the cyclin binding site were observed. Phosphorylation modulates the network of salt bridge composed of conserved residues and induces a subtle reorientation of helices H9 and H10 as measured by changes in relative angles. This results in the structural modification of H9 corresponding, in essence, to a straightening of the helix. This straightening of H9 alters the relative position of the N-terminal end of H9 that, along with loop L8–9, makes up the cyclin H binding site. The distance between the N-terminal part of H9 and the C-terminal part of H4 in the phosphorylated RARα is increased, which opens up the region of loop L8–9 and the N-terminal part of H4. These changes were also characterized by the increase in angle formed between H9 and H4.
Taken together, the consequences of the salt-bridge reorganization are small and induce subtle structural changes. Coupling these small population shifts to an allosteric mechanism is not without precedent. Indeed, while the structural changes are small and unlikely to be measurable by standard methods in structural biology, it has been shown in other systems that small changes introduced, for example, by point mutations can lead to significant changes in protein dynamics [52]. In one study, point mutations that were shown to change the specific volume of the protein cyclic AMP receptor protein led to more significant changes in protein compressibility and flexibility as measured by H/D exchange. The changes in specific volume for cAMP were on the order of 0.1%; we estimated changes in RARα from our simulations to be on the same order, around 0.13%. In addition, we also measured an increase in the number of cavities in the average structures of the phosphorylated receptor as measured by the program FPOCKET [53] (data not shown).
These fine structural alterations induced by phosphorylation affects, in turn, the internal dynamics of the receptor by altering the collective motions as shown by the quasi-harmonic analysis of our simulations. The small changes in conformation that occur upon phosphorylation can not be attributed to a specific quasi-harmonic mode, as is often the case for large-scale conformational change, but the change in dynamics of loop L8–9 is captured by changes in the ensemble of low frequency modes. A consequence of these changes, due to phosphorylation of S369, is an increase in the atomic fluctuations of loop L8–9, the cyclin H binding site.
In summary, the scenario proposed here for the allosteric communication pathway is that phosphorylation induces a local ordering of the structure in the region around the phosphorylation site, specifically in loop L9–10, that leads to the release of loop L8–9 via modulation of a the R347-D256 salt bridge, which then permits a greater conformational freedom of L8–9. The observed allosteric communication is subtle in that local modifications in side-chain orientation perturbs the conformational ensemble accessible to the RARα LBD, and the conformational redistribution, although not associated with major structural changes, modifies the intrinsic dynamics of the LBD and favors signal transduction. Evidence exists for other proteins in which allostery occurs not on the backbone level but rather by a rearrangement of side-chains [54], [55]. In the case of RARα, an orchestrated rearrangement of the salt bridges and ion-pairs at either end of H9 was observed, modulating specific structural support and triggering the allosteric communication within the receptor. The changing distribution of ion-pairs upon phosphorylation described in this work illustrates the shift of the conformational ensemble.
Currently, there is no structural information available for the cyclin H/RARα complex and the results presented here are a first step towards the characterization of the interaction between these important binding partners. In the future, we will follow an integrative approach toward the building of an RARα/cyclin H complex in order to better understand how the changes observed here can lead to an enhanced binding affinity. We will characterize the complex and identify important residues for interaction, whose mutations can alter the affinity and thus participate in the non-genomic regulation of the receptor.
The X-ray crystal structure of the heterodimeric complex RXR/RAR complexed to all-trans RA and a 13-mer peptide from the Nuclear Receptor Coactivator 2 protein was used (PDBid: 3A9E) [34]. The calculations presented here were limited to the RARα LBD monomer (Figure 1). All heavy atoms were present in the experimental coordinate files. Prior to hydrogen atom placement, the protonation states of all His residues at physiological pH (7.4) were obtained using two different methods of pKa calculation: an empirical method related to the protein structure, PROPKA[56] and the H++ server method [57] based on the Poisson-Boltzmann equation. Both methods yield the same results for protonation states. The final construction of the proton positions, enforcing the protonation states determined in the above continuum dielectric calculations, was done using the HBUILD facility [58] in the CHARMM program [59], [60]. A first energy minimization using a distance-dependent dielectric coefficient and an epsilon of 4, consisted of 100 steps using Steepest Descent method followed by 1000 steps of Adapted Basis Newton-Raphson minimization method, was realized in order to eliminate strong steric contacts before system solvation. The parameters for the phosphorylated form of serine, as well as for the retinoic acid ligand were the same as those previously used [26]. Both forms of RARα (phosphorylated and non-phosphorylated), in complex with the RA ligand and the co-activator peptide, were subject to the same protocol.
Explicit solvent molecular dynamics simulations of RARα were done using the NAMD program [61] and the all atom force field CHARMM27 [62] with CMAP corrections [63]. After the energy minimization described above, a cubic box of equilibrated TIP3P water molecules with a box length of approximately 75 Å per side was centered on the protein center of mass. Waters overlapping the protein complex were removed. These explicit solvent systems were neutralized by Na+/Cl− counterions, additional ion pairs were added to yield a final physiological ionic strength of approximately 0.15 M. Simulations were carried out under periodic boundary conditions and the long-range electrostatic interactions were treated with the Particle Mesh Ewald (PME) algorithm [64]. All bonds between hydrogens and heavy atoms were constrained using the SHAKE algorithm [65], and an atom-based switching function with a cutoff of 12 Å was applied to the van der Waals non-bonded interactions. An integration time step of 1 fs was used for all simulations.
A combined energy mimization-molecular dynamics protocol was used to prepare the solvated system for the molecular dynamics simulations. The water molecules were first relaxed around the fixed protein by 1000 steps of Conjugate Gradient (CG) energy minimization using a constant dielectric coefficient and an epsilon of 1, followed by heating to 600 K over 23 ps, 250 steps of CG, and finally heating over 25 ps to reach the temperature of 300 K. Next, the positional constraints were removed and the entire system was subject to 2000 steps of CG, heating to 300 K over 15 ps. Finally, a production run of 50 ns was simulated, the first 10 ns of the simulations were eliminated from the analysis to ensure a good convergence. This protocol was repeated three times for both the unphosphorylated (unp-RARα) and phosphorylated (p-RARα) form of the nuclear receptor leading to 150 ns total simulation time for each form.
The initial structure used in the simulations of the phosphorylated receptor was that of the WT receptor with a phosphorylated serine in position 369. To further ensure that the simulations of the phosphorylated LBD explore the new local energy basin thoroughly and that no systematic drift of the structure is present during the simulations, we assessed the convergence of the simulations using the ensemble-based approach developed by Lyman and Zuckermann [27]. The procedure can be described as follows: (1) a cut-off distance dc is defined for the calculations and a reference structure S1 is picked randomly from the trajectory, (2) S1 and all the structures less than dc from it are removed from the trajectory, constituting the set S1. These two steps are repeated until there are no conformations left in the trajectory. Next, the trajectory is divided in two groups, in our case according to the time intervals from 10 to 30 ns and from 30 to 50 ns. All the structures in the half-trajectories are clustered according to the set of reference structures, leading to a unique set of structures for a set of given reference structures.
In our case, we clustered each trajectory half using a Cα-RMSD of 1.4 Å as the cutoff distance. This ensured a reasonable total number of reference structures. Lone structures result from the absence of any other structure within the specified cutoff. Here, the criterion used as a measure of good convergence of the simulations was when a low number of lone structures was found in a given trajectory, as was similarly done in earlier work [28]. The convergence calculations were repeated three times for each of the six trajectories using different seeds for the random number generation. For the unphosphorylated and phosphorylated simulations, the number of lone structures found in each set is 2 and 3, respectively. Although this type of convergence evaluation cannot formally rule out the possibility of conformational changes occurring on a timescale longer than that of the simulations, it is a useful test to determine whether a representative ensemble of structures has been generated for both the non-phosphorylated and phosphorylated LDB of RARα with no systematic drifts in the simulations.
To characterize the low frequency, collective motions of RARα, as well as changes in these motions as a function of phosphorylation, a quasi-harmonic analysis (QHA) of the molecular dynamics trajectories was performed. The QHA was carried out over the final 40 ns of the simulation using the quasi-harmonic command in the VIBRAN module of the CHARMM program [62]; all the modes were calculated in this analysis (3xN atoms) corresponding to 12,285 modes for the phosphorylated form and 12,276 modes for the unphosphorylated one.
For each simulation, the root-mean-square coordinate difference (RMSD) was calculated, as well as the backbone atomic root mean square fluctuations (RMSf), which were averaged by residue. The RMSf were calculated from the molecular dynamic simulations over the time frames from 10 to 50 ns. These calculated fluctuations were compared to the atomic fluctuations calculated from experimental B-factors.(1)where t refers to a specific timeframe, is the reference position of atom i (average structure over the time considered), the position of atom i at the time t and T refers to the total number of timeframes used in the calculation of the average, which is related to the time interval for the averaging. Atomic fluctuations were compared to the experimental B factors using:(2)Cross-correlation coefficients Cij [66] between residues assess the nature of inter-residue motion, that is whether relative motions between the residues i and j are correlated or anti-correlated. Cross-correlation coefficients can be calculated from normal modes, from quasi-harmonic modes, as well as from molecular dynamic simulations following the equation:(3)where ri and rj are the displacements from the mean position of residues i and j, respectively. From the Cij correlation coefficients, which are organized as a matrix, a cross-correlation map was calculated using a color-coded 2D representation. In this representation, Cij = 1 identifies correlated motions and Cij = −1 anti-correlated motions. These values give us additional information concerning the global collective motions of RARα.
In this work, we calculated the cross-correlation coefficients directly from the molecular dynamics simulations of 50 ns. Each trajectory was evaluated in blocks of 100, 500 and 1000 ps. For each of these blocks, a mean structure was calculated and the Cij correlation coefficients were calculated for the backbone atoms. Correlation maps were obtained by averaging the Cij over all time interval blocks corresponding to the time interval. For the different time intervals, the results were essentially unchanged, so only the results from 500 ps interval are presented here.
From the multitude of LBD crystal structures, it is known that salt bridges play an important structural role. Here, in order to characterize salt bridge formation and stability during the molecular dynamics simulations, we used the “Salt Bridges” plugin for the VMD [67] program, which determines whether a salt bridge is formed. The salt bridge is considered formed if the distance between any oxygen atom of an acidic residue side-chain and the nitrogen atoms of a basic residue side-chain is within a distance of 3.2 Å. The distances were calculated over the last 40 ns of each simulation.
A more detailed structural analysis was done in order to assess any changes to the relative orientation of the secondary structural elements, in particular the helices present in the LBD. The relative orientation of two helices was expressed in terms of an orientation angle as defined in the Chothia-Levitt-Richardson algorithm [32] implemented in the CHARMM program. Axes for helices H4, H9 and H10 were determined based on the respective Cα atoms and the cylinder most closely approximating a helix on these atoms was calculated. From this, their relative orientations were determined over the course of the simulations.
A second analysis of helices, which quantified the bend of a helix as the radius of a circle needed to fit the α-carbons [68], was done; the smaller the radius, the larger the helix bend. The software SIMULAID [31], and specifically the TRAJELIX [30] tool, was used to calculate this radius.
Histograms of the distances and the angles obtained are plotted using the R project package [69]. After evaluating the normality of the angles and distances distributions using the Shapiro-Wilk test, we performed a Student's test to assess the statistical difference between the averages of values obtained from the non-phosphorylated and phosphorylated simulations. In supplementary material, we show the distributions of the helix radii from the individual simulations of the un-phosphorylated and phosphorylated LBD (Figure S3). The average radii for the individual simulations are 18.4, 17.8 and 18.4 Å for the unphosphorylated receptor and 18.5, 18.5 and 18.3 Å for the phosphorylated LBD (Figure S3). For reference, the radius calculated using the X-ray structure of helix H9 is 18.5 Å and the value for an ideal alpha-helix of the same sequence built by the PyMol program is 16.6 Å. Thus, in the crystal structure environment, helix H9 of the unphosphorylated receptor is more extended than that of an ideal helix. This is likely due to the alpha-helical sandwich organization of the ligand binding domain of nuclear receptors. One simulation of the unphosphorylated receptor showed a clear shift toward lower values, while the other two show a more subtle shift of the radius toward the ideal value.
|
10.1371/journal.pcbi.1004578 | Towards a Molecular Understanding of the Link between Imatinib Resistance and Kinase Conformational Dynamics | Due to its inhibition of the Abl kinase domain in the BCR-ABL fusion protein, imatinib is strikingly effective in the initial stage of chronic myeloid leukemia with more than 90% of the patients showing complete remission. However, as in the case of most targeted anti-cancer therapies, the emergence of drug resistance is a serious concern. Several drug-resistant mutations affecting the catalytic domain of Abl and other tyrosine kinases are now known. But, despite their importance and the adverse effect that they have on the prognosis of the cancer patients harboring them, the molecular mechanism of these mutations is still debated. Here by using long molecular dynamics simulations and large-scale free energy calculations complemented by in vitro mutagenesis and microcalorimetry experiments, we model the effect of several widespread drug-resistant mutations of Abl. By comparing the conformational free energy landscape of the mutants with those of the wild-type tyrosine kinases we clarify their mode of action. It involves significant and complex changes in the inactive-to-active dynamics and entropy/enthalpy balance of two functional elements: the activation-loop and the conserved DFG motif. What is more the T315I gatekeeper mutant has a significant impact on the binding mechanism itself and on the binding kinetics.
| Imatinib remains the most important and studied anti-cancer drug for cancer therapy in its new paradigm. Due to its inhibition of the Abl kinase domain, imatinib is strikingly effective in the initial stage of chronic myeloid leukemia with more than 90% of the patients showing a complete remission. However, the emergence of drug resistance is a serious concern. Here, we investigate the molecular mechanism of drug-resistant mutations which, despite the importance and the adverse effect on cancer patients prognosis, is still debated. Our extensive molecular simulations and free energy calculations are consistent with an allosteric effect of the single-point drug-resistance-causing mutations on the conformational dynamics. Two partially independent conformational changes play a role. Our findings might help the design of anti-cancer therapies to overcome drug resistance and be used to predict the clinical relevance of new drug-resistant mutants found by genetic screenings of tumor samples.
| The revolutionary discovery of the potent anticancer drug imatinib (Gleevec, 2001) [1] had a huge impact on cancer therapy. This drug has a striking efficacy in the early stages of chronic myeloid leukemia (CML), with 90% of patients showing remission [2, 3]. Imatinib targets the Abl tyrosine kinase (TK), constitutively active in CML due to a chromosomal translocation [4]. Unfortunately, most patients in an advanced stage of the disease suffer from relapse due to the onset of drug-resistance [5]. Even if, next-generation kinase inhibitors (KIs) are available, or in clinical trials [6], their efficacy might also be affected by drug resistance responses. Among different mechanisms, the development of resistance-inducing mutations is the most relevant in tyrosine kinases [6]. Mutations occur in highly conserved positions on the protein [7], frequently shared by several kinases [8], suggesting a conserved kinome-wide mechanism. Unfortunately, the molecular mechanism of mutation-mediated resistance are only partially understood. In the case of the widely studied “gatekeeper” mutant, found in several TKs (T315I in Abl) [9], three mechanisms have been proposed. The direct one involves the abrogation of a crucial hydrogen bond formed by imatinib. A second hypothesis posits that the observed shift towards the active form, which was reported in Abl and several other TK bearing the gatekeeper mutation, would allow the natural substrate ATP to outcompete the inhibitors. [10–13] Very recently, a third mechanism has been proposed for Abl T315I whereby the suppression of an induced fit effect involving the p-loop would be responsible for the decreased binding affinity of imatinib. [14] It is probable that the gate-keeper mutations have a combined effect on the binding of inhibitors, changing their binding mode and affecting at the same time the conformational changes [10, 11]. The importance of the conformational changes in the mode of action of drug-resistant mutations [15, 16] is also confirmed by the fact that many of them are far away from the binding site (Fig 1), and thus act allosterically by disfavoring the drug-binding conformation and favoring active form [8, 17–19]. The link between conformational changes and allosteric regulation in TKs is well established. For instance, in the case of Src (a close homologue of Abl) the gatekeeper mutation has been shown to allosterically affect remote regulatory motifs [20].
Indeed, TKs can exist in a dynamic equilibrium between multiple conformations [21–23], differing by the conformation of the “activation loop” (A-loop), of the conserved DFG motif and of the αC-helix (Fig 1). While the features of the active, catalytically-competent, state are shared among different TKs, and include an A-loop in an extended (“open”) conformation, inactive conformations can be multiple and highly diverse, all sharing an (at least partially) “closed” A-loop. Type II inhibitors, as imatinib, target a particular inactive conformation, known as “DFG-out”, where the aspartate is flipped, taking the place of the phenylalanine and pointing out of the ATP binding site [24] and opening an adjacent “allosteric” pocket. When both the DFG is in the ‘out’ conformation and the A-loop is not fully extended, the drug can enter the cavity and adopt a bridge position above the DFG, occupying both the catalytic site and the allosteric pocket [25]. The DFG-out conformation has been observed in many kinases but, despite a common binding mode, imatinib binds strongly only to some of them [26]. In contrast to the induced-fit mechanism proposed for Abl, free energy calculations performed with different and independent approaches on several TKs are consistent with a conformational selection mechanism [23, 27, 28]. Indeed, the different affinity to Abl and Src can be explained by the thermodynamic penalty to adopt the drug-binding “DFG-out” conformation [29]. Our simulations on Abl and Src also revealed a correlation between the flexibility of specific functionally-relevant structural elements and the DFG-out penalty [29].
To investigate the molecular mechanism of drug-resistant mutations, and whether or not changes in the conformational landscape, have a role in it, here we performed enhanced-sampling atomistic molecular dynamics (MD) simulations and free energy calculations on a pool of wild-type (WT) TKs for which imatinib has various strengths (IC50), and several Abl Imatinib-resistant mutants. In most drug resistant mutants we find extensive changes in the conformational free energy landscapes associated with two functionally-important conformational changes: the DFG-flip and the closed to open A-loop switch. The changes between the relative free energies of the active (DFG-in, extended A-loop) and inactive (DFG-out, closed A-loop) states are in the main due to entropic contributions arising from the fast (sub-μs) dynamics of proximal structural elements. It is thus not surprising that we find a correlation between the sub-μs flexibility of specific structural elements in the active state and drug resistance. Indeed, altering the sub-μs dynamics has an effect on the binding of Imatinib, as also shown by mutagenesis and calorimetry experiments. The gatekeeper mutant T315I is a notable exception, as it affects both the binding mechanism itself and the conformation of the A-loop equilibrium. To the best of our knowledge this is the most in-depth computational study addressing the molecular roots of resistant mutants of Abl.
The kinase structures were retrieved from the Protein Data Bank (PDB entries 2G1T, 2SRC, 1PKG, 3KMM, 2WGJ and 2ITW). Missing residues were added using the software Modeller [30], according to the respective Uniprot sequences. We used the Amber99SB*-ILDN [31, 32] force field, including backbone corrections by Hummer and Best [33], with explicit TIP3P [34] water molecules. The unbiased MD simulations were carried out with the ACEMD program [35] running on GPUs. The systems were minimized with 10000 steps of conjugated gradient and equilibrated in the isothermal-isobaric (NPT) ensemble for 10 ns, using a Berendsen barostat at 1 atm. The temperature was kept at 305 K by a Langevin thermostat. A 400 ns production run was carried out for all the systems in the canonical (NVT) ensemble. The runs for Abl, Src and Kit were extended to a total length of 1 μs in both the DFG-in and DFG-out conformations. The simulations of the five Abl mutants (G250E, E279K, H396P, E450K and T315I) were carried out with the same setup in both the DFG-in and DFG-out conformation, after mutating in-silico the respective residue in the Abl structure. To capture the sub-μs dynamics, the RMSF of the Cα were averaged on 30ns non-overlapping windows after discarding the first 100ns of each run. We also compared the diffusion in the space described by the first two principal component analysis vectors (see S9 Fig).
Since the conformational changes of TKs take place on time scales longer than those accessible by standard MD simulations, here we used a combination of enhanced sampling approaches. Parallel Tempering—Well Tempered Metadynamics [36, 37] in the well tempered ensemble (WTE) [38, 39] (PTmetaD) was chosen due to its proven ability to fully converge complex conformational free energy surfaces such as those relevant in kinases (including the DFG-flip). Indeed, the PTmetaD approach (both in the standard and WTE variants) has already successfully used to study the conformational dynamics and its associated free energy landscape in many kinases [12, 13, 29, 40, 41] and other flexible proteins [39, 42]. PTmetaD was performed using the software Gromacs 4.5 [43] and the PLUMED plugin [44], using an integration step of 2 fs. The particle mesh Ewald algorithm was used for electrostatic interactions. Temperature coupling was done with the V-rescale algorithm [45]. The WTE allowed the use of a reduced number of replicas compared to standard PTMetaD [12, 39]. An average exchange probability of 24% was obtained using 5 replicas in the temperature range 305–400 K. We used the same four collective variables (CVs) that were used to reconstruct the free energy surface (FES) associated with the DFG flip in Src and Abl [29].
They are shown in S6 Fig and defined in the following (SRC numbering): CV1 is the distance between the centre of mass of Asp404 (DFG motif) and Lys295. CV2 is the distance between the centre of mass of Phe405 and the Cβ of Ile293. CV3, is a function of 3 dihedral angles f(ϕ404, ψ405, ψ408) ranging from 3, when the three dihedral arguments correspond to the DFG-Asp-in position, to 0, when they are in the DFG-Asp-out conformation. CV4 is the distance between the centre of mass of residues Asn381, …, His384 and residues Ala408, …, Ile411 of the activation loop, known to form a β-sheet in the active conformation. The height of the Gaussians was set at 2.0 kJ/mol with a deposition rate of 1/2000 steps and a bias factor of 5. The Gaussian width used for the CVs was 0.1 for the dihedral similarity (CV3) and 0.3 Å for all the others.
A minimum of 400 ns of sampling per replica in the NVT ensemble were needed to reach full convergence of the free energy. In the case of G250E and T315I, as the convergence was slower than in the other cases, we performed more than 600 ns and 1200 ns of sampling per replica, respectively. The total sampling time amounted to more than 14 μs across all mutants. The free energy surface reconstruction was obtained from the PT-metaD-WTE by reweighting the fixed potential energy bias and using two independent approaches: integrating the bias or using the time independent estimator of Ref. [46]. The convergence of the free energy reconstruction was monitored by integrating the cumulative added bias as a function of time (50 ns intervals) and comparing the reconstruction to that obtained by the time-independent estimator (as shown in S11, S14–S18 Figs). Changes of all the CVs used were also monitored to guarantee that the system diffuses freely in the CV space and is able to visit all the basins several times (see S12 Fig).
The FES were also reprojected as a function of two other CVs describing the conformational change of the A-loop from open to closed. Two path collective variables [47] were built by using the open and closed crystallographic structures of Src (CV1) and Abl (CV2). The reweight was performed by using our python implementation of the approach of Ref. [46] (available on our homepage: https://www.ucl.ac.uk/chemistry/research/group_pages/prot_dynamics/. The entropic contribution to the DFG-in DFG-out free energy difference was computed by linear fitting of the free energy differences as a function of temperature. It has been shown that this approach is more accurate than competing ones. [48]. We also tested the robustness of the estimate with respect to the range of temperatures on both standard PT and WTE-PT (see S8 Fig).
The drug binding free energy surface was calculated using metadynamics with the same software described above. The General Amber Force Field (GAFF) was used for the ligand. The ligand charges were calculated at the HF level using a 631-G* basis set with the Gaussian03 [49] package. QM-level torsional scans with a step of 10 degrees were carried out for the ca-ca-n-c and ca-ca-c3-n3 dihedrals which appeared to have wrong torsional profiles. The profiles with the GAFF force field in vacuum were then fitted to the QM ones to refine the dihedral parameters. As customary in the case of ligand binding [50, 51], we performed a preliminary metadynamics run using sub-optimal geometrical CV to obtain an initial pathway for the setup of the path collective variables (PCV) [47]. We selected 23 frames from the lowest free energy path obtained in the preliminary run and optimized this initial guess using the methodology described by Branduardi et al. [47]. To take into account possible rearrangements, we included the Cα atoms of the A-loop and of the αC-helix in the definition of the PCVs. The 2 PCVs s and z were used to run a 300 ns metadynamics. The height of the Gaussians was set at 2.0 kJ/mol with a deposition rate of 1/2000 steps and a bias factor of 10. The Gaussian width used for the CVs was 0.1 for s and 0.003 for z. The free energy corresponding to the last leg of the unbinding mechanism (from the external binding pose to a fully solvated state) was computed again by using Well-tempered metadynamics following the approach of Ref. [52].
Src was expressed and purified following the procedure detailed in Ref. [53]. For the first two mutants, an auto-induction protocol was used to reach a good bacterial expression. Thermodynamic binding parameters for the association of imatinib to Src wild type and to the designed mutants were acquired using a VP-ITC microcalorimeter (MicroCal Inc.). The sample solution consisted of 2 ml 0.1 mM protein in MES buffer (pH 6.5) supplemented with 2% DMSO. The ligand solution consisted of 2 ml 1.6 mM imatinib in MES buffer (pH 6.5) with 2% DMSO. Both solutions were degassed for 5 minutes under vacuum. Titrations were conducted at 30°C, consisted of a first control injection of 1 μl followed by 37 injections of 8 μl over 16s, each spaced by an equilibration period of 480s. The samples were stirred at 266 rpm. Raw data were collected and corrected for ligand heats of dilution. A one site binding model was assumed and data were fit using MicroCal Origin software (version 7.0). All experiments were repeated in triplicate.
We ran MD simulations on wild-type (WT) Abl, five clinically-reported drug-resistant mutants of Abl and five WT tyrosine kinases (see Table 1) [29].
The aim is to assess if there is a correlation between the changes in sub-μs dynamics of the TK structure (both in WT proteins and drug-resistant mutants) and the binding affinity of imatinib. We probed the changes in sub-μs dynamics by comparing the root mean square fluctuations (RMSF) of the Cα averaged over 30ns-long time windows. Among the most common drug-resistant Abl mutants found in relapsing leukemia patients, we selected the gatekeeper mutant (T315I) and four others far from the binding site (allosteric) having a strong impact on imatinib binding (higher IC50 values, see Table 1) [16, 17, 54]. Apart from the widely discussed gatekeeper mutant (T315I), the mechanism of resistance of the others is completely unknown.
In addition to Abl and its mutants, we selected a pool of homologous tyrosine kinases that, while sharing the same imatinib binding mode, are inhibited by it with varying strength (see Table 1) [5, 55–59].
The RMSFs (Fig 2) show that all the Abl mutants and the imatinib weak-binders TKs (Src, Met and EGFR) are more rigid than WT Abl [29]. The regions most affected are the functionally-relevant regions of the N-lobe (p-loop, αC-helix and the β3-αC loop), the A-loop and, to a lesser extent, the αG-helix, whose flexibility profiles are more “Src-like” [29]. An analysis of the diffusion of the dynamics in the space defined by the first two principal component vectors of Abl is also consistent with the rigidification of the mutants (see S9 Fig).
The DFG-out states, with the notable exception of G250E and T315I, are even more rigid, indicating a probable entropic penalty with respect to the DFG-in states (see S7 Fig). The β3-αC loop, which is known to play a role in the allosteric activation of Abl by the SH2 domain [40, 60] is the most affected element.
The A-loop is more rigid in most mutants and in all the weak-binders group (Src, Met and EGFR), while it is comparatively more flexible in the strong imatinib binders (Abl, Kit and Lck). A-loop flexibility and IC50 are anti-correlated (Table 1). Finally, the αG-helix, an hotspot for allosteric regulation in TKs [20, 61], is very flexible in the weak-binding group and in E279K resistant mutant (see S2 Fig) and can experience unique conformational changes (see S2 Fig). The clinical drug-resistant mutants of Abl [16] are prevalently located in regions of the structure whose flexibility varies across weak and strong-binding TKs (see S1 Fig). This should not be surprising, as the observed flexibility changes affect regions that are either directly involved in the binding of type II inhibitors (the DFG-motif, A-loop [25] and p-loop. [62], see S1 Fig)) or allosterically connected to the former (αC-helix, αG-helix, β3-αC loop) [20, 40]. Thus, the change in the sub-μs dynamics of these regions could affect the entropic contribution to the relative stability of the inactive and active states, in turn decreasing the binding affinity of imatinib and other type II inhibitors that bind to an inactive state.
To understand whether inducing a change in the flexibility of relevant structural motifs is sufficient to increase the drug sensitivity, we designed a number of Src KD mutants and measured the changes in Kd.
We engineered mutations in four different locations, the P-loop (Q275), the β3-αC loop (P299) the A-loop (Q420) and the αG-helix (V461), to the equivalent residues in Abl and Lck. The residues to be substituted were chosen in the functionally-relevant regions showing a major variation in the sub-μs dynamics, between the weak and strong imatinib binders. We first performed MD simulations and compared the flexibility of various mutants to that of Src WT. Based on the RMSF we choose three candidates for the experiments. The first bears Q275A, P299Q and Q420E substitutions (Src to Lck), the second Q275G, P299E and Q420A (Src to Abl) and a third one with Q275A, P299Q and V461S (Src to Abl).
The RMSF shows an increased flexibility in the N-lobe (p-loop, β3-αC loop and αC helix) compared to Src WT (see S3 Fig). For the mutant Q275G/P299E/Q420A, also an increase in A-loop flexibility was identified. In the Q275G/P299E/V461S mutant was introduced a mutation on the αG-helix, that has provoked a suppression of the fluctuations in this area, resembling the αG-helix of Abl. The Kd appear to be anti-correlated with the RMSF of the A-loop (Table 2). As a cross-validation, when a chimeric protein with the entire αC-helix of Abl was produced, we observed a general suppression of the dynamics and an increased Kd of 14 μM. At difference with previous studies reporting similar Kd [26], the mutated residues are far from the binding site and do not interact with the drug.
If the correlation between changes in the sub-μs dynamics and imatinib affinity observed in the MD simulations and validated by experiments is due to entropic effects it must be reflected in the conformational free energy landscape (and in the entropic contribution to the relative stabilities of the DFG-in and -out states). As discussed above, Imatinib binds to an inactive state in which both the conformation of the DFG motif and that of the A-loop are important. In non-phosphorylated TKs the active-like “open” state in which the A-loop is extended, is marginally populated [12, 63, 64]. The DFG-out inactive state has a significant thermodynamic penalty and (according to the observed changes in the sub-μs dynamics) is in most cases entropically disfavored. In both inactive states the delicate balance between entropic and enthalpic contributions are crucial in defining their stability with respect to the open, DFG-in active state. This balance can be easily upset by drug-resistant mutations either directly or through the allosteric network. Thus a mutation affecting the entropy / enthalpy balance of the active and inactive states could easily shift the kinase towards active state, decreasing the binding affinity of type II inhibitors. Indeed, the stabilization of the active (extended A-loop) conformation has been experimentally observed for many TK gatekeeper mutants (including Abl T315I) [10, 11, 13, 65].
To further investigate this issue, we performed large-scale multiple-replica PT-MetaD simulations and computed the fully converged conformational free energy landscape associated with the DFG-flip of the five drug-resistant mutants. In Fig 3 we report the FES projected as a function of two collective variables (CVs) that distinguish the DFG-in from the DFG-out state [29], namely the distance between the DFG Asp404 and Lys295 and between the DFG Phe405 and Ile293 (Src numbering). As expected, for all mutants the global minimum of the FES corresponds to the DFG-in state (basin “IN” in Fig 3). The mutants also explore the DFG-out state (basin “OUT”), which in most cases corresponds to a well-defined metastable minimum. All mutations alter significantly the conformational free energy landscape. The free energy penalty of the DFG-out state increases going from Abl WT and G250E (ΔG of 4 ± 0.5 kcal/mol) to E450K (6 kcal/mol), H396P and E279K (7 ± 0.5 and 8 ± 0.5 kcal/mol). The gatekeeper T315I and (to a lesser extent) G250E are significant exceptions as the ΔG associated to their DFG-flip is very close to that of the WT. The increased thermodynamic penalty is due to a net entropic loss of the DFG-out state, as shown in Table 3 (the DFG-in/-out free energy differences as a function of the temperature are reported in S8 Fig). Again T315I and G250E are special cases as there is an entropic gain for the DFG-out state. The entropic contributions to the DFG flip are in excellent agreement with the observed changes in the sub-μs dynamics.
Indeed in H396P and E279K, where the A-loop is significantly more rigid, the flip has a larger free energy penalty. In the cases of G250E and the T315I gatekeeper mutant the penalty for the flip is comparable to that of the WT, and in agreement with the observed increased flexibility of the N-lobe and A-loop in the DFG-out state, there is an entropic gain associated with it. In both cases the DFG-out has a peculiar geometry where both Asp404 and Phe405 point outwards (S4 Fig) and an open active-like extended A-loop conformation is observed.
To quantify this aspect, we analyzed how the population of the “open” state is altered by the mutations. We re-projected the free energy surfaces (FES) as a function of path variables [47] describing the opening of the A-loop (Fig 4) with respect to Src (CV1) and Abl (CV2).
Abl, Src and E279K mainly populate semi-closed and closed conformations (basin B and C), while G250E, T315I and to a lesser extent E450K, H396P show a minimum in correspondence of the open A-loop (basin A, CV1 < 6). The A-loop in these and other mutants tends to form a second helix turn as in Src (S5 Fig). The residues involved are Asp391 to Ala395, which in the case of H396P are right before the mutation. The formation of this helix has been shown to stabilize the extended A-loop active state in TKs [10] and is also involved in the mechanism of oncogenic mutations [7, 12]. The stabilization of the active state is known to weaken the binding of ATP-competitive inhibitors to oncogenic mutants of EGFR [10, 12] and in the phosphorylated form of Abl [25, 26]. In the case of T315I the observed stabilization of the active state is in agreement with multiple experimental observations [10, 11, 13, 65].
Finally, we also observe a loss of structural integrity of the αC and αG helices in the DFG-out state of most mutants (S4 Fig) in agreement with previous proposals stating that conformational transitions in kinases are accompanied by local unfolding of secondary structural elements [66, 67].
Our results appear to be consistent with a significant impact of the resistant mutations on two different conformational changes: the DFG flip, disfavoring the drug-binding conformation (mainly mutants with rigidified A-loop and N-lobe), and the opening of the A-loop, possibly favoring the locking of the kinase in an active form. However, from the simulations on the unbound kinases, we cannot rule out an effect of the gatekeeper mutation on the proposed “induced-fit” mechanism [14]. Indeed, the predicted ΔG associated to the DFG-flip of T315I, which is very close to that of the WT, leaves two hypotheses open. Either the weaker binding of imatinib is due to the observed stabilization of the extended A-loop active state, or to the suppression of an induced fit effect, possibly acting on the p-loop conformation. To clarify this point and shed more light on the binding mechanism itself, we have computed the binding free energy of imatinib to Abl WT and Abl T315I along a physical association pathway. The use of such an approach, albeit it is significantly more expensive than an “end-point” free energy calculation (e.g. thermodynamic integration) has the advantage of reporting on free energy barriers and thus on the binding and unbinding kinetics. As expected, the crystallographic binding pose corresponds in both cases to the deepest minimum. Starting from that pose, the bidimensional (un-)binding free energy profiles show a substantial difference in imatinib’s mechanisms of binding to Abl WT and T315I (Fig 5). In the WT kinase the barrier to unbinding is lower (in agreement with the observed differences in unbinding kinetics) and there is one main exit path. It is also interesting to note that we find an “external binding pose” from which the inhibitor slides to its final crystallographic binding pose. In the external binding pose the DFG motif is already in the “out” position. In T315I we observe two different unbinding / binding pathways (B’ and E, Fig 5 and S10 Fig) somewhat similar to previous proposals. [68] In both of them the p-loop has a significant role. A pose similar to the external binding pose observed in the WT kinase is present (C’ in Fig 5) but it is slightly shifted towards the αC helix and more stable than in the WT. Thus the gain in free energy from the external binding pose to the final (crystallographic) pose is lower, in agreement with the observed increase of the IC50. When the unbinding free energy from C to a fully solvated state is accounted for (see S13 Fig), the overall free energy difference from the unbound fully solvated state to the crystallographic pose is around 13 kcal ± 2 (the larger convergence error is due to the algorithms used). When the free energy penalty of the DFG flip (≃ 4 kcal) is subtracted, we get 9 kcal ± 2, in agreement with the calorimetric and IC50 experimental data. It is thus clear that the effect of the T315I gatekeeper mutation is dual. It both affects the binding mechanism and stabilizes the active, extended A-loop conformation.
We have studied the interplay between sub-μs dynamics and conformational dynamics in TKs, and how these are influenced by drug-resistant mutations impacting Type-II inhibitors binding. Overall, our simulations show that drug-resistant mutations have a significant effect on both the sub-μs dynamics and the conformational free energy landscape. They affect the energy and the population of the DFG-out inactive state and of the open A-loop active-like state. Since type II inhibitors bind to the inactive kinase, a more accessible DFG-out state leads to a stronger binding of imatinib and other type II inhibitors. On the contrary, a more populated active-like state, in which the A-loop is elongated, increases the affinity towards ATP and disfavors type II inhibitors.
The selected drug-resistant mutants fall in two partially overlapping categories: those that have a significantly higher free energy penalty for the DFG-out state (E279K, H396P) and those (G250E, E450K, T315I) that populate the A-loop open active-like state. What is more the T315I gatekeeper mutant has a significant impact on the binding mechanism itself and on the binding kinetics.
The mutations affect the free energy differences associated to the conformational changes mainly by changing the sub-μs dynamics and consequently the entropy / enthalpy balance of the different states. Thus, relative short MD simulations, by revealing changes in the sub-μs dynamics, might be used to predict the impact of new mutations on Imatinib resistance.
The important role played by the entropic penalty of the DFG-out state must also be kept in mind when comparing low temperature and room temperature experiments. On the whole, we characterized the mechanism of action of several drug-resistant mutants of Abl. We have shown the link between fast and slow dynamics in these complex systems, providing a deeper understanding of the thermodynamics, kinetics and allosteric regulation of type II inhibitor binding TKs. In perspective, our results could help the design of fast and predictive computational approaches to predict the effect of yet unknown mutations of TKs.
|
10.1371/journal.pgen.1000721 | Evolutionary Convergence and Nitrogen Metabolism in Blattabacterium strain Bge, Primary Endosymbiont of the Cockroach Blattella germanica | Bacterial endosymbionts of insects play a central role in upgrading the diet of their hosts. In certain cases, such as aphids and tsetse flies, endosymbionts complement the metabolic capacity of hosts living on nutrient-deficient diets, while the bacteria harbored by omnivorous carpenter ants are involved in nitrogen recycling. In this study, we describe the genome sequence and inferred metabolism of Blattabacterium strain Bge, the primary Flavobacteria endosymbiont of the omnivorous German cockroach Blattella germanica. Through comparative genomics with other insect endosymbionts and free-living Flavobacteria we reveal that Blattabacterium strain Bge shares the same distribution of functional gene categories only with Blochmannia strains, the primary Gamma-Proteobacteria endosymbiont of carpenter ants. This is a remarkable example of evolutionary convergence during the symbiotic process, involving very distant phylogenetic bacterial taxa within hosts feeding on similar diets. Despite this similarity, different nitrogen economy strategies have emerged in each case. Both bacterial endosymbionts code for urease but display different metabolic functions: Blochmannia strains produce ammonia from dietary urea and then use it as a source of nitrogen, whereas Blattabacterium strain Bge codes for the complete urea cycle that, in combination with urease, produces ammonia as an end product. Not only does the cockroach endosymbiont play an essential role in nutrient supply to the host, but also in the catabolic use of amino acids and nitrogen excretion, as strongly suggested by the stoichiometric analysis of the inferred metabolic network. Here, we explain the metabolic reasons underlying the enigmatic return of cockroaches to the ancestral ammonotelic state.
| Bacterial endosymbionts from insects are subjected to a process of genome reduction from the moment they interact with their host, especially when the symbiosis is strict (the partners live together permanently) and the endosymbiont is maternally inherited. The type of genes that are retained correlates with specific metabolic host requirements. Here, we report the genome sequence of Blattabacterium strain Bge, the primary endosymbiont of the German cockroach B. germanica. Cockroaches are omnivorous insects and Blattabacterium cooperates with their metabolism, not only with essential nutrient metabolism but also through an efficient use of amino acids and the nitrogen excretion by the combination of a urea cycle and urease activity. The repertoires of functions that are maintained in Blattabacterium are similar to those already observed in Blochmannia spp., the primary endosymbiont of carpenter ants, also an omnivorous insect. This constitutes a nice example of evolutionary convergence of two endosymbionts belonging to very different bacterial phyla that have evolved a similar repertoire of functions according to the host. However, the current set of genes and, more importantly, those that were lost in the process of genome reduction in both endosymbiont lineages have also contributed to a different involvement of Blattabacterium and Blochmannia in nitrogen metabolism.
| In 1887, Blochmann first described symbiotic bacteria in the fatty tissue of blattids [1]. Later, Buchner [2] suggested that symbionts are involved in the decomposition of metabolic end-products from the insect host. A classic example is the cockroach. Several pioneering studies correlated the presence of cockroach endosymbionts with the metabolism of sulfate and amino acids [3],[4]. These endosymbionts were classified as a genus Blattabacterium [4], belonging to the class Flavobacteria in the phylum Bacteroidetes [5] and they live in specialized cells in the host’s abdominal fat body. Apart from cockroaches, they were only found in the primitive termite Mastotermes darwiniensis [6]. Phylogenetic analyses for the Blattabacterium-cockroach symbiosis supported the hypothesis of co-evolution between symbionts and hosts dating back to an ancient feature of more than 140 million years ago [7],[8]. Recently, genome sizes of the Blattabacterium symbionts of three cockroach species, B. germanica, Periplaneta americana, and Blatta orientalis were determined by pulsed field gel electrophoresis as approximately 650±15 kb [9]. Similarly, the authors demonstrated the sole presence of Blattabacterium strains in the fat body of those cockroach species by rRNA-targeting techniques. Phylogenetic analyses based on 16S rDNA also confirmed the affiliation of these endosymbionts to the class Flavobacteria [9]. Therefore, they are phylogenetically quite distinct from the majority of intensively studied insect endosymbionts that belong to the phylum Proteobacteria, mainly class Gamma-Proteobacteria. Recently, the highly reduced genome of “Candidatus Sulcia muelleri” (from now S. muelleri), an insect endosymbiont belonging to the class Flavobacteria has been also completely sequenced [10].
Primary endosymbionts such as Buchnera aphidicola or Wigglesworthia glossinidia complement the metabolic capacity of aphids or tsetse flies, respectively that feed on different nutrient-deficient diets [11]. There are also examples of metabolic complementation between two co-primary endosymbionts and their hosts. This is the case of S. muelleri, living in the sharpshooter Homalodisca vitripennis, which coexists with another Gamma-Proteobacteria endosymbiont, “Candidatus Baumannia cicadellinicola” (hereafter B. cicadellinicola). Both have developed a metabolic complementation to supply the host with the nutrients lacking in the limited xylem diet [12]. Another example is the case of B. aphidicola and “Candidatus Serratia symbiotica”, co-primary endosymbionts of the cedar aphid Cinara cedri that complement each other in the provision of essential nutrients [13],[14].
Omnivorous insects also harbor endosymbionts. It is the case, for example, of ants of the genus Camponotus and their primary endosymbionts, the Gamma-Proteobacteria “Candidatus Blochmannia floridanus” [15] and “Candidatus Blochmannia pennsylvanicus” [16] (from now B. floridanus and B. pennsylvanicus, respectively). In this association endosymbionts play an important role in nitrogen recycling [17].
Evolutionary convergences are generally considered as evidence of evolutionary adaptation. The study of endosymbiont evolution could provide examples of evolutionary convergences if we were able to show that very distant phylogenetic groups present similar functional repertoires and metabolic capabilities when they have evolved endosymbiosis in organisms having similar feeding behaviors. This may be the case of Blochmannia (a gamma-proteobacterium) and Blattabacterium (a flavobacterium) that have independently evolved in carpenter ants and cockroaches, two omnivorous insects.
In this study, we determine the genome sequence of an endosymbiotic flavobacterium, Blattabacterium strain Bge, primary endosymbiont of the German cockroach B. germanica. We have also inferred the metabolism to try to understand why cockroaches excrete ammonia, instead of being uricotelic like other terrestrial invertebrates, thus breaking the so-called “Needham's rule” [18], a question that has puzzled physiologists for a long time. Finally, we compare the inferred metabolism with the corresponding one of B. floridanus, the primary endosymbiont involved in nitrogen recycling in the carpenter ant Camponotus floridanus, an insect that has also a complex diet.
The general features of the genome of Blattabacterium strain Bge (CP001487) and their comparison with those of other selected bacteria are shown in Table 1. The size of the circular chromosome is 637 kb, and the G+C content is 27.1%. Only 23.4 kb are not-coding and they are distributed in 480 intergenic regions with an average length of 49 bp.
The overall coding density (96.3%) is the highest among insect endosymbionts known to date, indicating a highly compact genome. It is surprisingly higher than the most reduced insect endosymbiont “Candidatus Carsonella ruddii” (93.4%) [19]. In addition, 1.5 kb correspond to 139 overlapping regions with an average length of 11 bp. Of these overlaps, 94 (67.6%) are between genes on the same strand and 1 to 70 bp long. The other 45 cases (32.4%) involve two genes on opposite strands and are between 2 and 50 bp long. Of these, only in one case the two genes overlap with their start regions, whereas in the rest the overlap is in the terminal region of the genes. On the other hand, in “Ca Carsonella ruddii” 92% of the 126 overlaps are in tandem orientation, and thus on the same strand, and only five cases are between opposite strands, involving the termini and starts of the overlapping genes.
Assembly of the pyrosequencing data gave highly reliable contigs that combined with the data from Sanger sequencing resulting in a single contig, representing the entire genome. Probably due to the formation of a secondary structure, only a 33 bp stretch in an intergenic region upstream of the GroEL gene was not covered by pyrosequencing data but only by Sanger reads. Furthermore, annotation of the ORFs allowed a clear assignation of protein functions even in cases with only weak similarities with existing database entries. Not a single case of a possible host gene incorporated in the symbiont genome was found. Neither had we found coding sequences affiliated with Blattabacterium strain Bge outside the genome that could have been assigned to the host genome.
A total of 627 putative genes have been assigned (Figure S1), 586 of which are protein coding genes (CDS), 40 are RNA-specifying genes (34 tRNAs, 3 rRNAs located in a single operon, one tmRNA, and the RNA components of RNase P and the Signal recognition particle). The only pseudogene found corresponds to the protein component of RNase P. This gene coding for 118 amino acids is disrupted by an in-frame stop codon at amino acid position 53. The RNase P proteins of the free-living F. psychrophilum [20], Flavobacterium johnsoniae (http://genome.jgi-psf.org/flajo/flajo.info.html) and Gramella forsetii [21] contain a lysine residue at that position. Therefore, it is possible that the stop codon has been generated by an A–T point mutation in position 157 of the nucleotide sequence. Despite this mutation, the RNase P could be functional as it has been described that in vitro the RNA component can act enzymatically without a functional protein component [22]. Regarding the coding genes, it is interesting that, despite the compactness of the genome, there are eight gene duplicates: miaB, rodA, serC, lpdA, ppiC, argD, hemD, and uvrD.
No specific sequence of the origin of replication (oriC), such as dnaA boxes, was found in the genome [23]. Likewise dnaA, which codes for the protein that initiates replication by binding to such sequences, was also absent. Thus, the putative origin of replication was determined by GC skew analysis. The transitional region where the GC skew changes from negative to positive one (Figure S2) showed the position of replication origin to be in the gene dapB. It is worth mentioning that neither dnaA nor any of the genes normally adjacent to the replication site in bacteria (dnaN, hemE, gidA, hemE, and parA) have been found in this genome. However, Blattabacterium strain Bge, has retained recA, which could trigger replication by an alternative mechanism [15],[23].
We have inferred the metabolism of Blattabacterium strain Bge from its complete genome (Figure 1). Blattabacterium strain Bge possesses a limited capacity for nutrient uptake with only one ABC-type transport system, which may be specialized in fructose transport because this bacterium, contrary to the other sequenced endosymbionts, seems unable to use glucose as a nutrient. On the other hand, Blattabacterium strain Bge also codes for a glycerol uptake facilitator that enables transport of solutes, such as O2, CO2, NH3, glycerol, urea, and water. Therefore, it is possible that Blattabacterium strain Bge obtains carbon from glycerol as a supplementary source.
A sodium/drug antiporter, NorM, is also encoded by this genome. This system of efflux drug transport is common among enterobacteria but not among flavobacteria. In this group it is only known for the free-living bacteria F. psychrophilum and G. forsetii. This system can act as a multidrug transport as well as transporting oligosaccharidyl lipids and polysaccharide compounds.
There is an array of metal ion homeostasis transporters. In Blattabacterium strain Bge, there is a Trk transport system, a uniporter of the monovalent potassium cation, which requires a proton motive force and ATP in order to function. Only W. glossinidia has a similar transport system, although the encoded subunits differ: trkA and trkB in Blattabacterium; trkA and trkH in W. glossinidia. Other solutes are also transported by symport systems. Blattabacterium strain Bge is able to uptake glutamate and aspartate via a proton symporter. Both metabolites play an important role in the metabolism of this bacterium (see below). A phosphate/sodium symporter is also present.
Regarding electron transport, the encoded NADH-dehydrogenase (ndh) oxidizes NADH without proton translocation. There is also a succinate dehydrogenase (sdhABD). Electrons are transferred to a membrane-bound menaquinone (MQ) and a molybdenum-oxidoreductase, which accepts electrons from the MQ. With these elements, a proton motive force can be generated.
Blattabacterium strain Bge seems to be able to reduce intracellular sulfate to sulfite. A number of genes required for sulfur assimilation present in the genome, include those encoding for the two subunits of the sulfate adenylyltransferase, cysN and cysD, the adenosine phosphosulfate (APS) reductase cysH and the sulfite reductase proteins cysI,J. There is a missing step for the conversion of adenosine-5′-phosphosulfate (APS) into 3′-phospho adenosine-5′-phosphosulfate (PAPS). The generated sulfite is reduced to sulfide further on and assimilated into the sulfur-containing amino acids L-cysteine and L-methionine.
Blattabacterium strain Bge is able to synthesize its own cell wall and plasma membrane. However, it has lost the entire pathway required for lipopolysacharide (LPS) biosynthesis, like all sequenced Buchnera strains and B. cicadenillicola. This property explains why Blattabacterium strain Bge, similarly to these bacteria, are surrounded by a host vacuolar membrane, as shown in the electron-microscopy images (Figure S3).
Regarding amino acid biosynthesis, Blattabacterium strain Bge has the genes encoding biosynthetic enzymes needed to synthesize 10 essential (His, Trp, Phe, Leu, Ile, Val, Lys, Thr, Arg, and Met) and 7 nonessential (Gly, Tyr, Cys, Ser, Glu, Asp, and Ala) amino acids. Thus, the endosymbiont metabolism relies on Pro, Gln and Asn supplied by the host. Also present is the complete machinery to synthesize nucleotides, fatty acids, and the cofactors folic acid, lipoic acid, FAD, NAD, pyridoxine, and riboflavin. Finally, genes encoding enzymes for the synthesis of siroheme and menaquinone were also identified.
With respect to the metabolism of carbohydrates, genome analysis of Blattabacterium strain Bge indicates the presence of a truncated glycolysis pathway, since the genes that encode for phosphofructokinase (pfkA) and pyruvate kinase (pyk) are missing, as well as any sugar phosphorylating system except for fructose. Therefore, the pathway begins with fructose-1 phosphate and continues with the canonical enzymatic steps until the synthesis of phosphoenolpyruvate (PEP). Given the lack of pyruvate kinase genes, Blattabacterium strain Bge must produce pyruvate via the malic enzyme (NADP+-dependent malate dehydrogenase). Additionally, a complete non-oxidative pentose phosphate pathway is encoded in Blattabacterium strain Bge. As it is the case with Wigglesworthia, the glycolytic enzymes seem to be involved in gluconeogenesis rather than glycolysis complementing the non-oxidative pentose phosphate pathway [24].
In summary, although Blattabacterium strain Bge genome shows a strong reduction in gene number in all the functional categories, compared to their free-living relatives (see below), the core of essential functions and pathways is particularly well preserved.
The protein genes of Blattabacterium strain Bge were classified according to COG categories (Figure 2, Table 2). This distribution was compared with those of twelve selected bacteria: four Flavobacteria, which included three free-living species (F. psychrophilum, F. johnsoniae and G. forsetii) and the endosymbiont S. muelleri, and eight Proteobacteria endosymbionts, seven Gamma-Proteobacteria (B. floridanus, B. pennsylvanicus, B. cicadellinicola, B. aphidicola Aps, B. aphidicola Cce, S. glossinidius, and W. glossinidia) and one Alfa-Proteobacterium (Wolbachia sp. from Drosophila simulans). Taking the observed distribution of COG categories for Blattabacterium strain Bge as the expected distribution followed by each of the other bacteria examined, the hypothesis of equal distribution was rejected in all but the carpenter ant endosymbionts, Gamma-Proteobacteria B. floridanus and B. pennsylvanicus (Table 2). These results suggest that it is the hosts’ diet (cockroaches and carpenter ants are both omnivores) rather than phylogenetic closeness which is more strongly linked with the type of genes retained. This appears to be a clear case of functional evolutionary convergence in a broad sense. The proximity between the endosymbionts from omnivorous hosts was also confirmed when a dendrogram was created using the matrix of Kulczynski phenetic distances (Figure 3A). To locate the phylogenetic position of Blattabacterium strain Bge and compare it with the COG-based functional analysis, we used a phylogenetic tree based on 16S rDNA gene sequences (Figure 3B). As expected, the 16S rDNA gene analysis clearly separate Bacteroidetes from Proteobacteria phyla. Blattabacterium strain Bge clusters monophyletically within the Bacteroidetes phylum. The functional clustering differs clearly from the phylogenetic one.
A striking trait of this genome is the presence of a complete urea cycle (Figure 4). This feature has been described in few bacteria, and in only one member of the Bacteroidetes phylum, the cellulolytic soil bacterium Cytophaga hutchinsonii [25]. Moreover, to date, there are no reports of a complete urea cycle in an endosymbiont. The Blattabacterium strain Bge genome also retains the genes for the catalytic core of urease and we have detected urease activity in endosymbiont-enriched extracts of cockroach fat body (see below).
The genome of Blattabacterium strain Bge has two urease genes, ureAB and ureC, coding for the catalytic subunits, but lacks all genes for the accessory proteins supposedly required to produce an active enzyme in most bacteria. The ureAB fusion is not a novel situation since fused urease genes have also been described in other bacterial genomes, as it is the case of the free-living Flavobacterium C. hutchinsonii [25]. Regarding the lack of accessory genes, a similar situation is found in Bacillus subtilis cells expressing urease activity, which are able to grow with urea as sole nitrogen source [26]. To corroborate the presence of an active urease in Blattabacterium strain Bge, we performed an enzymatic assay on crude extracts of the endosymbiont-enriched fraction of the B. germanica fat body. Figure S4 shows a representative result for the urease assay. Although the detected specific activity under our experimental conditions was low (2 mU mg−1 protein; 1 U of urease corresponds to the formation of 1 µmol of ammonia per min), it was reproducible. Urease activity was also reproducibly detected in endosymbiont extracts from P. americana fat body (data not shown).
To further study the inferred metabolism in relation to nitrogen economy, we carried out a stoichiometric analysis of the reactions involved in the Krebs and urea cycles as well as other directly related reactions, such as urease, the malic enzyme, and their links to amino acid utilization (Figure 1 and Figure 4). Our results strongly suggest a key involvement of the endosymbionts in nitrogen metabolism and excretion in the German cockroach, in addition to their role in providing essential amino acids and coenzymes to the host. It is also worth mentioning that the endosymbiont metabolism relies on a supply of Gln from the host to cater for all its biosynthetic needs, including the urea cycle. Stoichiometric analysis shows that eleven out of fourteen elementary modes produce ammonia (Table S1). It follows that the metabolic network of Blattabacterium strain Bge could potentially use amino acids efficiently as energy and reducing-power sources, generating nitrogen waste in the form of ammonia (Figure 4).
Urease genes are also present in the Blochmannia endosymbiont genome [15] and the biochemical function of the urease in the carpenter ant endosymbionts is completely different from Blattabacterium. Studies of gene expression [27] and feeding experiments with 15N-labelled urea [17] in carpenter ants corroborate the role of urease in the transfer of nitrogen from dietary urea into the hemolymph amino acid pool. This requires an endosymbiont glutamine synthase to act as an essential step in nitrogen conservation during amino acid anabolism. Thus, although carpenter ants are omnivorous, their bacterial endosymbionts may upgrade their diet via an efficient nitrogen economy [17]. German cockroaches are also omnivorous; however, their endosymbionts lack genes encoding a glutamine synthase-like activity, a clear indication that the metabolic function of urease is not the same in the German cockroach and carpenter ant endosymbionts because generated ammonia cannot be re-assimilated. Therefore, although we have revealed a functional convergence between the cockroach and carpenter ant endosymbionts, which is probably due to their hosts’ omnivorous diets, they differ greatly from a metabolic viewpoint in detail, particularly in terms of nitrogen metabolism.
Traditionally, Blattabacterium endosymbionts have been postulated to be involved in the metabolism of uric acid in cockroaches. For instance, uric acid accumulation has been observed in aposymbiotic cockroaches [28],[29]. Metabolic use of nitrogen derived from fat body urates has been observed in B. germanica under certain conditions (e.g., in females on low-protein diet [30] and consumption of empty spermatophores by starved females [31]). Interestingly, fat body endosymbionts have been involved in uric acid degradation to CO2 in experiments with the wood cockroach Parcoblatta fulvescens injected with 14C-hypoxanthine [32]. Although involvement of gut microbiota cannot be completely ruled out, endosymbiont metabolism seemed more likely [33]. However, our results show that the endosymbiont genome does not code for any activity related to either the synthesis or the catabolism of urates. Therefore, and contrary to early reports based on putative cultured endosymbiotic bacteria [29], Blattabacterium strain Bge cannot participate in the metabolism of this nitrogen compound directly. Since uricase activity has been detected in the fat body of the cockroach [28],[34],[35], the host could contribute with uric-derived metabolites to the nitrogen economy of the endosymbiont which, in turn, would produce ammonia and carbon dioxide as final catabolic products.
The genome sequencing, metabolic inference, detection of a urease in the endosymbiont and the stoichiometric analysis of the central pathways of Blattabacterium strain Bge shed light on a whole series of hitherto unexplained classical physiological studies on ammonotelism in cockroaches [33],[36],[37]. Contrary to the speculation that some terrestrial invertebrates, like gastropods, annelids [36] and isopods [38], exploit ammonia excretion as “a return to the cheapest way” [38] to eliminate nitrogen, the case of the German cockroach and its bacterial endosymbionts indicates that this might not be the case. The evolution of terrestrial-living metazoa has favored the emergence of uricotely (e.g. the majority of insects) and ureotely (e.g. mammals) as water-saving strategies. Meanwhile, ammonotely, the ancestral character present in aquatic animals, has classically been considered maladaptive for terrestrial animals [18]. Symbiosis seems to play a role in this “return” of cockroaches to ammonotely by providing new enzymes required for this new nitrogen metabolism. Thus the metabolic capabilities acquired by symbiogenesis [39] afford to explore new ecological niches and dietary regimes.
B. germanica (Blattaria: Blattellidae) was reared in the Entomology laboratory (Cavanilles Institute for Biodiversity and Evolutionary Biology, University of Valencia). The cockroaches were kept in the laboratory at 25°C and fed with a mixture of dog food (2/3) and sucrose (1/3).
The bacterial endosymbionts were extracted from the fat body of B. germanica females. To do so, cockroaches were killed by a 15 to 20 min treatment with ethyl acetate and the bacterial cells were separated from the fat body as in [15]. An enriched fraction of bacteriocytes is then obtained that is used to extract total DNA following a CTAB (Cetyltrimethylammonium bromide) method.
The complete genome sequence of Blattabacterium strain Bge was obtained by a hybrid sequencing approach based on ABI 3730 sequencers and the pyrosequencing system (454; Life Science). To construct shotgun libraries, DNA fragments were generated by random mechanical shearing with a sonicator and posterior separation in a pulsed field gel electrophoresis. Insert sizes of 1–2 kb and 3–5 kb were purified and cloned into vector from XL-TOPO PCR cloning kit. Plasmid DNA was extracted using 96-well plates (Millipore) with the PerkinElmer MULTIPROBE II robot according to the manufacturers. DNA sequencing was performed on an ABI PRISM 3730 Genetic Analyzer (Applied Biosystems). In the initial random sequencing phase 9,227 sequences were obtained with 1.5-fold sequence coverage. Given the lack of joining between sequences, which may have been due to a large number of sequences from the host, a strict sequence analysis was performed with a specific bioinformatic tool called a Categorizer. It carries out a sequence classification method based on n-mers composition to correctly distinguish between Blattabacterium strain Bge and contaminating host sequences. This classifier was trained with sets of sequences identified from Blattabacterium strain Bge and the host. With these sets, we constructed a feature vector or model representing the 4- to 7-mers usage pattern of each organism. Then the n-mers composition of each read was compared with these generated models with a k-nearest neighbor clustering algorithm (KNN).
Although the number of retrieved host sequence reads was higher than the one of Blattabacterium strain Bge sequences for both sequencing approaches, the pyrosequencing approach generated enough sequences to close the gaps identified with the first method. The tool Gap4 from Staden Package [40] was used for the total assembly.
Fat body of B. germanica was isolated and prefixed in a 2.5% paraglutaraldehyde fixative mixture buffered with 0.1 M phosphate at pH 7.2 (PB). Prefixation was performed at 4°C for 24 h and then rinsed several times in PB. To avoid the loss of this dispersed tissue, the fat body was placed in agar (2%) forming small blocks. After prefixation, these blocks were fixed in 2% osmium tetroxide for one hour, dehydrated in graded alcohol and propylene oxide, stained in a saturated uranyl acetate solution 2% and embedded in araldite to form the definitive blocks. Thin sections (0.05 µm) were made using the Reichert-Jung ULTRACUT E (Leica) ultramicrotome, and then were stained with uranyl acetate and lead citrate. A JEOL-JEM 1010 electron microscope was used for the analysis.
The putative coding regions (CDSs) in the Blattabacterium strain Bge genome were identified with the GLIMMER3 program [41]. This program was first trained with closely related organism sequences from the Flavobacteria group. The coding sequence model obtained was then used by GLIMMER3 to scan the genome to predict potential coding regions by considering the putative existence of initiation codons and ORF length. Start and stop codons of each putative CDS were curated manually through visual inspection of the Blattabacterium strain Bge Genome Browser, a database specially designed for this symbiont. The putative coding proteins were initially analyzed by reciprocal best hits to determine orthology between genes of the Blattabacterium and those from bacteria belonging to the Flavobacteria group. According to these criteria, two genes are orthologs when a gene in one genome matches as the best hit with a gene in the other genome. Sequences that could not be assigned to any function in comparison with flavobacterial genomes were identified by searching a non-redundant protein database using BLASTX [42]. Final annotation was performed using BLASTP comparison with proteins in the NCBI and Pfam domains identified using the Sanger Centre Pfam search website. Non-coding RNAs were identified by different approaches. The tRNAscan program was used to predict tRNAs, as well as other small RNAs, like tmRNA, the RNA component of the RNase P. Signal Recognition Particle RNA were identified by programs like ARAGORN, BRUCE and SRPscan, as well as consulting the Rfam database [43]–[45].
In the absence of a diagnostic cluster of DnaA boxes, the origin of replication was identified by GC-skew calculated as (C−G)/(C+G) using the program OriginX [46]. The origin is located in the transitional region where the GC-skew changes from negative to positive values.
The ORFs orthologous to known genes in other species were catalogued based on non-redundant classification schemes, such as COG (Clusters of Orthologous Groups of Proteins). A metabolic network was reconstructed using the automatic annotator server from KAAS-KEEG [47]. According to our genome annotation, each pathway was examined checking the BRENDA [48] and EcoCyc databases [49].
Comparison between the COGs distribution of each species with that of the Blattabacterium strain Bge was carried using chi-square tests. To avoid the problem of multiple testing, we applied the Bonferroni correction so that for each individual test the significance level was 0.05/12 = 0.0042. That is, if the p-value is lower than 0.0042 then the hypothesis is rejected. The first p-value corresponds to the standard chi-square test (Chi2 p-value, df = 19). Due to the asymptotic nature of this test, expected frequencies should be higher than 5. However, we might expect some frequencies with low values. To correct this situation we also performed a Monte-Carlo version of this test (MC p-value). We performed 19,999 simulations under the null hypothesis, which together with the observed Chi2 statistics constituted a set of 20,000 values. The MC p-value cannot be lower than 1/20,000 = 5.00E-5.
The Kulczynski distance between species 1 and 2 is given by 1−0.5(Σjmin(y1j,y2j)/Σjy1j + Σjmin(y1j,y2j)/Σjy2j) where j (from 1 to 20) refers to the corresponding normalized COG categories (from 0 to 1). The dendrogram was derived from the corresponding distance matrix by applying a complete clustering method in which the distance between clusters A and B is given by the highest distance between any two species belonging to A and B, respectively. The statistical significance of the clusters of the dendrogram was evaluated by bootstrap analysis based on 100,000 replicates.
The sequences of 16S rDNA were aligned with MAFFT (v6.240) [50] program. The positions for the phylogenetic analysis were derived by Gblocks v0.91b [51]. In total, 1530 nucleotides were selected. The phylogenetic reconstruction was carried out by maximum likelihood using the PHYML program [52]. The best evolutionary model chosen by MODELTEST [53] was a GTR + Gamma (G) + I (Proportion invariant). Bootstrap values were based on 1000 replicates.
Abdominal fat bodies from dissected B. germanica adult females were homogenized with a Douce homogenizer adding a 50 mM HEPES buffer containing 1 mM EDTA, pH 7.5. The crude extract was centrifuged for 25 min at 6000 rpm at 4°C, and the pellet was resuspended with the homogenization buffer. The supernatant and a crude extract of cockroach heads (host tissue without endosymbionts) were used in control experiments. The resuspended pellet or bacteria-enriched fraction was treated with lysozyme (3.5 U mL−1) for 30 min at 4°C and sonicated for 5 sec. Urease activity was determined incubating the extract at 37°C with 110 mM urea. At different time intervals the reaction was stopped by adding 1 vol. 10% trichloroacetic acid and the produced ammonia was measured by the colorimetric Berthelot method [54] as described in [55]. The protein content was measured with a Nanodrop ND1000 equipment.
Stoichiometric analysis (using METATOOL) [56] was performed on the central pathways directly involved in amino acid catabolism, including the Krebs and urea cycles. Information about the reversibility of reactions was checked in the BRENDA database [48]. The input file for METATOOL is available upon request to the corresponding author.
The genome was sent to GenBank and has been assigned accession number CP001487.
|
10.1371/journal.pgen.1000873 | Deciphering Normal Blood Gene Expression Variation—The NOWAC Postgenome Study | There is growing evidence that gene expression profiling of peripheral blood cells is a valuable tool for assessing gene signatures related to exposure, drug-response, or disease. However, the true promise of this approach can not be estimated until the scientific community has robust baseline data describing variation in gene expression patterns in normal individuals. Using a large representative sample set of postmenopausal women (N = 286) in the Norwegian Women and Cancer (NOWAC) postgenome study, we investigated variability of whole blood gene expression in the general population. In particular, we examined changes in blood gene expression caused by technical variability, normal inter-individual differences, and exposure variables at proportions and levels relevant to real-life situations. We observe that the overall changes in gene expression are subtle, implying the need for careful analytic approaches of the data. In particular, technical variability may not be ignored and subsequent adjustments must be considered in any analysis. Many new candidate genes were identified that are differentially expressed according to inter-individual (i.e. fasting, BMI) and exposure (i.e. smoking) factors, thus establishing that these effects are mirrored in blood. By focusing on the biological implications instead of directly comparing gene lists from several related studies in the literature, our analytic approach was able to identify significant similarities and effects consistent across these reports. This establishes the feasibility of blood gene expression profiling, if they are predicated upon careful experimental design and analysis in order to minimize confounding signals, artifacts of sample preparation and processing, and inter-individual differences.
| As a major defence and transport system, blood cells are capable of adjusting gene expression in response to various clinical, biochemical, and pathological conditions. Here, we expand our understanding about the nature and extent of variation in gene expression from blood among healthy individuals. Using a large representative sample of postmenopausal women (N = 286) in the Norwegian Women and Cancer (NOWAC) postgenome study, we investigated blood gene expression changes due to normal inter-individuality (age, body mass index, fasting status), and exposure variables (smoking, hormone therapy, and medication use) at proportions and levels found in real life situations. Host genes were found to vary by inter-individual (i.e. fasting, BMI) and exposure (i.e. smoking) factors, and these gene lists may be used as a basis for further hypothesis development. Our study also establishes the feasibility of blood gene expression profiling for disease prediction, diagnosis, or prognosis, but underscores the necessity of care in study design and analysis to account for inter-individual differences and confounding signals.
| There is growing evidence that transcriptome analysis of peripheral blood cells is a valuable tool for determining signatures related to disease [1]–[5] and drug-response [6]. Differences in blood gene expression may also reflect the effects of a particular exposure, such as smoking [7], metal fumes [8], or ionizing radiation [9]. In our previous research, we studied gene expression profiles from whole blood related to hormone therapy (HT) use in postmenopausal women [10] and identified specific challenges raised by inter-individual variability when isolating signals associated with defined exposure levels. Although blood gene expression profiling promises molecular-level insight into disease mechanisms, there remains a lack of baseline data describing the nature and extent of variability in blood gene expression in the general population. Characterizations of this variation and the underlying factors that most influence gene expression amongst healthy individuals will play an important role in the feasibility, design and analysis of future blood-based studies investigating biomarkers for exposure, disease progression, diagnosis or prognosis [11].
Several studies [12]–[18] have reported that technical variables such as collection, transportation, storage of blood samples, RNA isolation method and choice of microarray platform, in addition to biological effects, can influence gene expression profiles. These technical factors associated with the processing and preparation of human blood and subsequent microarray hybridization represent significant challenges in the analysis of variability.
Furthermore, a few previous studies have used microarrays to analyze blood from healthy volunteers and found that inter-individual sample variation was associated with sex [18], age [13],[18], the time of day the sample was taken [18],[19], and the proportion of the different cell populations comprising the blood sample [13],[18],[20]. However to date, all such studies have focused on gene expression profiles generated from a small set of samples not representative of the general population using different blood cell subtypes. For several reasons including the small sample sizes, these studies have been restricted to the analysis of a small number of variables simultaneously, thus ignoring possible interaction and confounder effects.
Finally, an understanding of these causes of variability would represent a significant step forward in the identification and evaluation of the disease and disease risk biomarkers. Most if not all genes are involved in molecular pathways that provide mechanistic insight in response to exposure or disease development. Pathway depictions are usually simplified, ignoring interactions with other pathways, and we often have incomplete knowledge about the specific interplay of the many elements in almost any particular system.
Using a large representative sample set of postmenopausal women in the Norwegian Women and Cancer (NOWAC) postgenome study [21],[22] processed via a standardized blood collection procedure and via an experimentally validated microarray platform [23], we investigate here the baseline variability of whole blood gene expression profiles. This represents the first comprehensive cross-sectional analysis of blood gene expression changes related to multiple inter-individual and exposure variables, and opens the new research discipline of systems epidemiology [24]. In this setting, we investigated blood gene expression changes due to technical variability, normal inter-individuality, and exposure variables at proportions and levels relevant to real life situations, and establish that these effects are mirrored in the blood transcriptome.
Peripheral blood is an ideal surrogate tissue as it has the potential to reflect responses to changes in the immediate and distant environments by alterations of gene expression levels. Given the number of factors that influence gene regulation and expression, it is not surprising that often more than one strong signal is present in any given high-dimensional dataset. The external validity of NOWAC as a representative sample of the Norwegian female population has been verified in several methodological analyses and found to be acceptable [35]. Studies of the internal validity, including reliability, have been undertaken for dietary questions [36],[37], menopausal status, and use of HT [36],[38], whereas validation of variables measuring physical activity remain ongoing.
In addition to technical variability, substantial differences in gene expression profiles were identified between individuals with respect to exposure. Overall, the functional enrichment of significant single genes and gene set enrichment analyses show that high-throughput gene expression studies implicate similar (although not identical) underlying biology across several studies. Whereas age did not induce a large effect in blood gene expression for our cohort of postmenopausal women aged from 48 to 62 years, pathways and gene sets affected by smoking and, to a lesser extent both BMI and fasting, are numerous and interconnected. Some expression profiles associated with these variables may also be associated with other factors (e.g., lower levels of exercise, age). A host of new candidate genes for regulation by inter-individual (fasting, BMI) and exposure (smoking) factors were identified which could be used as a basis for hypothesis development.
Several processes associated with smoking were involved in cardiovascular regulation by G-coupled receptors (i.e. purinergic, adrenergic beta-1, urotensin II or thromboxan A2 receptors) or protein activity (i.e. thrombospondin type-1, fibronectin type-3). Consistent with previous observations that smoking reduces olfactory sensitivity in a dose- and time-dependent manner [39],[40], we find that smoking significantly impairs blood gene expression of olfactory receptors. We also observed that smokers have deregulated gene expressions of several P450 cytochromes which catalyse mono-oxygenase activity that can both toxify and detoxify carcinogenic compounds. As established in normal lung [41] and rats [42], smokers tend to have a small increase in NAD(P)H:(quinone-acceptor) oxidoreductase compared to non-smokers.
Two previous studies [7],[31] have examined the effects of cigarette smoking on leukocyte gene expression in circulation and both of the associated signatures had the most significant enrichment scores over all gene sets considered here. Inflammatory responses previously associated with smoking [7] were up-regulated in the blood expression of smokers in our dataset. Lending support that smoking has immune and inflammatory effects, specific blood cell gene signatures [13],[18] (i.e. increased monocytes and decreased red blood cell and natural killer cell signalling) were differentially expressed according to smoking status. This is consistent with previous observations showing that the total numbers of peripheral leukocytes differ by smoking status [43],[44]. Core genes up-regulated in non-smokers from the enriched hormone-related gene sets [10],[33] were predicted to be involved in neuroactive ligand-receptor interactions like prostaglandin receptors. Elevated prostaglandin E2 synthesis has been previously reported in smokers in comparison with non-smokers [45],[46]. The predicted gene network also reflects the effect of smoking on hormone levels with increased secretion of prolactin and glucagon [47]. Two pathways related to exercise [32] were also found up-regulated in non-smokers, which may simply be due to an underlying prevalence of active exercisers in non-smokers [48].
In our study, we found BMI class associated with blood gene expression changes involved in several immune processes including diabetes type I. It has been reported that several immune functions are dysregulated in obesity [49],[50] and both genetic and environmental factors such as obesity have been implicated as triggers in the pathogenesis of diabetes. The role of autoimmunity in the origins of type I diabetes is well-known, including a role in latent autoimmune diabetes in adults [51] and several observations suggest that autoimmunity may be part of type II diabetes [52]–[55]. Finally, two pathways related to exercise [32] were also up-regulated in women with normal BMI which may be due to a higher prevalence of physical exercise than in overweight/obese women.
Of all the variables considered, fasting was associated with the largest number of genes, but few genes were identified as core genes possibly due to the limited number of fasting women (N = 28) at the time of blood sampling. Selection of core genes aims to select a subset of true positives which work together (possibly in similar pathways) towards significance of the set. The significant core genes associated with fasting were generally involved in gene expression regulation and chromatin modification [56]–[58]. Much of our understanding of the effects of nutrition on chromatin structure has been gleaned from model organisms, especially S. cerevisiae, C. elegans, Drosophila, and mice [59]. In humans, two previous studies were unable to characterize acute effects of food intake in blood gene expression profiles [13],[18]. One putative 784-gene signature exists [34], however only 49 genes associated with fasting overlap with this signature. This may simply be due to chance.
Due to a significant interaction between HT and MED within our profiles, further analyses with a larger sample size are needed in order to investigate the different categories of medications, HT regimens and hormone levels, as well as their interactions in blood.
Differences between the genes identified and the interpretation of results in the various studies discussed here are likely to have resulted from technical differences in the array platforms used, the subset of blood cells analyzed, and the chosen analytical procedures. Several studies [12]–[18] examined how gene expression profiles of blood samples are affected by technical variables. Specific blood sample collection methods result in the isolation of different blood cell subpopulations. White blood cells have been defined as the most transcriptionally active of all cell types in blood and may give the most sensitive gene expression profiles in response to defined factors [60]. In large epidemiological studies, RNA stabilization is compulsory and PAXgene tubes have been found satisfactory to stabilize and enable RNA extraction from whole blood cells [61]. While high proportions of globin RNA could reduce sensitivity with respect to certain microarray platforms [60],[62],[63], we previously investigated two globin reduction protocols and determined that they were not beneficial when Applied Biosystems (AB) microarrays are used [23]. We found that RNA extraction and one variable related to RNA degradation (i.e. time between blood collection and freezing) had a significant global effect on blood gene expression profiles. In addition to normalization preprocessing, our results suggest that technical variability should not be ignored and possible adjustment for technical sources of variability should be considered in any analysis. Techniques such as surrogate variable analysis [64] may adjust for hidden sources of heterogeneity and large-scale dependence in gene expression studies [65]. As an example in our study, 25 significant surrogate variables were highly correlated to the strongest identified technical sources of noise, array lot number (canonical correlation r2 = 0.95), time between blood collection and freezing (canonical correlation r2 = 0.62) and RNA extraction (canonical correlation r2 = 0.43).
After adjustment for technical variability, our analysis demonstrates the ability to find significant similarities between studies by focusing on the biological implications of the gene sets from each individual study, rather than the specific single genes that met the criteria for significant differential expression in each individual study. They lend support to the idea that blood gene expression studies can indeed detect exposure-specific differences and that failure to consider this type of biological variation can result in the misidentification of genes when investigating predictive, diagnostic or prognostic signatures in blood.
In conclusion, this study extends the limited baseline information currently available that describes normal patterns of variation in blood gene expression. The data generated have been made freely available and should represent a useful resource for the design of future studies including power calculations. Our results confirm the feasibility of identifying signatures of inter-individual factors (e.g. fasting, BMI) and exposure factors (e.g. smoking) in blood-based gene expression profiles, and reinforces the need for proper study design, sample preparation, and technical analysis.
We have received approval from the Regional Committee for Medical Research Ethics for the collection and storing of questionnaire information and blood samples. The informed consent formula explicitly mentions that the blood samples can be used for gene expression analyses as well as large-scale genotyping.
All data are stored and handled according to the permission given by the Norwegian Data Inspectorate. The Directorate of Health and Social affairs (SHD) has given us an exemption from the confidentiality of information in national registers.
Before use of the biological material, a request has been sent to the regional ethical committee for Northern-Norway. Use of biological material requires permission according to laws pertaining to biotechnology and gene technology, both of which are administered by the SHD.
The women are part of the Norwegian Women and Cancer (NOWAC) study (http://uit.no/kk/NOWAC/) consisting of 172471 women who were 30 to 70 years of age at recruitment from 1991 to 2006 [22]. The NOWAC postgenome cohort study [21] consists of approximately 50,000 women born between 1943 and 1957, randomly drawn in groups of 500 from the NOWAC registers, who gave blood samples between 2003 and 2006 and filled in a two-page questionnaire. The two-page questionnaire included questions regarding menopausal status, weight, height; past week exposure to smoking, HT, oral contraceptives, other MED, omega-3 fatty acid, soy or other dietary supplements; and details concerning blood specimen collection (date, hour, posture). Women included in the present study received a blood collection kit and an accompanying two-page questionnaire by mail in April 2005. Among the group of 500 women, 444 (89%) returned both citrate and PAXgene blood RNA (PreAnalytiX GmbH, Hembrechtikon, Switzerland) tubes; 3.3% declined to participate, 0.7% had died or migrated and 7% did not respond. Samples were included in the study according to the following inclusion criteria: the donor was postmenopausal (99 donors excluded), blood was successfully collected in one PAXgene tube and in two plasma collection tubes (8 donors excluded), and the samples were frozen within 3 days from blood collection (9 donors excluded). Based on these criteria, 328 PAXgene blood samples were included for RNA extraction.
PAXgene blood RNA tubes were thawed at room temperature for 4 h. 500 µL of blood was removed and stored on −70°C for future use. Total RNA was isolated using the PAXgene Blood RNA Isolation Kit, according to the manufacturer's manual. RNA quantity and purity was assessed using the NanoDrop ND-1000 spectrophotometer (ThermoFisher Scientific, Wilmington, Delaware, USA). The absorbance ratio of 260 nm and 280 nm (A260/A280) was between 1.93 and 2.1 for all samples included for further analysis. The Experion automated electrophoresis system (BioRad, Hercules, CA, USA) and the RNA StdSens Analysis Kit was used to evaluate RNA integrity of a randomized 32% of the samples, according to the instruction manual. The electropherograms were inspected for clear ribosomal peaks. We were not able to analyze any numerical criteria corresponding to electrophoresis patterns, because this information was not available. Thirty nine samples were excluded from further analysis due to insufficient RNA purity, yield or integrity. RNA samples were kept at −70°C until further use.
After exclusion based on study design and RNA quality and quantity criteria, samples were analyzed using the Applied Biosystems (AB) expression array system (Foster City, Lousiana, USA). 500 ng total RNA was used for amplification by the NanoAmp RT-IVT labeling kit from AB for one round of amplification, in accordance with the manufacturer's manual. Briefly, the 1st strand of cDNA was synthesized by reverse transcription using the T7-oligo (dT) primer, followed by 2nd strand synthesis. The double-stranded cDNA was purified, and used as template for in vitro transcription (IVT). During IVT, digoxigenin (DIG)-labeled UTP was incorporated into the cRNA. The quantity and purity of the cRNA was measured on the NanoDrop ND-1000, and the cRNA was stored on −70°C until further use. 10 µg of DIG-labeled cRNA was fragmented and hybridized to AB Human Genome Survey Microarray V2.0, in accordance with the Chemiluminescence Detection Kit Protocol.
The AB Human Genome Survey Microarray V2.0 contains 277 control probes and 32,878 probes for the interrogation of 29,098 genes. AB Expression System software was used to extract signal intensities, signal to noise ratios (S/N) and flagging.
A total of 304 arrays including 15 technical replicates were analyzed. Data analysis was performed using R (http://cran.r-project.org), an open-source-interpreted computer language for statistical computation and graphics, and tools from the Bioconductor project (http://www.bioconductor.org), adapted to our needs. Using R, we set the expression intensity to “missing” for genes with flagging value >8191 (threshold recommended by the microarray manufacturer). For a set of technical replicate arrays from the same subject, we excluded the array with the least number of probes that had a S/N exceeding 3. Furthermore, arrays (N = 3) where less than 40% of the probes had a S/N≥3 were also removed from the analysis. Individual probes were not considered, if the S/N exceeded 3 in less than 50% of the samples. After sample and probe filtration, we proceeded with a log2 transformation, quantile normalization and imputation of missing values using 10-nearest neighbourhood method [66]. A total of 286 arrays and 16185 probes are analyzed. Microarray data have been deposited at Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/geo) accession number GSE15289.
The global ANCOVA [25] was carried out by comparison of linear models via the extra sum of squares principle to test for the univariate and multivariate association between global expression values and technical variables. All significant technical variables with a permuted p-value <0.001 identified in the ANCOVA multivariate analysis were included in the gene-wise linear model selection as random (array lot number, RNA extraction date) and fixed (time between blood collection and freezing) variables.
Forward-backward variable selection was used to select gene-wise model based on BIC. Linear mixed models were used to test the association of each gene with the significant technical and all biological variables. The z-score from the global test [26] was used to select core probes that most strongly explain the difference between groups setting a FDR [67] threshold which maximizes the discovery of true positives (weight = 2) versus false positives (weight = 1) associated with each variable. Gene set enrichment analysis was conducted using the global test [26], which offers the opportunity to compare two or more groups while taking into account the association between probe sets as well as their individual effects. When testing several gene sets curated from the literature, we adjusted for multiple testing using FDR [67]. Functional clustering and gene networks prediction were performed with the Database for Annotation, Visualization, and Integrated Discovery (DAVID) at http://david.abcc.ncifcrf.gov/ [27], and the Human Experimental/Functional Mapper (HEFalMp) [28] at http://function.princeton.edu/hefalmp, respectively.
|
10.1371/journal.pgen.1005110 | DNA Polymerase ζ-Dependent Lesion Bypass in Saccharomyces cerevisiae Is Accompanied by Error-Prone Copying of Long Stretches of Adjacent DNA | Translesion synthesis (TLS) helps cells to accomplish chromosomal replication in the presence of unrepaired DNA lesions. In eukaryotes, the bypass of most lesions involves a nucleotide insertion opposite the lesion by either a replicative or a specialized DNA polymerase, followed by extension of the resulting distorted primer terminus by DNA polymerase ζ (Polζ). The subsequent events leading to disengagement of the error-prone Polζ from the primer terminus and its replacement with an accurate replicative DNA polymerase remain largely unknown. As a first step toward understanding these events, we aimed to determine the length of DNA stretches synthesized in an error-prone manner during the Polζ-dependent lesion bypass. We developed new in vivo assays to identify the products of mutagenic TLS through a plasmid-borne tetrahydrofuran lesion and a UV-induced chromosomal lesion. We then surveyed the region downstream of the lesion site (in respect to the direction of TLS) for the presence of mutations indicative of an error-prone polymerase activity. The bypass of both lesions was associated with an approximately 300,000-fold increase in the mutation rate in the adjacent DNA segment, in comparison to the mutation rate during normal replication. The hypermutated tract extended 200 bp from the lesion in the plasmid-based assay and as far as 1 kb from the lesion in the chromosome-based assay. The mutation rate in this region was similar to the rate of errors produced by purified Polζ during copying of undamaged DNA in vitro. Further, no mutations downstream of the lesion were observed in rare TLS products recovered from Polζ-deficient cells. This led us to conclude that error-prone Polζ synthesis continues for several hundred nucleotides after the lesion bypass is completed. These results provide insight into the late steps of TLS and show that error-prone TLS tracts span a substantially larger region than previously appreciated.
| Genomic instability is associated with multiple genetic diseases. Endogenous and exogenous DNA-damaging factors constitute a major source of genomic instability. Mutations occur when DNA lesions are bypassed by specialized translesion synthesis (TLS) DNA polymerases that are less accurate than the normal replicative polymerases. The discovery of the remarkable infidelity of the TLS enzymes at the turn of the century immediately suggested that their contribution to replication must be tightly restricted to sites of DNA damage to avoid excessive mutagenesis. The actual extent of error-prone synthesis that accompanies TLS in vivo has never been estimated. We describe a novel genetic approach to measure the length of DNA synthesized by TLS polymerases upon their recruitment to sites of DNA damage. We show that stretches of error-prone synthesis associated with the bypass of a single damaged nucleotide span at least 200 and sometimes up to 1,000 nucleotide-long segments, resulting in more than a 300,000-fold increase in mutagenesis in the surrounding region. We speculate that processive synthesis of long DNA stretches by error-prone polymerases could contribute to clustered mutagenesis, a phenomenon that allows for rapid genome changes without significant loss of fitness and plays an important role in tumorigenesis, the immune response and adaptation.
| Genomic stability is continuously threatened by endogenous and exogenous DNA-damaging factors. Unrepaired lesions stall the replication machinery, because the highly selective active sites of replicative DNA polymerases cannot accept abnormally shaped nucleotides [1, 2]. The bypass of replication impediments is facilitated by specialized translesion synthesis (TLS) polymerases. In humans, these include the Y-family enzymes Polη, Polι, Polκ, and Rev1, and the B-family enzyme Polζ. The yeast Saccharomyces cerevisiae has homologs of Polη, Rev1 and Polζ [3]. A more open active site allows the TLS polymerases to accommodate a variety of DNA lesions and catalyze synthesis on damaged templates [4, 5]. While important for tolerating DNA damage, TLS is a highly mutagenic process because of the miscoding potential of the damaged nucleotides and the inherently lower fidelity of the specialized polymerases. It is a major source of environmentally induced mutations and a significant contributor to spontaneous mutagenesis. Particularly, yeast and mammalian cells lacking Polζ or its partner Rev1 are completely deficient in mutagenesis induced by most DNA-damaging agents [3, 6, 7].
TLS is a two-step process that involves insertion of a nucleotide opposite the lesion and extension of the resulting distorted primer terminus. The insertion can be performed by a replicative polymerase or one of the TLS polymerases, depending on the type of lesion. With the exception of cis-syn cyclobutane pyrimidine dimers, where Polη is also able to facilitate the extension step, extension of the aberrant primer terminus is usually catalyzed by Polζ [8–10]. This unique role of Polζ underlies its absolute requirement for damage-induced mutagenesis: the extension of primer terminus containing a wrong nucleotide is essential for conversion of the misincorporation into a permanent change in the DNA. While the molecular details of the insertion and extension steps have been studied extensively for a variety of lesions, the subsequent processes leading to the replacement of the TLS polymerases with accurate replicative enzymes are poorly understood. The low fidelity of the TLS polymerases suggests that their contribution to DNA synthesis past the lesion site must be tightly regulated. The coordination of TLS with the ongoing DNA replication has been considered in the context of the following two models [6, 11]. In the model referred to as “polymerase switching” or “TLS at the fork”, the stalled replicative polymerase hands the primer terminus over to a TLS polymerase to allow for the bypass to occur and then returns to resume high-fidelity replication. In this scenario, TLS operates at the fork and provides for continuous synthesis of the daughter strand. In contrast, in the “gap-filling” model, stalling of the replicative polymerase is followed by a quick re-priming of replication downstream of the blocking lesion, which leaves a gap between the site of the lesion and the site of the restart. TLS polymerases then act post-replicatively to facilitate filling of these gaps. The possibility of direct switching from a replicative to a TLS polymerase and back is in excellent agreement with the biochemical properties of these enzymes and their behavior on templates containing a single replication-blocking lesion in vitro [12, 13]. A bulk of evidence, however, suggests that TLS in vivo might predominantly occur as postreplicative gap filling (discussed thoroughly in [6, 11, 14]). Perhaps the strongest arguments are the lack of a noticeable decrease in the rate of fork progression in TLS-deficient mutants [15–17] and the accumulation of discontinuities in the newly synthesized DNA observed in multiple studies of UV-irradiated E. coli, yeast and mammalian cells [17–22]. The possibility of TLS polymerases acting in a gap-filling mode is illustrated by the participation of human Polκ in DNA synthesis during nucleotide excision repair (NER) [23] and by a recent study suggesting that Polζ is required for filling of lesion-containing, NER-generated gaps in non-dividing yeast cells [24]. Electron microscopy showed that gaps of up to 1,000 nucleotides are left behind the forks in replicating UV-treated yeast [17]. Nonetheless, all of the existing evidence for each of the TLS models is indirect. Whether TLS polymerases actually operate in these gaps, which undoubtedly exist, and what proportion of TLS events occur at the fork rather than postreplicatively, remains to be established.
The mode of TLS could be an important factor that determines the extent of synthesis performed by the error-prone TLS polymerases. In the case of direct polymerase exchange at the fork, it would seem rational for the replicative polymerase to return to the primer terminus as soon as its activity is no longer impeded by the lesion. In vitro, the eukaryotic replicative polymerases Polδ and Polε can resume processive DNA synthesis once a TLS polymerase elongates the primer by two to five nucleotides past the lesion [13]. It has been argued that the stretches of TLS synthesis should be just sufficiently long to prevent degradation of the TLS product by the proofreading activity of a replicative polymerase and not much longer to limit the accumulation of mutations in the downstream region [12]. Replicative DNA polymerases can sense distortions in the duplex DNA within 3–6 base pairs from the primer terminus [12, 13, 25]. Therefore, extending the error-prone synthesis further than a few nucleotides past the lesion would provide little benefit for the efficiency of TLS while delaying the fork progression and increasing the mutation load. In contrast, in the gap-filling model of TLS, the progression of replication does not require switching from a TLS to a replicative polymerase once the lesion is bypassed. A replicative polymerase may not be readily available outside of the replication fork, and the TLS polymerases could be well positioned to fill a large portion of the gap or even close it completely. A possibility of error-prone TLS proceeding well beyond the lesion and generating additional “untargeted”, or “hitchhiking”, mutations in the adjacent region has been considered previously [26, 27] but was never thoroughly investigated. In eukaryotes, the involvement of Polζ in the extension step during the bypass of most lesions makes this polymerase a likely candidate for continuing the synthesis past the lesion. The fidelity of purified Polζ in vitro was estimated to be 5.6 x 10-4 mutations per nucleotide synthesized [28]. While this is dramatically lower than the fidelity of the replicative enzymes, Polζ is still the most accurate among the eukaryotic TLS polymerases. It could be the best choice for synthesis of extended DNA stretches if it were to be done by a TLS polymerase.
In this study, we aimed to determine how much DNA is synthesized in an error-prone manner after Polζ completes the bypass of a plasmid-borne tetrahydrofuran lesion (THF) or a UV-induced chromosomal lesion in vivo. We reasoned that if Polζ-dependent TLS is accompanied by extensive copying of the adjacent undamaged DNA, these regions should exhibit an increased frequency of mutation. We developed new genetic assays that allowed us to identify and isolate the products of Polζ-dependent TLS occurring in vivo and examine the region surrounding the lesion site for the presence of mutations characteristic of a TLS polymerase activity. The results argue that Polζ copies several hundred nucleotides of DNA past the lesion after completing the bypass, which leads to a 300,000-fold increase in the mutation rate in this region. This demonstrates that TLS in eukaryotic cells is associated with mutagenesis not only at the lesion site, but in the extended adjacent area as well. In addition, these findings provide support for the previously hypothetical role of Polζ in filling the daughter strand gaps formed opposite lesions in replicating DNA.
This study aimed to determine whether Polζ-dependent TLS in vivo is associated with low-fidelity DNA synthesis past the lesion, and if so, define the size of DNA region copied in an error-prone manner. We first developed a system for the analysis of TLS through THF, an abasic site analog. Mutagenic bypass of both natural and artificial abasic sites in yeast requires Polζ for the extension step [29, 30] and is, thus, a good model of Polζ-dependent TLS. We have constructed a double-stranded plasmid containing a THF lesion at a specific position in the URA3 gene, a yeast replication origin ARS4, and the LEU2 gene for selection of cells containing the plasmid (Fig. 1A). Replication of this plasmid was studied in apn1Δ apn2Δ strains lacking the two yeast apurinic/apyrimidinic endonucleases to prevent removal of the lesion by the base excision repair system. Replication of the damaged strand in the yeast cells could be accomplished through several pathways, including TLS, error-free bypass utilizing an undamaged homologous sequence as a template, or a recombination-dependent mechanism. To distinguish the products of Polζ-dependent TLS from other events, we took advantage of the earlier observations that TLS through THF predominantly involves an A or C incorporation across from the lesion [31–33]. The THF was designed to replace a C in the wild-type URA3 sequence, such that an A or C insertion opposite it would produce a Ura- phenotype. Accurate bypass utilizing an undamaged template or repair of the lesion prior to replication (presumably infrequent in the apn1Δ apn2Δ strains but possible, Ref. [34]) would produce a DNA strand with a G at this position and a Ura+ phenotype (Fig. 1B). TLS events, thus, result in the formation of half-sectored colonies, where the Ura- and Ura+ halves result from the copying of the THF-containing and the complementary undamaged strands, respectively (Fig. 1C). Non-TLS events result in the formation of Ura+ colonies. According to the previous reports, a small percentage of TLS cases involve a T or G insertion opposite the THF (e.g. 1% and 8% of all TLS events, respectively, during replication of a double-stranded THF-containing plasmid; Ref. [31]). The T insertion results in a Ura- phenotype in our system and, therefore, is identified as a TLS event. The G insertion preserves the wild-type URA3 sequence and cannot be distinguished from non-TLS events. Accordingly, TLS products containing a G across from the THF position were not included in the analysis described below. The half-sectored colony phenotype was observed in ~1% of the transformants with the THF-containing plasmids (a total of 394 sectored colonies among ~40,000 transformants analyzed). While this is consistent with the previous observations that the great majority of THF bypass events occurs through mechanisms other than TLS [31], non-sectored colonies in our system could also result from replication fork uncoupling and copying of the undamaged strand [17]. To detect TLS-associated mutations, we isolated the plasmids from the Ura- part of the half-sectored colonies and analyzed a 1.7-kb region beginning at the original lesion site and extending in the direction of TLS, as well as a 550-bp region upstream of the lesion site in respect to the direction of TLS, by Sanger sequencing. The sequence corresponding to the oligonucleotide used to construct the THF-containing plasmid (20 nt before and 26 nt after the lesion) was excluded from the analysis, because mutations in this region could result from errors during the in vitro synthesis of the oligonucleotide. All of the plasmids analyzed (a total of 394) contained a base substitution at the THF position, confirming the mutagenic TLS event. The distance from the lesion site at which additional mutations were found served as an estimate of the length of the error-prone TLS tracts in this system.
Because of the accumulation of random damage in the ssDNA used for the plasmid construction, replication of the THF-containing plasmid is expected to result in a high rate of mutations in the vector sequence not related to the THF bypass. To evaluate the contribution of these “background” mutations to the overall mutagenesis in the 1.7-kb region downstream of the THF site, we designed a control plasmid constructed via the same procedure as the THF-containing plasmid but with no deliberately introduced lesion. To be able to analyze the progeny of the same strand that is replicated via error-prone TLS in the THF bypass assay, we engineered the two strands of the control plasmid to confer different phenotypes. One strand contains the wild-type URA3 sequence, while the other contains a three-nucleotide substitution (the ura3-103,104 mutation) at the position equivalent to that of the THF lesion, resulting in a three-base mismatch in the plasmid (Fig. 1D). Approximately 20% of transformants with this plasmid produced half-sectored colonies consistent with the replication of the plasmid and segregation of the URA3 and ura3-103,104 alleles into the daughter cells. The proportion of half-sectored colonies only mildly increased in mismatch repair (MMR)-deficient msh2Δ strains (from 21% to 33% on average), indicating that the “bubble”-type mismatch is inefficiently corrected by MMR. The non-sectored colonies could possibly result from the presence of more than one plasmid copy in some cells, which could preclude segregation of the Ura- phenotype, loss of a fraction of daughter plasmids and/or repair of the heteroduplex by an unknown mechanism distinct from MMR. We did not attempt to distinguish between these possibilities and used only half-sectored colonies for the analysis of control replication products. All of the plasmids isolated from the Ura- part of the half-sectored colonies contained the ura3-103,104 allele sequence at the site of the of the original mismatch, as expected from accurate copying of the Ura- strand. DNA sequence analysis of the surrounding region provided an estimate of the frequency of mutations associated with our method of plasmid construction.
A total of 394 THF bypass products and 456 products of the control plasmid replication were analyzed by DNA sequencing. As expected, the majority of TLS events resulted in an A (243/394; 62%) or C (80/394; 20%) incorporation opposite the lesion. T incorporation occurred in 18% of all cases (71/394). A total of 18 mutations were found in the downstream region at distances between 34 and 1529 nucleotides from the THF (Fig. 2A; Table 1). These “hitchhiking” mutations were noticeably concentrated within an approximately 220-nucleotide segment immediately adjacent to the lesion. Although 11 mutations were found among the 456 control plasmids downstream of the lesion site, their distribution was significantly different from that in the TLS products. Mutations in the control plasmids were randomly distributed throughout the sequenced region, with none of the 11 mutations occurring within the first 220 nucleotides, in contrast to ~40% in the THF bypass products (p = 0.0045, Fisher’s exact test). The rate of mutation in the 220-nucleotide region next to the THF site constituted 8.1 x 10-5 per nucleotide (Table 2). This exceeds the genome-wide mutation rate in yeast by approximately 300,000-fold and is close to the rate of errors reported for copying of undamaged DNA by purified Polζ in vitro (5.6 x 10-4, [28]). The frequency of mutations immediately upstream of the lesion site did not differ from that in the control plasmids (Fig. 2A), consistent with the idea that the patch of increased mutagenesis resulted from error-prone DNA synthesis initiated at the lesion site. The frequency of mutations downstream of the lesion was reduced to the background level as the distance from the lesion exceeded 220 nucleotides. Accordingly, the types of mutations in these distant regions were very similar to those in the control plasmids (predominantly C→T transitions and -1 deletions). In contrast, only one C→T transition and no -1 frameshifts were found in the 220-bp region adjacent to the lesion site (Fig. 2B, Table 1). We, therefore, concluded that mutations present in the TLS products outside of the 220-bp region must have resulted from damage of ssDNA during the plasmid construction, and only those observed within the 220-bp region are indicative of error-prone DNA synthesis associated with the THF bypass. We also sequenced the 220-bp region in 47 THF bypass products and 57 control plasmids recovered from msh2Δ strains to determine whether errors made during TLS-associated synthesis are corrected by MMR. We observed no increase in the frequency of untargeted mutations over that in MMR-proficient strains (Table 2), indicating that MMR does not operate in TLS tracts. This is in agreement with a previous report that MMR does not efficiently correct errors made by Polζ [35]. Taken together, these observations suggest that the error-prone synthesis typically continues for approximately 200 nucleotides after the THF bypass is completed. Because of the high level of background mutagenesis in this system, we cannot exclude a possibility that a minor fraction of TLS events could involve more extensive low-fidelity synthesis.
As described previously, sensitivity of the plasmid TLS assay is limited by the high background likely resulting from spontaneous damage during the plasmid construction. To overcome this limitation and to ascertain that the extended stretches of error-prone DNA synthesis is not a peculiar feature of the plasmid assay, we next developed an approach to study mutagenesis associated with the Polζ-dependent bypass of a chromosomal DNA lesion. Because creating a unique site-specific abasic site in a yeast chromosome is technically challenging, we chose to use a lesion induced by UV irradiation at a specific dipyrimidine sequence in the chromosomal URA3 gene. Cis-syn cyclobutane pyrimidine dimers and (6–4) photoproducts are major types of DNA lesion induced by UV irradiation and can be generated at any of the four pyrimidine doublets, TT, CT, TC, and CC [36]. Error-prone bypass of UV lesions in vivo, like that of abasic sites, occurs via a Polζ/Rev1-dependent pathway [37–39]. We introduced a single nucleotide substitution at position 764 of the chromosomal URA3 gene (the ura3-G764A mutation) that leads to a Ura- phenotype and creates a dipyrimidine sequence (TC), a possible site for UV lesion formation (Fig. 3). The ura3-G764A strains can revert to the Ura+ phenotype via several base substitutions at the 3' C or 5' T of the dinucleotide (Fig. 3A). The occurrence of either substitution upon UV irradiation of yeast cells indicates that the lesion formation and the mutagenic TLS have taken place at this site. The frequency of the ura3-G764A reversion increased in a dose-dependent manner in wild-type strains, but not in rev3Δ mutants lacking Polζ (Fig. 4). This indicated that UV irradiation readily induces lesions that are bypassed via Polζ-dependent synthesis to produce the Ura+ revertants. DNA sequence analysis of the URA3 locus of 165 independent revertants obtained after irradiation with 60 J/m2 UV light showed that all of the revertants contained base substitutions at 3' C, 5' T or both positions of the dipyrimidine at the site of ura3-G764A mutation (S1 Table). Thus, the system is highly efficient in the identification of products of mutagenic TLS through a site-specific chromosomal lesion. To determine the extent of error-prone synthesis associated with the bypass of this lesion, total DNA was isolated from the 165 Ura+ revertants, and 2.5-kb regions upstream and downstream from the reversion site in respect to the direction of TLS were amplified by PCR and sequenced (Fig. 3B).
Similar to the THF bypass, TLS through the chromosomal UV lesion was frequently accompanied by additional mutations in the 2.5-kb region downstream of the lesion site. In 12 cases of reversion at the ura3-G764A site, a mutation at the 5’ or 3’ nucleotide of the TC doublet was associated with a mutation at the next G presumably not involved in the lesion formation (+1 position; S1 Table). Because the accuracy of nucleotide incorporation at this position is likely severely affected by the distorted DNA structure at the damaged site, we did not include these mutations in the calculation of mutation rate in the adjacent region. A total of 15 additional mutations were found in the 165 TLS products at distances between 16 to 2155 nucleotides downstream of the reversion site (Fig. 5; Table 3). As in the case of the THF bypass, the untargeted mutations were noticeably concentrated in the region immediately adjacent to the lesion, but the hypermutated stretch now extended as far as 1000 nucleotides from the presumed lesion position (Fig. 5). The mutation rate in this 1000-bp segment constituted 6.7 x 10-5 per nucleotide (Table 2), which is similar to the level of mutagenesis we observed in the products of the THF bypass. To confirm that these mutations are associated with TLS at the site of ura3-G764A mutation, we repeated the experiment but selected for cells that underwent mutation at the CAN1 locus rather than the Ura+ reversion. The CAN1 is located ~83 kb away from the ura3-G764A allele in the same chromosome V. We detected only one DNA sequence change in the 1-kb region downstream of the ura3-G764A mutation site among 161 independent UV-induced Canr mutant (Fig. 5). This indicated that the untargeted mutations in the Ura+ revertants were, indeed, related to the TLS through a nearby lesion and did not simply reflect a high level of mutagenesis in the genome of irradiated cells. The frequency of mutations upstream of the ura3-G764A site was low and similar to that in the Canr controls, consistent with the idea that the error-prone synthesis initiated at the site of ura3-G764A mutation. The rare mutations we observed in the Canr controls and in the Ura+ revertants outside the hypermutated 1-kb region likely resulted from additional UV-induced lesions. According to previous estimates, the dose of 60 J/m2 used in our experiments is expected to induce approximately one lesion per 1–2 kb [40, 41]. The frequency of mutation we observed in the Canr controls (which is indicative of the average frequency of mutagenic lesions in the genome of cells undergoing UV-induced mutation) is about 100-fold lower. This is consistent with the notion that the majority of lesions are repaired in NER-proficient cells, and only a fraction of the remaining lesions are mutagenic.
Polζ is required for the extension step during the mutagenic bypass of THF and UV-induced lesions [29, 33, 38, 39] and could, thus, be responsible for the error-prone synthesis in the downstream region. To exclude the possibility that the observed high rate of mutations resulted from synthesis by the other low-fidelity yeast polymerase, Polη, we sequenced the 1-kb region downstream of the ura3-G764A mutation site in 165 Ura+ revertants obtained in the rad30Δ background (S2 Table) after exposure to the same dose of UV irradiation (60 J/m2). The reversion frequency was not significantly affected by the inactivation of RAD30. Five mutations were found at distances of 505–1026 nucleotides from the reversion site, which corresponds to a mutation rate of 3 x 10-5. This is similar to what we observed in the RAD30+ strain and argues against a major role of Polη in mutagenesis downstream of the lesion site. In contrast, Polζ appeared to be required for the generation of untargeted mutations. Although very little induced mutagenesis could be seen in rev3Δ strains lacking Polζ (Fig. 4), the frequency of Ura+ revertants at the dose of 60 J/m2 exceeded the spontaneous reversion frequency approximately three-fold, so the rare revertants resulting from Polζ-independent bypass could be recovered. Sequencing of the 1-kb region downstream of the ura3-G764A mutation site in 231 Ura+ revertant obtained in the rev3Δ background (S3 Table) detected no untargeted mutations. This indicated that the long stretches of hypermutation downstream of the lesion are specifically associated with Polζ-dependent TLS and strongly implicate Polζ in the generation of these mutations. Similar to the THF bypass experiments, sequencing of the 1-kb region from UV-induced Ura+ revertants obtained in the msh2Δ background showed that the frequency of untargeted mutations was not increased in the absence of MMR (Table 2), confirming that MMR does not correct errors in TLS tracts.
Previous studies of the Polζ-dependent TLS focused primarily on events at the lesion site [8, 29, 31, 42–47]. These studies have established Polζ as a key player in the extension of distorted primer termini resulting from nucleotide insertion opposite lesions by other polymerases. This function underlies the renowned requirement of Polζ for mutagenesis at the damage location. In the present work, we used novel genetic assays where the products of TLS can be identified phenotypically to demonstrate that the bypass of the THF and the UV-induced lesions is associated with a dramatic increase in mutagenesis in the adjacent region. The following arguments suggest that this mutagenesis likely results from continuous synthesis of long DNA stretches by Polζ. First, the essential role of Polζ in the extension step of the bypass puts this polymerase in a perfect position to carry on synthesis beyond the lesion. Second, the mutation rate in the region downstream of the lesion (~10-4 per bp) is comparable to the rate of errors observed during Polζ-dependent copying of undamaged DNA in vitro (5.6 x 10-4 per bp; Ref. [28]). Third, the untargeted mutations disappear in Polζ-deficient rev3Δ strains. Fourth, the level of untargeted mutagenesis remained high in rad30 mutants lacking Polη, the only other TLS polymerase in yeast capable of copying complex templates. We can also exclude a reduction in the fidelity of replicative polymerases as the cause of TLS-associated mutagenesis, because this would be expected to result in a genome-wide elevation of the mutation rate, which is not observed.
The rate and specificity of mutation downstream of the lesions argues that it results from the copying of adjacent undamaged DNA by Polζ and not from its recruitment to sites of secondary lesions in single-stranded regions formed after the replication arrest. The frequency of mutation triggered by the formation of ssDNA alone, without additional mutagenic treatment, is at least an order of magnitude lower than that observed in our experiments [48]. The ssDNA-mediated mutagenesis is also characterized by the abundance of C→T changes in the exposed strand (~43% of all base substitutions; Ref. [48]), which are completely absent in the 200-bp segment adjacent to the THF site in our study (Table 1). While the UV lesion bypass products contained C→T transitions (Table 3), the mutation rate downstream of the lesion site still greatly exceeded that expected from spontaneous damage in persistent ssDNA stretches (Table 2 and Ref. [48]). Therefore, we conclude that the distribution of mutations in the vicinity of the site-specific lesions (Figs. 2A and 5) directly reflects the extent of continuous synthesis by Polζ. Based on this, we estimate that, in vivo, Polζ copies at least 200 and sometimes up to 1,000 nucleotides of undamaged template upon completing the lesion bypass. This nicely parallels previous electron microscopy studies showing that the single-stranded gaps left behind the replication forks in UV-irradiated yeast are typically smaller than 400 nucleotides, but longer gaps (more than 1000 nucleotides) could be seen in a fraction of replication intermediates [17]. If TLS, as it is currently viewed, occurs predominantly in these gaps, Polζ must be filling a substantial portion of the gaps. Further studies would be needed to determine whether Polζ is solely responsible for the gap filling or if it is later replaced by a replicative polymerase such as Pol δ. The TLS assays we developed can be used as a tool to address this question if mapping of TLS stretches is complemented by accurate measuring of ssDNA regions accumulating next to the lesions.
We observed a notable difference in the distribution of mutations in the THF versus the UV lesion bypass assay, which might reflect a previously unappreciated aspect of TLS regulation. Mutations associated with the THF bypass concentrated within 220 nucleotides from the lesion (Fig. 2; Table 1). In contrast, mutations that accompanied the UV lesion bypass were distributed nearly randomly across the adjacent 1-kb region, becoming less frequent only beyond that point (Fig. 5). This indicates that the bypass of the chromosomal UV lesion is associated with longer stretches of error-prone synthesis. It seems likely that the extent of Polζ-dependent synthesis may be limited by the length of the single-stranded region remaining after re-priming of replication downstream of the lesion. This length could vary depending on the lesion position in the leading or lagging strand template. Stalled lagging strand synthesis is likely restarted with the initiation of the next Okazaki fragment, so the size of single-stranded gaps on the lagging strand would not exceed the size of Okazaki fragments (140–175 nucleotides according to the recent estimates, Refs. [49, 50]). However, replication restart on the leading strand requires additional regulatory mechanisms, and the re-priming might occur at a greater distance from the lesion. Although the direction of replication through the URA3 gene in our THF bypass assay is unknown, we believe the lesion is likely to be encountered by the lagging strand machinery in the majority of cases. The THF lesion is located at comparable distances from the centromere-proximal and centromere-distal sides of the replication origin ARS4 in the pRS315-URA3 OR2 plasmid (Fig. 1A). Because of the inhibitory effect of the repetitive centromeric sequences on the fork progression [51], the lesion-containing region is likely to be first approached by the fork coming from the centromere-distal side of ARS4. This would put the THF lesion in the lagging strand template. In contrast, the UV-induced lesion in chromosome V is likely located in the leading strand template. The end of the URA3 gene in a genetically unmanipulated chromosome V corresponds to the beginning of the replication termination zone (the region between 117 and 123 kb in Fig. 3B; Refs. [52, 53]). The LEU2 insertion (Fig. 3B) presumably moves the termination zone away from the URA3, placing the site of lesion formation in the leading strand template at a distance of at least 1 kb from the termination zone. The lesion position in the opposite DNA strands in the THF and the UV lesion bypass assays could potentially explain the differences in the length of DNA fragments synthesized in an error-prone manner. While this explanation at present remains hypothetical, the assays we described here could easily be adapted in the future to explore the role of replication fork dynamics and asymmetry in regulating the extent of Polζ-dependent synthesis.
The findings described here bring a new twist to understanding the consequences of DNA damage. The extent of error-prone synthesis downstream of the lesions is such that, regardless of whether the lesion occurs in a functionally important position, the probability of inactivating a nearby gene is extremely high. We estimate that, with ~7% of TLS tracts containing an additional mutation beyond the lesion site (this study), and ~1/3 of all base substitutions and the majority of frameshifts in coding regions affecting the gene function [54], a TLS tract spanning a coding region will destroy the gene in ~4% of cases. While targeted mutations are undoubtedly the primary cause of damage-induced mutagenesis, given the high proportion of essential genes (e. g. 1/3 of all genes in yeast), the untargeted mutations contribute to DNA damage sensitivity and pose an upper limit to the absolute number of unrepaired lesions that can be tolerated by a cell. Extended tracks of error prone synthesis can also lead to the accumulation of multiple mutations in a localized area without causing hypermutability across the genome. The localized hypermutability provides a mechanism for rapid genome changes while minimally affecting fitness and is believed to play an important role in several biological processes, including tumorigenesis, immune response and adaptation (discussed in [55–60]). While a major cause of clustered mutations was suggested to be the enzymatic deamination of cytosines in ssDNA, processive synthesis of long stretches of DNA by error-prone polymerases could conceivably contribute to this phenomenon as well. The two processes, in fact, are interrelated: the excision of uracil resulting from cytosine deamination by uracil DNA glycosylases produces abasic sites [61]. Subsequently, the deaminase-induced mutagenesis is, in part, mediated by Polζ-dependent TLS [57]. The long stretches of error-prone synthesis are also likely relevant to other situations where mutagenic processes promote adaptation, evolution or human disease and where the role of clustered mutations is yet to be established. For example, Polζ/Rev1-dependent TLS is believed to be responsible for the acquired drug resistance and the development of secondary tumors in patients undergoing chemotherapy with DNA-damaging agents [62–65]. The ability of TLS enzymes to generate multiple mutations in extended stretches of DNA likely accelerates the emergence of chemoresistance. Future molecular characterization of therapy-resistant tumors could help clarify the role of the TLS-associated localized hypermutability in tumor evolution.
The haploid Saccharomyces cerevisiae strains PS1001/PS1002 (MATα ade5 lys2-Tn5-13 trp1-289 his7-2 leu2-3,112 ura3Δ apn1Δ::loxP apn2Δ::loxP) and OK29/30 (MATα ade5-1 lys2::InsEA14 trp1–289 his7-2 leu2-3,112 ura3-G764A-LEU2) were used in the THF and UV lesion bypass assays, respectively. PS1001 and PS1002 are two independent isolates of the same genotype derived from CG379Δ [66, 67] by disruption of the APN1 and APN2 genes by the loxP-LEU2-loxP and loxP-kanMX-loxP cassettes, respectively. The cassettes were PCR-amplified from pUG73 [68] and pUG6 [69], and the disruption was followed by the Cre/loxP-mediated marker removal [69]. OK29 and OK30 are two independent isolates of the same genotype engineered as follows using E134 (same as OK29/30, but ura3-52; [70]) as the starting material. First, a Ura+ derivative of E134 (named E134+) was obtained by M. R. Northam in our laboratory by transformation with a PCR fragment containing the wild-type URA3 gene. The ura3-G764A mutation was created by site-directed mutagenesis in a yeast integrative vector containing the URA3 and LEU2 genes cloned into pUC18 [71], yielding pUC18-ura3-G764A-or1. OK29 and OK30 were then constructed by replacing the wild-type chromosomal URA3 gene of E134+ with the ura3-G764A-LEU2 cassette amplified by PCR from pUC18-ura3-G764A-or1. The primers for amplification had 20 bp of homology to pUC18 regions flanking the cassette at the 3’ end and 45 bp of homology to the chromosome V sequences upstream and downstream of the URA3 gene at the 5’ end. The LEU2 insertion next to the ura3-G764A in OK29/30 does not affect the function of the URA3 gene or the Ura- phenotype conferred by the mutation. The rev3Δ, rad30Δ and msh2Δ mutants of OK29/30 and msh2Δ mutants of PS1001/1002 were constructed by transformation with PCR-generated DNA fragments carrying the kanMX cassette flanked by short sequence homology to REV3, RAD30 or MSH2.
The pRS315-URA3 OR2 plasmid [72] containing the URA3 gene cloned into the HindIII site of pRS315 [73] was kindly provided by Dr. Youri Pavlov (University of Nebraska Medical Center, Omaha, U.S.A). In addition to the URA3, it carries the LEU2 selectable marker, the yeast autonomous replicative sequence ARS4, a yeast centromere sequence, and the f1 phage origin of replication. The single-stranded DNA (ssDNA) form of pRS315-URA3 OR2 contains the transcribed URA3 strand. Escherichia coli F’ strain DH12S (Invitrogen) and M13KO7 helper phage (New England Biolabs) were used for isolation of the pRS315-URA3 OR2 ssDNA. The E. coli strains XL10-Blue and MC1061 (Invitrogen) were used for plasmid rescue from yeast cells and for propagation of plasmid DNA.
The single-stranded pRS315-URA3 OR2 phagemid was purified as described in [74] with some modifications. The DH12S strain transformed with the phagemid was grown in LB medium to an optical density at 600 nm of 0.05 and then infected with M13KO7 at a final concentration of 1 x 108 pfu/ml. On the following day, the bacteriophage particles were precipitated from the culture supernatant by stirring in 4% polyethelene glycol—0.5 M NaCl at 4°C for 1 h and subsequent centrifugation at 4,000 x g for 30 min. The pellets were washed with 10 mM Tris-HCl and resuspended in 10 mM Tris-HCl pH 8.0. The cell debris was then removed by centrifugation at 60,000 x g for 15 min at 4°C. To pellet the phage particles, a subsequent overnight centrifugation was performed under the same conditions. The pelleted bacteriophage particles were resuspended in 10 mM TE buffer. To remove residual fragments of bacterial DNA or RNA that could anneal to the pRS315-URA3 OR2 ssDNA, the bacteriophage particles were treated with 120 U/ml T4 DNA polymerase (New England Biolabs) and 5 μg/ml RNAse A (USB) in NEB2 buffer (New England Biolabs) at 37°C for 2 h. The treatment was done in the absence of dNTPs to utilize the 3’-exonuclease rather than the DNA polymerase activity of T4 DNA polymerase. The reaction was stopped by incubation with 5 μg/ml Proteinase K (Sigma-Aldrich) at 55°C for 30 min. The pRS315-URA3 OR2 ssDNA was then purified from pre-cooled bacteriophage particles by three sequential extractions with phenol, two extractions with phenol-chloroform, and one extraction with chloroform, followed by ethanol precipitation. Samples were shaken gently to prevent shearing of the DNA. Purified ssDNA was stored in 10 mM TE buffer at -80°C.
The double-stranded plasmid containing a site-specific THF lesion and a control undamaged plasmid were constructed by annealing oligonucleotides 5’-AGGTTACGATTGGTTGATTATGACACXCGGTGTGGGTTTAGATGACA-3’ (Oligos etc), where “X” is THF, and 5’-AGGTTACGATTGGTTGATTATGACACGGCGTGTGGGTTTAGATGACA-3’ (IDT), respectively, to the pRS315-URA3 OR2 ssDNA and synthesizing the remainder of the second strand by T7 DNA polymerase. The oligonucleotides are complementary to the URA3 nucleotides 579–625. The control oligonucleotide contains three bases (underlined) that do not match the wild-type URA3 sequence and produce a triple CCG → GGC substitution (the ura3-103,104 allele) resulting in a Ura- phenotype. The oligonucleotides were PAGE-purified and annealed to the pRS315-URA3 OR2 ssDNA by incubating a two-fold molar excess of the oligonucleotide with the 400 ng of ssDNA at 72°C for 2 min in T4 DNA ligase buffer (New England Biolabs) and then cooling slowly to room temperature. The whole volume of the annealing mix was then incubated with 10 U of T7 DNA polymerase, 200 μM dNTPs, 4 mM ATP and 10 U of T4 DNA ligase (New England Biolabs) in T7 DNA polymerase buffer at 37°C for 1.5 h. The reactions were then treated with Proteinase K (Invitrogen) at 37°C for 20 min. The covalently closed double-stranded plasmids were isolated from 0.8% agarose gel by centrifugation through premade Sephadex-10 columns (Pharmacia Fine Chemicals) as previously described [75].
Double-stranded THF-containing and control plasmids were introduced into the yeast cells by polyethylene-glycol-mediated transformation [76]. The strains were grown at 30°C in rich liquid YPDAU medium [77] prior to transformation. Transformants were selected on synthetic complete medium without leucine (SC—leu), and three-day-old colonies were replica-plated on synthetic complete medium without leucine and uracil (SC −leu—ura) to score half-sectored phenotype. Total yeast DNA was purified from the Ura- part of the half-sectored colonies using the MasterPure Yeast DNA Purification Kit (Epicentre). To isolate plasmids from the total DNA samples, 5–7 μl of each sample was used for transformation of the XL10-Blue or MC1061 E.coli strain, and plasmid DNA was purified from individual bacterial colonies by using the High-Speed Plasmid Mini Kit (IBI Scientific). A portion of the plasmid comprising 550 nucleotides upstream and 1.7 kb downstream of the THF position (in respect to the direction of the presumed TLS), as well as the corresponding region in the progeny of the control plasmid, was analyzed by DNA sequencing.
To measure the frequency of UV-induced ura3-G764A reversion, appropriately diluted overnight cultures of the ura3-G764A mutants were plated on synthetic complete medium and SC—ura medium and irradiated immediately with 254 nm UV light at doses indicated in Fig. 4. The plates were incubated for five days at 30°C. Mutant frequencies were then calculated as the ratio of the number of revertants on selective plates to the number of colonies on synthetic complete plates multiplied by the dilution factor.
To isolate UV-induced Ura+ revertants or canavanine-resistant (Canr) mutants of the OK29 and OK30 strains or their rev3Δ, rad30Δ or msh2Δ derivatives, the strains were streaked for single colonies on YPDAU plates and grown for three days at 30°C. Liquid cultures were then started in YPDAU from the individual colonies and grown to the stationary phase. A total of 200 μl of two-fold concentrated saturated cultures were spread on a SC—ura plate or SC—arg supplemented with 60 μg/ml L-canavanine, irradiated immediately with 60 J/m2 of 254 nm UV light, and incubated for seven days (for Ura+ revertants) or five days (for Canr mutants) to allow for colony formation. One revertant or Canr mutant was randomly picked from each plate for DNA sequence analysis. Total yeast DNA was isolated from the revertants and Canr mutants using the MasterPure Yeast DNA Purification Kit (Epicentre). A 5 kb-region comprising 2.5 kb upstream and 2.5 kb downstream of the ura3-G764A mutation site was amplified by PCR using Pfu DNA polymerase kindly provided by Dr. Farid Kadyrov (Southern Illinois University School of Medicine, Carbondale, U.S.A.) and sequenced.
|
10.1371/journal.ppat.1006623 | Role of the central lysine cluster and scrapie templating in the transmissibility of synthetic prion protein aggregates | Mammalian prion structures and replication mechanisms are poorly understood. Most synthetic recombinant prion protein (rPrP) amyloids prepared without cofactors are non-infectious or much less infectious than bona fide tissue-derived PrPSc. This effect has been associated with differences in folding of the aggregates, manifested in part by reduced solvent exclusion and protease-resistance in rPrP amyloids, especially within residues ~90–160. Substitution of 4 lysines within residues 101–110 of rPrP (central lysine cluster) with alanines (K4A) or asparagines (K4N) allows formation of aggregates with extended proteinase K (PK) resistant cores reminiscent of PrPSc, particularly when seeded with PrPSc. Here we have compared the infectivity of rPrP aggregates made with K4N, K4A or wild-type (WT) rPrP, after seeding with scrapie brain homogenate (ScBH) or normal brain homogenate (NBH). None of these preparations caused clinical disease on first passage into rodents. However, the ScBH-seeded fibrils (only) led to a subclinical pathogenesis as indicated by increases in prion seeding activity, neuropathology, and abnormal PrP in the brain. Seeding activities usually accumulated to much higher levels in animals inoculated with ScBH-seeded fibrils made with the K4N, rather than WT, rPrP molecules. Brain homogenates from subclinical animals induced clinical disease on second passage into “hamsterized” Tg7 mice, with shorter incubation times in animals inoculated with ScBH-seeded K4N rPrP fibrils. On second passage from animals inoculated with ScBH-seeded WT fibrils, we detected an additional PK resistant PrP fragment that was similar to that of bona fide PrPSc. Together these data indicate that both the central lysine cluster and scrapie seeding of rPrP aggregates influence the induction of PrP misfolding, neuropathology and clinical manifestations upon passage in vivo. We confirm that some rPrP aggregates can initiate further aggregation without typical pathogenesis in vivo. We also provide evidence that there is little, if any, biohazard associated with routine RT-QuIC assays.
| Differences in the folding and packing of the prion protein into aggregates are thought to be the molecular basis for variability in prion transmissibility. Recombinant prion protein aggregates prepared without non-protein cofactors typically are less tightly packed within residues ~90–160 and are much less infectious than bona fide tissue-derived prions. Substitution of a central cluster of 4 positively charged lysines within residues 101–110 of the prion protein with uncharged residues allows the formation of synthetic recombinant prion aggregates with larger tightly packed cores. Inoculation of these mutant and wild-type synthetic aggregates into animals demonstrated that both the central lysine cluster as well as templating with bona fide prions, strongly influence both the initiation of self-propagating transmissible misfolding of the endogenous prion protein and the resultant neuropathological and clinical outcomes. The results also confirm that some recombinant prion protein aggregates can replicate in vivo without causing classical prion disease pathogenesis.
| Prion diseases, or transmissible spongiform encephalopathies (TSEs), are fatal neurodegenerative diseases of infectious, genetic, or spontaneous origin. A common feature of these diseases is the misfolding and aggregation of the cellular prion protein (PrPC) into TSE-associated (Scrapie, or PrPSc) forms. PrPSc is often resistant to digestion by proteinase K (PK) and, as such, may also be called PrPRes. PK digestion typically leaves a resistant core of residues ~90–231, attributed to ordered aggregation and the transformation of PrPC’s alpha-helical and natively disordered regions into tightly packed assemblies with high beta-sheet content [1, 2]. Conversion of PrPC to PrPSc is driven by a templated seeded polymerization mechanism in which PrPSc recruits and catalyzes the refolding of PrPC into growing PrPSc aggregates [3–7].
Since the inception of the hypothesis that an altered form of prion protein is the infectious agent of TSEs, researchers have attempted to produce prion infectivity from PrP molecules alone under chemically defined conditions in vitro. Various in vitro conversion reactions have been developed [4, 8–18]. These techniques exploit the self-replicating ability of PrPSc using natural PrPC or similarly soluble and protease-sensitive PrP (PrPSen) substrates from various sources. PrPC substrates for these reactions can be contained in crude brain homogenates [8] or cell lysates [19], or be purified from various sources [4, 20]. In addition, recombinant PrPSen (rPrPSen) substrates can be expressed and isolated from E. coli [11, 12, 13, 21]. Sonicated protein misfolding cyclic amplification (PMCA) reactions using brain homogenate as a source of PrPC substrate can recapitulate many attributes of the initial PrPSc seed and generate substantial levels of infectivity [8]. Purified PrPC or rPrPSen substrates have also been converted by PMCA to infectious forms on their own [14], but the generation of high-titered rPrP prions that more fully recapitulate the properties of the initial PrPSc seed has required the presence of cofactors such as RNA and/or phospholipids [21–25].
Prion-seeded polymerization of rPrPSen also occurs in shaken reactions in the absence of cofactors or denaturing conditions at near neutral pH, hereafter referred to as real time quaking-induced conversion (RT-QuIC) conditions [16, 17]. Sano et al [26] demonstrated that some strain properties can be conserved transiently in recombinant PrPRes (rPrPRes) generated under RT-QuIC conditions, but with low infectivity. rPrPRes generated under such conditions has many characteristics of infectious PrPSc including partial PK-resistance, seeding capabilities, and high beta-sheet secondary structure. However, these conversion products tend to have smaller PK-resistant cores comprised of relatively low molecular weight C-terminal fragments of ~8–12 kDa from the region of residues ~160–231 [27, 28]. To date there have been no reports of robust infectivity from rPrPRes generated under these, or similar, non-denaturing conditions at neutral pH and without cofactors. The Sano study showed a modest increase in infectivity in a single round of RT-QuIC amplification; however, the infectivity was lost in subsequent serial RT-QuIC reactions [26]. The authors attribute this to rapid growth of off-pathway aggregates that overtake the faithful templating activity of the original prion seeds.
Reduced specific infectivity in certain rPrP amyloid preparations has been attributed to incomplete refolding of residues 90-~160, which are more tightly packed and highly protease resistant in more infectious forms of PrPSc [9, 14, 15, 29, 30]. Tight packing of the 90-~160 region in the absence of cofactors is impeded by a highly conserved cluster of four lysines within residues 101–110 (the central lysine cluster or CLC) [28, 31–33]. We have hypothesized that this is due to repulsion of their cationic side chains in the absence of ion pairing with suitable anionic residues, cofactors or salts. In any case, wild-type rPrP amyloids formed without cofactors tend to have PK-resistant cores that are restricted to C-terminal residues ~160–231 [28, 32] Mutations of the 4 lysines to alanines (K4A) or asparagines (K4N), which neutralize the lysines within the CLC, allow formation of RT-QuIC products with N-terminally extended PK-resistant cores and with infrared spectra more reminiscent of bona fide PrPSc in terms of β-sheet amide I vibrations [32, 33]. In this study, we have tested whether these scrapie-seeded RT-QuIC conversion products are infectious for rodents. Furthermore, we tested the influence of CLC mutations on the resulting disease phenotypes.
We generated RT-QuIC conversion products seeded with either scrapie brain homogenate (ScBH) or normal brain homogenate (NBH) using recombinant hamster (residues 23–231) wild-type (WT), K4N, or K4A rPrPsen as substrates, hereafter to be called ScBH(WT)RTQ, ScBH(K4N)RTQ, ScBH(K4A)RTQ, NBH(WT)RTQ, NBH(K4N)RTQ and NBH(K4A)RTQ products, respectively. To ensure that the ScBH(WT)RTQ, ScBH(K4N)RTQ, and ScBH(K4A)RTQ products to be used for inoculations did not contain residual infectivity from the initial scrapie seed, we performed three serial rounds of tube-based RT-QuIC-like reactions with 1000-fold dilutions between rounds. These rounds diluted the ScBH seed (a 10−6 brain tissue dilution) to 10−12, a dilution that was far beyond the limit of detection of our rodent bioassays and RT-QuIC [16]. PK-treated and untreated RT-QuIC reaction products from each round were analyzed by Western blot using the polyclonal R20 antiserum directed against C-terminal residues 218–231 (Fig 1). No PrP was detected (Fig 1A, red arrows) in reactions without any rPrPSen substrate that were seeded initially with 10−6 scrapie brain and then subjected to serial RT-QuIC rounds. In the reactions containing rPrPSen substrate we detected predominant ~10 and 12 kDa C-terminal PK-resistant bands in the ScBH(WT)RTQ products similar to our previous findings [32], whereas the NBH(WT)RTQ product showed no PK-resistant rPrP (Fig 1B). In contrast, both the ScBH(K4N)RTQ and NBH(K4N)RTQ products contained multiple PK-resistant bands, but the banding profile differed between the two. Whereas the ScBH(K4N)RTQ and ScBH(K4A)RTQ products had predominant ~12, 13 and 17 kDa bands, the NBH(K4N)RTQ products had a predominant 12-kDa band, with much weaker bands at ~10 and 17 kDa, consistent with previous observations (Fig 1B, [32] [33]). Similar Western blot results were obtained in two additional independent 3-round RT-QuIC reactions with the WT, K4N and K4A rPrP substrates. These results confirmed that both scrapie seeding and the CLC mutation affected the PK-resistant cores of the rPrP aggregates produced with shaking under mild conditions in the absence of cofactors and denaturants.
The various RT-QuIC conversion products were inoculated intracerebrally into hamsters or Tg7 transgenic mice that over-express hamster PrPC [34]. We will call these first-passage (P1) animals ScBH(WT)P1, ScBH(K4N)P1, NBH(WT)P1 and NBH(K4N)P1 according to their inoculum. No clinical signs of prion disease were seen by blinded observers within 652 and 492 days post inoculation (dpi) in any of the P1 hamsters or mice, respectively (Table 1 and S1A & S1B Table). To investigate the possibility of subclinical pathology we collected brains at various time points for RT-QuIC, immunoblotting, and histopathological analyses. Moreover, animals that became sick or injured without apparent outward signs of prion disease were also collected and analyzed when possible.
First, we used end-point dilution RT-QuIC [16] to measure the prion seeding activity in the brains of the animals. Neither the wild-type nor mutant rPrP products of reactions seeded with NBH (i.e., NBH(WT)RTQ and NBH(K4N)RTQ) elicited any detectable seeding activity after either >652 or 492 d incubations in the brains of P1 hamsters or Tg7 mice, respectively (S1A & S1B Table). This was not surprising in the case of the NBH(WT)P1 animals because there was no apparent rPrPRes or ThT-positive amyloid in the inoculated NBH(WT)RTQ preparation (Fig 1 and S1 Fig). However, in the case of NBH(K4N)P1 animals, the inoculum contained partially PK-resistant products. Additionally, both the NBH(K4N)RTQ product (Fig 1 and S1 Fig) and the brains from the Tg7 NBH(K4N)P1 mice collected 2 h after inoculation (n = 2) had RT-QuIC seeding activity, but that activity was not detected in brains collected at later times post-inoculation (S1 Table). Thus, the NBH(K4N)RTQ seeding activity failed to be retained or propagate using endogenous PrPC in vivo.
In contrast, seeding activity was detectable in the brains of P1 animals inoculated with in vitro products seeded with ScBH (i.e. ScBH(WT)RTQ and ScBH(K4N)RTQ) (Table 1, S2 Fig and S1A & S1B Table). In the case of the ScBH(WT)RTQ inoculum, seeding activity was detected in all ScBH(WT)P1 animals at >350 dpi, but none of the 6 hamsters (Fig 2A) and only 1 out of 12 Tg7 mice (Fig 2C) showed more than a 10-fold increase in seeding activity over the levels measured in a control group 2 h after inoculation. Much stronger evidence for propagation of seeding activity was observed after inoculations with the ScBH(K4N)RTQ product, with 3 out of 4 hamsters (Fig 2B) and all 4 Tg7 mice (Fig 2D) harvested after ≥392 days showing increases in seeding activity of ~500–100,000 fold. Reinforcing the effects of the CLC mutations, we also saw persistence of seeding activity in all 5 hamsters inoculated with an ScBH(K4A)RTQ product and increased levels of seeding activity in 2 out of 5 of those animals (S1A Table). The observed effects were dose-dependent because the seeding activity was generally weaker in the brains of the animals inoculated with a 1:10 dilution of the ScBH seeded conversion products in inoculation buffer compared with those diluted 1:1 (S1A & S1B Table). Altogether, these results indicated that both scrapie seeding and the CLC mutations enhanced the in vivo propagation capacities of products generated in vitro in the absence of cofactors or denaturing conditions.
Prompted by RT-QuIC detection of prion seeding activity in the brains of ScBH(WT)P1 and ScBH(K4N)P1 animals, we analyzed P1 brain tissue for the presence of PrPRes made from endogenous PrPC using a C-terminal PrP antiserum (R20). None of the ScBH(WT)P1 or ScBH(K4A)P1 hamsters showed any evidence of PrPRes (Fig 3 [A463], Table 1, and S1A Table). However, one of the six ScBH(K4N)P1 hamsters (A459-2) had a distinct PrPRes banding pattern with three bands at ~14, 17, and 22 kDa (Fig 3 and Table 1). This banding pattern differed from the ~20, 25, and 30 kDa banding pattern of bona fide 263K ScBH PrPRes that was used to initially seed these RT-QuIC reactions (Fig 3). Furthermore, all four ScBH(K4N)P1 Tg7 mice showed the same PK-resistant banding pattern with the same molecular weights of ~14, 17, and 22 kDa (Fig 3 and Table 1). One of the six ScBH(WT)P1 mice (B991-1) also showed a similar PrPRes banding profile (Fig 3 Table 1). However, no PrPRes was detected in the uninoculated or seed-only control animals or NBH(WT)P1 or NBH(K4N)P1 animals (Fig 3 and Table 1). Epitope mapping (S3 Fig) of the PrPRes detected in the ScBH(WT)P1 and ScBH(K4N)P1 animals indicated that the PrPRes generated in vivo following inoculation with ScBH(WT)RTQ and ScBH(K4N)RTQ products had a PK-resistant core that was smaller than that of 263K PrPRes, and spanned approximately from the R18 epitope (residues 143–156) (S3D Fig) to the C-terminal R20 epitope (S3E Fig). Deglycosylation with PNGase F showed that the multiple bands containing this PK-resistant core were due to differences in N-linked glycosylation (Fig 4). This confirmed that rather than being residual unglycosylated rPrP inoculum, the PrPRes being detected in these brains was made endogenously from glycosylated host-derived PrPC.
Blinded histopathological analyses of P1 brains were performed to compare astrogliosis (anti-GFAP staining), spongiosis (hematoxylin and eosin staining), and abnormal PrP deposition (anti-PrP antibody EP1802Y staining). The three ScBH(K4N)P1 hamsters with the highest RT-QuIC seeding activities (≥ ~107 SD50/mg brain; S1A Table) displayed mild histolopathological signs of prion disease (Figs 5 and 6, Table 1, S1A and S2 Tables). One hamster (A457-1) had enhanced PrP staining in the lateral ventricle along ependymal cells and adjacent parenchyma near the ventricle but no astrogliosis or spongiosis (S2 Table). Another (A459-2) had widespread diffuse PrP staining and focal vacuolation in the cerebral cortex (Fig 5), but no astrogliosis (S2 Table). A third hamster (A459-1) showed clear PrPRes deposition, mild spongiform change and moderate astrogliosis in the cerebral cortex and hippocampus (Table 1, Fig 6, S1A and S2 Tables). The two ScBH(K4A)P1 hamsters with the highest seeding activity (≥ ~106 SD50/mg brain; S1A Table) had either subtle diffuse PrPRes deposition in the caudal region of the cerebral cortex or plaque-like deposits lining the meninges and ependymal cells. Neither of these hamsters showed signs of astrogliosis or spongiosis (Fig 5, S1A and S2 Tables). None of the ScBH(WT)P1 hamsters displayed any histopathology despite having detectable, but relatively low, seeding activity (~103−105 SD50/mg brain). Additionally, none of the NBH(WT)P1 or NBH(K4N)P1 hamsters or hamsters inoculated with any of the RT-QuIC products diluted 1:10 showed any histopathological signs of prion disease. Importantly, the neuropathology and PrPRes deposition pattern observed in the three ScBH(K4N)P1 and two ScBH(K4A)P1 hamsters, although different from controls, lacked the severity of neuropathology commonly observed in the thalamus of animals inoculated with our standard 263K scrapie prion stock (Fig 5). Paired with our previous observations, the histopathology suggested that laboratory contamination was not an explanation for the unique pathology we observed.
More obvious histopathological signs of prion disease were observed in all four of the Tg7 mice inoculated with the ScBH(K4N)RTQ product (Figs 6 and 7, S5 Fig, Table 1, S1B and S2 Tables). These mice had prominent spongiform lesions and severe astrogliosis. PrP staining revealed small punctate and granular aggregates in the hypothalamus as well as diffuse punctate staining in the hippocampus, and cortex (S2 Table). One of the mice (B988-1) had additional PrP aggregates in the cerebellum. In contrast to the PrPRes distribution pattern observed in ScBH(K4N)P1 mice, positive control Tg7 mice inoculated with 263K ScBH had localized neuropathology and PrPRes in the thalamus and brain stem (S4 Fig and S2 Table) and minimal pathology in the cortex and hippocampus (Fig 7 and S5 Fig). Moreover, each of the ScBH(K4N)P1 mice with histological lesions were also positive for PrPRes by Western blot and each had high (≥ 5x108 SD50/mg brain) RT-QuIC seeding activity (S1 Table). One ScBH(WT)P1 mouse showed histological signs of prion disease which included small punctate and granular PrP aggregates, along with mild spongiform change and weak astrogliosis (Figs 6 and 7, S5 Fig, Table 1, S1B and S2 Tables). Three additional ScBH(WT)P1 mice showed weak, localized astrogliosis but no other histopathological signs of TSE disease. None of the uninoculated mice or the mice inoculated with NBH(WT)RTQ, NBH(K4N)RTQ or 10-fold dilutions thereof showed any histopathology (Table 1 and S1B Table). Thus, the ScBH(K4N)RTQ product elicited neuropathological lesions that were distinct from those of 263K scrapie-inoculated animals and absent in animals receiving the other inocula.
The transmission properties of P1 brain homogenates was assessed by second passages (P2) into Tg7 mice. Brain homogenates from 3 of the 4 ScBH(K4N)P1 mice (B987-1, B987-2, and B988-2) and the one ScBH(WT)P1 mouse (B991-1) with strong seeding activity and detectable PrPRes were used to inoculate the P2 Tg7 mice (S1C Table). Regarding the ScBH(K4N)P2 mice, three died without overt clinical signs of disease, one at 101 and two at 121 dpi. However, these mice had unused nestlets, which is a common early manifestation of scrapie disease in mice [35]. At 143 dpi, two additional ScBH(K4N)P2 mice showed signs of prion disease, including pronounced myoclonus, slight hyperactivity, head tilting, sideways gait, ungroomed coats, and unused nestlets (Table 2 and S1C Table). These and 5 additional non-clinical ScBH(K4N)P2 mice were euthanized at 143 dpi for further biochemical and histological analysis (see below). Brains from the four remaining ScBH(K4N)P2 mice were harvested at 433 dpi after prolonged neurological signs including myoclonus, unused nestlets, poor grooming, and jerky movements for >5 months, and, in one case, designation by our standard criteria as having clinical prion disease. All of the ScBH(K4N)P2 mice, regardless of time of death, had abundant RT-QuIC seeding activity (~107−109 SD50/mg tissue; results from representative ScBH(WT)P2 and ScBH(K4N)P2 mice are shown in S6 Fig). However, because the inoculum delivered ~105 SD50/mg tissue to the brain, we could not establish that overall seeding activity levels increased post-inoculation by more than ~10,000-fold, with the latter being the accumulation at 433 dpi (Fig 2F). All of the ScBH(K4N)P2 mice also had PrPRes with ~14, 17, and 22 kDa representing un-, mono- and di-glycosylated bands, respectively (Fig 8, Table 2 and S1C Table), similar to the bands seen in the corresponding P1 mice (Fig 3).
Regarding the ScBH(WT)P2 mice, two were harvested at 143 dpi with possible clinical signs. These mice had RT-QuIC seeding activities of SD50s ≥5.00x107 (Fig 2E and S1C Table) and PrPRes (Table 2 and S1C Table). The PrPRes in these mice was indistinguishable from that of the ScBH(K4N)P2 mice shown in Fig 8. At 245–251 dpi the three remaining ScBH(WT)P2 mice displayed clinical signs and higher prion seeding activities (S1C Table and Fig 2). Interestingly, immunoblot analysis of brain homogenates of these same mice revealed, in addition to the truncated ~14, 17, and 22 kDa bands seen in the ScBH(K4N)P2 and 143-dpi ScBH(WT)P2 mice (without PNGase F treatment), higher molecular weight bands reminiscent of the ones found in 263K ScBH-inoculated mice (Fig 8; bracket). Following PNGase F treatment of this material we detected two predominant bands: the ~14 kDa band observed in the ScBH(K4N)P2 mice and a ~20 kDa band similar to 263K PrPRes (Fig 8; blue and red arrows, respectively). Importantly, none of the NBH(WT)P2, NBH(K4N)P2 or uninoculated control mice had any clinical signs or seeding activity in their brains. Furthermore, immunoblot analysis confirmed the absence of PrPRes (Table 2 and S1C Table).
After second passage, all of the ScBH(K4N)P2 and ScBH(WT)P2 mice showed at least two out of three of the neuropathological lesions examined, namely spongiosis, astrogliosis and PrP deposition but with varied intensity and distribution (Figs 6 and 9, S7 Fig, S1C and S2 Tables). The most consistent differences between the ScBH(K4N)P2 and ScBH(WT)P2 were i) the level of spongiosis at the relatively early 143 day timepoint (Fig 6) and ii) the more focal lesions in the ScBH(K4N)P2 mice compared to the ScBH(WT)P2 mice (S2 Table). The ScBH(K4N)P2 mice that displayed the prolonged clinical phenotype, as mentioned above, also had higher levels of spongiosis, astrogliosis, and PrP deposition (Figs 6F and 9, S7 Fig and S2 Table) than those that displayed a more acute phenotype. None of the NBH(WT)P2, NBH(K4N)P2 or uninoculated age-matched control mice showed any indication of TSE disease by histopathology (Table 2 and S1C Table).
Collectively, these results indicated that upon second passage, infections initiated by ScBH(K4N)RTQ and ScBH(WT)RTQ, but not their analogous NBH-seeded products, induced a serially transmissible prion disease. However, resulting diseases gave distinct clinical, biochemical and histopathological manifestations compared to classical 263K scrapie, indicating that both scrapie templating and mutation of the CLC lysines affect the properties of the synthetic prion strains produced in vitro from rPrPSen alone without cofactors, denaturants or physiologically implausible pHs.
Many, if not most, proteins are capable of assembling into self-propagating amyloid fibrils under certain conditions [36, 37]. A key question is whether such fibrils can be initiated and propagated under physiological conditions. Infectious prion aggregates must be able to survive in the host long enough to initiate formation of new aggregates at a rate that exceeds that of any clearance processes. This requires that the prions gain access to suitable pools of normal substrate molecules for recruitment into growing prion aggregates. In the case of prions made of PrP molecules, the interactions between incoming prions and PrPC is likely to be constrained by the GPI-anchored membrane association and heavy glycosylation of most potential PrPC substrate molecules, as well as the local availability of suitable cofactor molecules. If the infecting aggregate is bona fide PrPSc coming from another mammal that expresses wild-type PrPC, then the PrPSc is derived from endogenous PrPC molecules and thus must have a conformation that can accommodate them. However, this might not be true of synthetic aggregates that are made de novo in in vitro reactions from recombinant PrP molecules that lack post-translational modifications. Such assemblies might be packed into conformations that, upon inoculation, are less able to accommodate new monomers that are membrane-bound or laden with bulky glycans. In such cases, the aggregate replication rate might be diminished, and unable to outpace clearance mechanisms and accumulate to toxic levels within the lifespan of the host. Our current data provide evidence that CLC mutations can not only affect the conformation of rPrP amyloids but also their ability to propagate utilizing endogenous wild-type PrPC in vivo and cause pathology.
Here, and in other studies in which prions were made from recombinant PrP in physiologically compatible buffers but in the absence of natural cofactors, initial templating by PrPSc enhanced the infectivity of the rPrP fibrils [14, 26]. In our case, the ScBH(WT)RTQ and ScBH(K4N)RTQ products both were retained or propagated in the host and, on second passage, caused clinical prion disease. By contrast, the NBH(WT)RTQ and NBH(K4N)RTQ did not. The lack of infectivity in NBH(K4N)RTQ was not solely due to a lack of seeding activity, because it propagated in serial RT-QuIC reactions and generated amyloid with a PK-resistant core, albeit one that differed somewhat from those of ScBH(WT)RTQ and ScBH(K4N)RTQ (Fig 1). Furthermore, the level of inoculated NBH(K4N)RTQ seeding activity was 2–3 logs lower than that of the ScBH(WT)RTQ and ScBH(K4N)RTQ inocula (S1 Fig). However, seeding activity was detectable in NBH(K4N)P1 Tg7 mice 2 h post-inoculation at levels 1–2 logs lower than that of the ScBH(K4N)P1 Tg7 mice, yet the untemplated inocula did not amplify in vivo. Thus, either the untemplated conformation, which was distinct from the ScBH templated conformation [32], or the lower inherent seeding activity of the NBH(K4N)RTQ product, may have rendered it non-pathogenic upon inoculation into rodents. Perhaps spontaneously arising (NBH-seeded) amyloids, even if partially PK-resistant, have a conformational architecture that is more easily targeted for destruction, or unable to accommodate endogenous PrPC or cofactor molecules efficiently enough to outpace degradation. Our data suggest that this is less of a limitation when the rPrPRes amyloid is templated by bona fide PrPSc.
Nonetheless, scrapie templating alone, without cofactors, was insufficient to enable replication of fully pathogenic prions in vitro using recombinant WT, K4N or K4A PrP because neither ScBH(WT)RTQ, ScBH(K4N)RTQ, nor ScBH(K4A)RTQ induced clinical disease on first passage. However, one hamster and 5 of the PrPC overexpressing Tg7 mice inoculated with either ScBH(WT)RTQ or ScBH(K4N)RTQ accumulated high seeding activities (>108 SD50/mg) (Fig 2; S1 Table). Thus, consistent with conclusions of previous studies [38–40], the accumulation of high levels of seeding activity was not necessarily coincident with clinical disease.
From a practical perspective, it is notable that we detected increases in seeding activity only when (i) certain ScBH-seeded RT-QuIC products were inoculated into transgenic mice that massively overexpress PrPC or (ii) when unnatural mutant RT-QuIC products [ScBH(K4N)RTQ or ScBH(K4A)RTQ] were inoculated into wild-type hamsters. The fact that the ScBH(WT)RTQ products failed to propagate above discernable input levels in wild-type animals suggests that there is little, if any, biohazard associated with performing RT-QuIC assays under routine clinical circumstances in which WT rPrP substrates would be used by humans expressing normal amounts of PrPC. Indeed, in our bioassays 2.5 μg of RT-QuIC reaction product was inoculated directly into the brains of hamsters without any apparent pathological effect, making these RT-QuIC products at least 106−109-fold less pathogenic than bona fide 263K PrPSc (e.g. [1]) and PrPSc-seeded products of brain homogenate-based PMCA reactions [41].
Interestingly, Kim et al induced clinical disease on first passage into hamsters inoculated with recombinant hamster 90–231 rPrPRes generated in sonicated PMCA reactions containing 0.1% SDS and 0.1% Triton X100 [14]. In the present study, the ScBH(WT)RTQ and ScBH(K4N)RTQ products were shaken rather than sonicated and contained only 0.002% SDS. Whether sonication, the much higher detergent content, or the use of N-terminally truncated PrP in the Kim et al study was most responsible for the greater pathogenicity of the rPrP PMCA products is not clear. Notably, SDS has an anionic moiety that might ion-pair with the cationic CLC to facilitate tighter packing in this region. The hydrophobic tail of SDS could also play an important role, analogous to previously observed effects of detergents [30, 42] and lipids [31, 43] on the infectious properties and protease-resistant core of recombinant PrP amyloids.
Although scrapie seeding enhanced the pathogenicity of both the ScBH(WT)RTQ and ScBH(K4N)RTQ products, the mutant ScBH(K4N)RTQ material produced more extensive seed amplification and prion pathogenesis on first passage into both hamsters and Tg7 mice (Tables 1 and 2 and S1A and S1B Table). On second passage in the mice, both of these infections caused clinical disease and extensive signs of prion pathogenesis, although the appearance of spongiosis and abnormal PrP was slower in the ScBH(WT)RTQ mice (Fig 6). Given that single preparations of ScBH(WT)RTQ and ScBH(K4N)RTQ products were inoculated in the first passage, it remains possible that these differences observed in vivo might have been due to stochastic events during in vitro propagation that were not determined by the WT or K4N sequences specifically. However, the biochemical distinctions between the ScBH(WT)RTQ and ScBH(K4N)RTQ products shown in Fig 1B were consistently observed in three independent 3-round RT-QuIC reactions. Thus, at present the simplest explanation for the in vivo phenotypes is that they are linked in some way to the observed structural/conformational differences between the ScBH(WT)RTQ and ScBH(K4N)RTQ products. Nonetheless, it is difficult to exclude the possibility that the infectious agents produced in the RT-QuIC reactions were low-abundance rPrP species rather than the bulk of the PK-resistant forms that were detected by immunoblotting. Indeed, none of the combined RT-QuIC products that we have generated showed the same predominance of the 17-19-kDa band that is exhibited by PK- and PNGase F-treated 263K PrPSc (as shown in Fig 4). Thus, although our biochemical and bioassay data support the idea that the CLC residues and mutations thereof influence misfolding and transmissibility of PrP structures formed in the absence of cofactors, we cannot discriminate clearly between the possibilities that (i) the ScBH(WT)RTQ and ScBH(K4N)RTQ products are relatively uniform amyloids that are transmissible and ultimately pathogenic, but much less so than 263K PrPSc, (ii) the products are mixtures of conformers, only a subset of which are transmissible and/or pathogenic but, again, less so than 263K PrPSc because we never detected clinical disease on first passage, and (iii) the products are the same as described in ii) but with interference of seed propagation in vivo by non-infectious or non-pathogenic conformers in the preparation.
In any case, our findings raise the question of why the K4N and K4A mutations promote the formation of more pathogenic scrapie-templated seeds in vitro in the absence of cofactors. One possibility is that neutralization of the cationic lysine sidechains relieves the need for charge compensation by anionic cofactors in vitro in order to achieve a conformation within the ~90–160 region [32, 33] that more readily enables the conversion and incorporation of PrPC in vivo. However, these mutations in rPrP clearly did not allow for fully faithful propagation of the 263K PrPSc strain in RT-QuIC reactions. Thus, the latter must require other molecules or microenvironments that would be available in vivo. A key biochemical feature of the prions in the ScBH(K4N)P1 mice was their distinctive ~14-kDa PK-resistant core, a feature that was faithfully propagated in ScBH(K4N)P2 mice as well. These results suggest that multimers with only this 14-kDa C-terminal core can be infectious. However, the fact that the generation of these multimers was aided by inoculation of the ScBH(K4N)RTQ products with N-terminally extended cores provides evidence that structure that is outside the 14-kDa C-terminal core influences the templating activity and transmissibility of synthetic PrP amyloids. Interestingly, the ScBH(WT)P2 mice generated an additional ~20 kDa PrPRes species (Fig 8). This suggests that the ScBH(WT)RTQ inoculum may have allowed the strain to diverge and generate both forms of PrPSc, as has been described previously as “deformed templating” [44]. While the ~20 kDa species was only seen in the three ScBH(WT)P2 mice with clear clinical signs of TSE disease, we did not observe the ~20 kDa core in the clinical ScBH(K4N)P2 mice, indicating that this conformer is not required at immunoblot-detectable levels for the induction of clinical TSE disease.
Without PrPSc templating, neither the K4N nor the WT rPrP substrates formed aggregates that initiated seed amplification or pathology in vivo, despite the seeding activity observed in the NBH(K4N)RTQ products. This confirms that recombinant PrPC molecules do not readily adopt an infectious form in the absence of a PrPSc template to guide refolding and assembly. Altogether, the results of this study support the concept that tight folding in the vicinity of the CLC, which is near the N-termini of the PK-resistant and solvent-excluding cores of PrPSc protomers, is critical in the formation of pathogenic TSE prions. Moreover, the CLC lysines may inhibit the propagation of TSE prions in the absence of charge-compensating cofactors.
Recombinant hamster PrPC and mutants were purified as previously described [45]. Briefly, BL21(DE3) Rosetta 2 Escherichia coli containing the pET41 vector (EMD, Billerica, MA, USA) with the PrP sequence (Syrian hamster residues 23–231; accession no. K02234, or the PrP sequence including the lysine mutations described in [32]) were grown in Luria broth medium in the presence of kanamycin and chloramphenicol. Protein expression was induced using the autoinduction system and purified using nickel-nitrilotriacetic acid superflow resin (Qiagen) with an ÄKTA pure chromatography system (GE Healthcare Life Sciences) as previously described [45]. The purified protein was extensively dialyzed into 10 mM sodium phosphate buffer (pH 5.8) and stored at −80°C. Protein concentration was determined by measuring absorbance at 280 nm.
Scaled-up RT-QuIC reactions were performed as previously described [32]. Reaction mixtures were the same as for RT-QuIC with the exceptions of not including ThT and using the indicated substrates. Reactions were performed at a volume of 1 ml in a 1.5 ml tube in an Eppendorf Thermomixer-R at 700 rpm to generate RT-QuIC products. Reactions were run for 15 h with cycles of 60 s shaking and 60 s of rest throughout the incubation at 42°C. Tube-based 1000 uL RT-QuIC reactions were seeded with 2μL of a 10−6 263K prion infected (263K) or normal Syrian Golden hamster (NBH) brain tissue dilutions to initiate templated conversion of the substrate and as a specificity control, respectively. Reactions were limited to an incubation of 15 hours to minimize the incidence of spontaneous conversion in the WT rPrPSen reactions. Following a 15 hour incubation one microliter of the first round reaction was used to seed a 1 ml second round reaction resulting in a 1:1000 fold dilution of the seed. After a 15 hour incubation we repeated the process by taking one microliter of the second round conversion product to seed a third and final round at an additional 1:1000 dilution of the seed. Three serial passages of the scaled up reactions were performed to achieve a final brain tissue dilution of at least 10−12.
Brain homogenates were prepared as previously described [45] with the exception of being homogenized in 1X PBS and stored at -80°C. For RT-QuIC analysis BHs were serially diluted in 0.1% SDS (sodium dodecyl sulfate, Sigma)/N2 (Gibco)/PBS as previously reported [46].
RT-QuIC was performed as previously described [45]. Briefly, for each reaction well 2ul of the indicated brain dilution was added to 98 μL of a reaction mixture resulting in a final concentration of 10 mM phosphate buffer (pH 7.4), 300 mM NaCl, 0.1 mg/mL Ha90-231 rPrPSen, 10 μM ThT, and 1 mM ethylenediaminetetraacetic acid tetrasodium salt (EDTA). The reaction mix was loaded into each well of a black-walled 96-well plate with a clear bottom and reactions were seeded with 2 μL of the indicated dilution for a final reaction volume of 100 μL containing 0.002% SDS. Plates were sealed and incubated in a BMG FLUOstar plate reader for 30 h with cycles of 60 s shaking and 60 s of rest throughout the incubation at 55°C. ThT fluorescence measurements (450 ± 10 nm excitation and 480 ± 10 nm emission; bottom read) were taken every 45 min. Fluorescence reactions were judged to be positive or negative prior to 24 h as described previously [47]. Briefly, to compensate for differences between the plate readers, we averaged data from replicate wells and normalized to a percentage of the maximal fluorescence response of the instrument. The obtained values were plotted against the reaction times. Samples were judged to be positive if two or more wells crossed a 10% maximum ThT fluorescence threshold prior to the 24 h cutoff. Spearman-Kärber analysis was used to estimate the seeding dose at which 50% of the replicate wells became positive (SD50) [16, 48].
Prior to inoculation procedures, animals were anesthetized with isoflurane for restraint and pain reduction. Animals were euthanized using carbon dioxide asphyxiation using standard methods recommended by American Veterinary Medical Association guidelines (https://www.avma.org/KB/Policies/Documants/euthanasia.pdf). Inoculations were performed as previously described [49]. RT-QuIC reaction products were diluted either 1:1 v/v or 1:10 v/v in inoculation buffer (phosphate-buffered balanced saline supplemented with 2% fetal bovine serum) and brain homogenates (10% in PBS) were diluted to the desired final concentration in inoculation buffer. Four- to 6-week-old Tg7 mice, overexpressing Syrian golden hamster (SGH) PrP ~5 fold [34], or SGH were inoculated intracerebrally (i.c.) with 50 μl of the inocula. Animals were observed daily for signs of neurological disease including unused nestlets, poor grooming, kyphosis, ataxia, wasting, delayed response to stimuli, and somnolence and were euthanized on confirmation of progressive neurological disease or at designated time points.
The animal experimental protocol was reviewed and approved by the Rocky Mountain Laboratories Animal Care and Use Committee (Animal Study Protocol 2014–005 and 2016–039). The Rocky Mountain Laboratories are fully accredited by the American Association for Laboratory Animal Care and 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.
RT-QuIC reaction products containing 0.1 mg/ml rPrP were treated with 15 μg/ml PK (CALBIOCHEM) at 37°C for 1 h. 20% Brain homogenates in PBS were treated with 50 μg/ml PK at 37°C for 1 h. Reactions were stopped by diluting in sample buffer containing a final concentration of 4 M urea, 4% SDS, 2% β-mercaptoethanol, 8% glycerol, 0.02% bromophenol blue and 50 mM Tris-HCl; pH 6.8. Sample were then analyzed using Western blot techniques as previously described [32]. Equal volumes of PK-treated reactions were run on 10% Bis-Tris NuPAGE gels (Invitrogen). Proteins were transferred to an Immobilon P membrane (Millipore) using the iBlot Gel Transfer System (Invitrogen). Immunoblotting was carried out using the following α-PrP antibodies: R20 (epitope: residues 218–231), R30 (epitope: 89–103) [50], 3F4 (Millipore; epitope: 109–112), mAB132 (epitope: 119–127) [51], and R18 (epitope: 143–156) [50, 52] as previously described [32, 33].
As described previously [49], SGH or mice were euthanized, brains were removed, and the sagittal half contralateral to the site of inoculation was placed in 50 ml of 3.7% phosphate-buffered formalin for 3 to 5 days before dehydration and embedding in paraffin. New dissection tools were used for each dissection. Antigen retrieval was performed as previously described [49], followed by staining with rabbit monoclonal antibody (EP1802Y at a 1:6000 dilution) to PrP (Abcam). Astrocyte detection was performed by staining with polyclonal rabbit anti-glial fibrillary acidic protein (anti-GFAP; Dako). Slides were also stained for observation of overall pathology using a standard hematoxylin-eosin (H&E) protocol. All histopathology slides were analyzed by observers blinded to the animal inoculation groups. Histopathology was scored on a scale of 0 to 3 with 0 being no pathology in the entire brain and 3 being severe widespread pathology throughout the brain. Animals with a score above 0 in any of the three lesion types was marked positive for histopathology.
|
10.1371/journal.pgen.1000433 | A Genome-Wide Association Study Confirms VKORC1, CYP2C9, and CYP4F2 as Principal Genetic Determinants of Warfarin Dose | We report the first genome-wide association study (GWAS) whose sample size (1,053 Swedish subjects) is sufficiently powered to detect genome-wide significance (p<1.5×10−7) for polymorphisms that modestly alter therapeutic warfarin dose. The anticoagulant drug warfarin is widely prescribed for reducing the risk of stroke, thrombosis, pulmonary embolism, and coronary malfunction. However, Caucasians vary widely (20-fold) in the dose needed for therapeutic anticoagulation, and hence prescribed doses may be too low (risking serious illness) or too high (risking severe bleeding). Prior work established that ∼30% of the dose variance is explained by single nucleotide polymorphisms (SNPs) in the warfarin drug target VKORC1 and another ∼12% by two non-synonymous SNPs (*2, *3) in the cytochrome P450 warfarin-metabolizing gene CYP2C9. We initially tested each of 325,997 GWAS SNPs for association with warfarin dose by univariate regression and found the strongest statistical signals (p<10−78) at SNPs clustering near VKORC1 and the second lowest p-values (p<10−31) emanating from CYP2C9. No other SNPs approached genome-wide significance. To enhance detection of weaker effects, we conducted multiple regression adjusting for known influences on warfarin dose (VKORC1, CYP2C9, age, gender) and identified a single SNP (rs2108622) with genome-wide significance (p = 8.3×10−10) that alters protein coding of the CYP4F2 gene. We confirmed this result in 588 additional Swedish patients (p<0.0029) and, during our investigation, a second group provided independent confirmation from a scan of warfarin-metabolizing genes. We also thoroughly investigated copy number variations, haplotypes, and imputed SNPs, but found no additional highly significant warfarin associations. We present power analysis of our GWAS that is generalizable to other studies, and conclude we had 80% power to detect genome-wide significance for common causative variants or markers explaining at least 1.5% of dose variance. These GWAS results provide further impetus for conducting large-scale trials assessing patient benefit from genotype-based forecasting of warfarin dose.
| Recently, geneticists have begun assaying hundreds of thousands of genetic markers covering the entire human genome to systematically search for and identify genes that cause disease. We have extended this “genome-wide association study” (GWAS) method by assaying ∼326,000 markers in 1,053 Swedish patients in order to identify genes that alter response to the anticoagulant drug warfarin. Warfarin is widely prescribed to reduce blood clotting in order to protect high-risk patients from stroke, thrombosis, and heart attack. But patients vary widely (20-fold) in the warfarin dose needed for proper blood thinning, which means that initial doses in some patients are too high (risking severe bleeding) or too low (risking serious illness). Our GWAS detected two genes (VKORC1, CYP2C9) already known to cause ∼40% of the variability in warfarin dose and discovered a new gene (CYP4F2) contributing 1%–2% of the variability. Since our GWAS searched the entire genome, additional genes having a major influence on warfarin dose might not exist or be found in the near-term. Hence, clinical trials assessing patient benefit from individualized dose forecasting based on a patient's genetic makeup at VKORC1, CYP2C9 and possibly CYP4F2 could provide state-of-the-art clinical benchmarks for warfarin use during the foreseeable future.
| Warfarin is the most widely prescribed anticoagulant for reducing thromboembolic events that often give rise to stroke, deep vein thrombosis, pulmonary embolism or serious coronary malfunctions [1]. A combination of genetic and non-genetic factors cause Caucasians to exhibit 20-fold interindividual variation in required warfarin dose needed to achieve the usual therapeutic level of anticoagulation as measured by the prothrombin international normalized ratio or INR [2]–[4]. Thus, in the absence of information (genotypic, clinical, etc.) for predicting each patient's required warfarin dose, initial prescribed doses may be too low (risking thrombosis) or too high (risking over-anticoagulation and severe bleeding). Warfarin's risk of serious side effects, narrow therapeutic range, and wide interindividual variation in warfarin dose have focused attention on the need to better predict dose in the initial stage(s) of treatment.
We and others have shown that the warfarin drug target VKORC1 (vitamin K epoxide reductase complex, subunit 1) contains common polymorphisms that account for a major portion (∼30%) of the variance in required warfarin dose [5],[6], and we have recently evaluated ∼1500 Swedish patients of the Warfarin Genetics (WARG) cohort in the largest study to date showing likely patient benefit from genetic forecasting of dose [3]. The study confirmed that SNPs in VKORC1 and in the warfarin-metabolizing gene CYP2C9 (cytochrome P450, family 2, subfamily C, polypeptide 9) predict ∼40% of dose variance while non-genetic factors (age, sex, etc.) jointly account for another ∼15%. The robust and now widely replicated associations of warfarin dose with VKORC1 and CYP2C9 have provided one of the most successful applications of pharmacogenetics to date [7] and offer promise for genetic predication of required dose in a clinical setting [3].
Knowledge of major predictors of warfarin dose also impacts the methodology for finding further dose-related genes. In early candidate gene work with a small sample of 201 patients [8], we noted that univariate regression (with tested SNP as the only dose predictor) could statistically detect warfarin association with VKORC1 and with one of two non-synonymous CYP2C9 SNPs (*3) known to influence warfarin dose (Table 1 in [8]). However, a second non-synonymous CYP2C9 SNP (*2) with known but weaker influence on warfarin dose was not detected by univariate regression, but *2 was statistically significant in multivariate regression adjusted for the other known genetic and non-genetic predictors of dose (Table 3 in [8]). These empirical results in a small warfarin sample provided a signpost underscoring the potential importance of multivariate regression for detecting weak effects in studies now searching for additional warfarin genes across the genome.
A genome-wide association study (GWAS) enables a systematic search of the entire genome for genetic factors that cause any inherited trait. This method has successfully identified susceptibility loci for common diseases [9], and is beginning to be applied to pharmacogenomics. A recent warfarin GWAS in 181 patients did not detect other genetic factors with major effects on warfarin dose beyond VKORC1 [10] but was underpowered for identifying loci with a moderate contribution. We have now genotyped 325,997 SNPs in 1053 patients of the WARG cohort and here report the first GWAS that is sufficiently powered to detect additional genetic factors that may only modestly influence warfarin dose.
Figure 1A and the first line of Table 1 summarize results of testing 325,997 GWAS SNPs for association with warfarin dose by univariate regression. The strongest associations were at multiple SNPs in and near VKORC1 (Figure 1A) with the lowest p-value given by rs9923231 (P = 5.4×10−78). In prior fine-mapping of the VKORC1 locus [8], we identified rs9923231 as one of three SNPs located in introns or immediately flanking VKORC1 that exhibit almost perfectly concordant genotypes yielding pairwise linkage disequilibrium (LD) r2≈1 and which define the warfarin-sensitive A-T-T haplotype at rs9923231-rs9934438-rs2359612 (see also [11]). These highly concordant SNPs were the best predictors of warfarin dose in our previous study and in this GWAS analysis (p<5.4×10−78) and completely accounted for the dose variance explained by all other fine-mapping SNPs near VKORC1 [8]. The group of SNPs with the second lowest univariate p-values clustered around CYP2C9 which contains two non-synonymous exonic SNPs whose minor alleles (*2, *3) impair warfarin metabolism and are well known to be associated with warfarin dose. In our previous work [8], we discovered an unusual SNP (rs4917639) whose minor allele is almost perfectly associated with the “composite” CYP2C9 allele formed by combining *2 and *3 into a single allele. Indeed, the GWAS results (1053 subjects) confirmed that LD is nearly perfect (pairwise r2≈1.0) between rs4917639 and the composite of *2 and *3. Thus, the highly significant univariate result for rs4917639 (R2 = 0.121, p<3.1×10−31) reflects the combined effect of CYP2C9*2 rs1799853 (R2 = 0.038, p<8.8×10−13) and CYP2C9*3 rs1057910 (R2 = 0.080, p<4.5×10−17). Figure 1A therefore indicates p-values for this composite SNP as well as for *2 and *3.
Figure 1B and Table 1 (lines 2 to 5) show the results of multivariate regression analysis in which individual SNPs were tested for association with warfarin dose after adjustment for established genetic and non-genetic predictors of dose. The only SNP reaching genome-wide significance (p<1.5×10−7) was a non-synonymous SNP (rs2108622) in exon 2 of CYP4F2 (cytochrome P450, family 4, subfamily F, polypeptide 2) introducing a Val to Met amino acid change at position 433 (V433M). SNP rs2108622 predicts additional dose variance (∼1.1%) that is independent of the variance already explained by VKORC1 and CYP2C9. As noted in the Introduction, our early studies with a small sample of 201 Swedish patients failed to detect the weak CYP2C9*2 effect on dose by univariate regression but *2 was significant in multiple regression [8]. The results in Table 1 with rs2108622 of CYP4F2 show the same phenomenon with a p-value of 1.6×10−5 in univariate regression (line 1) but progressively lower p-values as known predictors are added to the multivariate model so that for the full model a p-value of 8.3×10−10 is achieved which is far below genome-wide significance (p<1.5×10−7). The CYP4F2 association was further confirmed by testing an independent replication panel of 588 Swedish warfarin patients who gave a multivariate p-value of 0.0029 and a total overall p-value of 3.3×10−10 when combined with the GWAS subjects (Table 2). During preparation of this paper, a candidate gene study of drug-metabolizing and transporter genes independently discovered the association of rs2108622 and CYP4F2 with warfarin dose, providing further confirmation [12].
To increase the power of our multivariate regression model and possibly detect additional weak effects, we added CYP4F2 (rs2108622) to the model as a predictor and conducted further analyses. First, we retested the GWAS SNPs, but no new SNPs reached genome-wide significance and there was also no apparent excess of SNPs at lower significance thresholds (Figure S1). We also tested warfarin association with haplotypes and with ungenotyped SNPs imputed at 2.2 million HapMap SNPs, but no haplotype or imputed SNP approached genome-wide significance in a genomic region not containing VKORC1, CYP2C9 or CYP4F2. To explore whether copy number variations (CNVs) detectable by the HumanCNV370 array might influence warfarin dose, we used rigorous quality control and retained 879 samples calling 2530 CNVs (see Materials and Methods). None of the CNV loci were significantly associated with dose after correction for multiple testing (lowest CNV p-value was 1.1×10−4 which exceeds 0.05/2530≈2.0×10−5). We note that probe density in many of the detected CNVs is not optimal for conducting association analyses and these results should therefore be viewed as preliminary.
Finally, after excluding SNPs near VKORC1, CYP2C9 and CYP4F2, we identified 40 other loci containing one or more GWAS SNPs with p-values below 2.0×10−4 and we genotyped 40 SNPs representing these loci in a follow-up sample of 588 Swedish warfarin patients. However none of the 40 loci replicated for association with warfarin dose, the lowest p-value being 0.04 which is not significant after correction for 40 tests (Table S1). Having not found evidence for any additional genetic modulators of dose, we examined the entire data set (GWAS plus followup samples) for evidence of statistical interaction between pairs of the established dose predictors (VKORC1, CYP2C9, CYP4F2, age, sex). None of the pairs exhibited statistically significant interaction after p-values were corrected for the 15 interaction tests (Table S2).
We also performed a GWAS for a secondary trait (“over-anticoagulation”) which we previously found was associated with VKORC1 and CYP2C9 in a candidate gene study [3]. By titrating warfarin dose, physicians attempt to achieve a target level of anticoagulation determined by a reading of 2.0 to 3.0 for the prothrombin international normalized ratio (INR), which is the ratio of time required for a patient's blood to coagulate relative to that of a reference sample. However over-anticoagulation (defined as an INR above 4.0) sometimes occurs and, using Cox regression, our GWAS tested for SNP association with the occurrence of over-anticoagulation in patients during the first 5 weeks of treatment (see Materials and Methods: Association testing of SNPs and haplotypes). We observed genome-wide significant association (p<1.5×10−7) at several SNPs in and around VKORC1 including rs9923231 (P = 8.9×10−9), but no other SNPs achieved genome-wide significance including CYP2C9*3 (p<4.0×10−5), CYP2C9*2 (p = 0.93), or the “composite” *2*3 SNP rs4917639 (p<0.007) (Figure S2). However we note that our previous candidate gene study evaluated a larger sample set (1496 WARG subjects) which yielded genome-wide significant association with over-anticoagulation for both VKORC1 rs9923231 (P = 5.7×10−11) and CYP2C9*3 (P = 1.5×10−9) [3]. To explore whether these SNPs might cause over-anticoagulation independent of altering the required (i.e., administered) warfarin dose, we added required dose to the Cox regression model as a predictor of over-anticoagulation, and found that both VKORC1 and CYP2C9*3 have a significant effect independent of dose (P<0.05) (Table S3).
We conducted the first GWAS sufficiently powered to detect DNA variants with a modest influence on the warfarin dose needed to achieve therapeutic anticoagulation. In univariate analysis of GWAS SNPs (Figure 1A), we identified extremely strong association signals (p = 10−78 to 10−13) at SNPs in and near VKORC1 and CYP2C9, two genes already known to explain ∼30% and ∼12% of warfarin dose variance, respectively. By applying multivariate regression adjusting for known genetic and non-genetic predictors of dose (Figure 1B), we also detected genome-wide significance of p<8.3×10−10 at CYP4F2 (rs2108622) that accounted for approximately 1.5% of dose variance. The increased power of multivariate regression to detect this modest effect is nicely illustrated in Table 1 which shows a higher univariate p-value for CYP4F2 (p<1.6×10−5) but progressively lower multivariate p-values as known predictors of dose are added to the regression model. We confirmed the CYP4F2 association in a second large sample set and the association was also reported by another group [12] during preparation of our work, thus fully establishing the genuine effect of CYP4F2 (see also [10] where CYP4F2 explained ∼1% dose variance with nominal p<0.043 significance). Although multivariate regression has not been widely used to increase power in other GWAS analyses because known genetic variants usually explain little phenotypic variance, the potential for power increase is perhaps obvious if known predictors do explain substantial variance. Thus multiple regression has, for example, been previously advocated for linkage analyses of line crosses [13],[14].
To estimate the multivariate regression power of our GWAS (1053 subjects), we used Equation 1 (see Materials and Methods) to calculate power to detect SNPs explaining specific magnitudes of variance () for warfarin dose (see Table 3). The table shows that power to achieve genome-wide significance (p<1.5×10−7) is essentially 100% for VKORC1 rs9923231 (), CYP2C9*3 () and CYP2C9*2 (), but power falls to ∼48% for CYP4F2 rs2108622 (). The table also shows that when CYP4F2 is added to the multivariate model, a SNP accounting for 1.5% or 1.0% of the dose variance would have ∼82% or ∼41% power of being detected, respectively. Therefore we estimate that our GWAS had at least 80% power to detect warfarin-associated variants explaining at least 1.5% of the dose variance but 40% or less power to detect genome-wide significance if a variant accounts for 1% or less dose variance.
However it is important to emphasize that these power estimates assume that the dose-altering DNA variant is genotyped and tested directly or is indirectly detected through a marker in sufficiently high LD to the dose variant that the marker's magnitude is detectable (Table 3). The assumption of directly testing the dose-altering variant is accurate for CYP2C9*2 and *3 which are each known to alter warfarin metabolism [15],[16] and is likely correct for CYP4F2 rs2108622 which, like *2 and *3, changes protein coding sequence. However, to explore whether other dose-altering variants might be undetected due to insufficient LD with genotyped GWAS SNPs, we determined the relationship between the variance observed at a marker () and at the causative variant () assuming pairwise LD of r2 between the two polymorphisms (see Materials and Methods: How Much Does Linkage Disequilibrium Attenuate Association with a Quantitative Trait?). The relationship is given by Equation 3 in Materials and Methods () which is analogous to Pritchard and Prezworski's relationship () for the number of cases () providing equal power in a case-control study that tests either the disease-causing SNP or a nearby marker [17]. To use the equation to estimate magnitudes for variants that might be undetected by our GWAS, we note that ∼90% of the GWAS SNPs had a minor allele frequency (MAF) above 10% in our warfarin subjects implying that a “rare” dose-altering variant (MAF≈1%–5%) would be covered at a likely maximum r2 of only ∼0.1 to ∼0.5. This low r2 coverage implies that rare variants could have values (0.05 to 0.02) easily detected by regression testing of the variant itself, but unlikely to be detected through a GWAS marker since maximum values could drop to 0.01 or much lower (see Equation 3 and Table 3). By contrast, “common” SNPs (MAF≥5%), which might also be dose variants, are covered by GWAS SNPs of this study at reasonably high r2 values in most instances (r2>0.8 or r2>0.5 for ∼60% or ∼80% respectively of common SNPs [18] and r2>0.9 for ∼90% of non-synonymous common SNPs [19] in HapMap Caucasians). We therefore conclude that our GWAS probably detected most common SNP variants explaining 1.5% or more of the warfarin dose variance, but may have failed to detect rarer variants that could individually explain up to 5% of dose variance. We further note that the HumanCNV370 array used in this study does not have the required marker complement to undertake a comprehensive GWAS of common CNVs.
As noted in the Introduction, the widely replicated warfarin dose associations with VKORC1 and CYP2C9 represent one of the most successful applications of pharmacogenetics to date. Our study together with that of Caldwell et al. [12] now also clearly demonstrates that CYP4F2 (rs2108622) is a third gene that influences warfarin dose, but our GWAS and statistical analysis also implies that additional common SNP variants that influence dose may not exist in Caucasian populations. However, Caucasians might carry common variants with effects smaller than CYP4F2 or rare variants whose effects are substantially larger than the ∼1% of dose variance explained by CYP4F2. Furthermore, other unidentified genes may influence warfarin dose in other ethnicities such as Asians or Africans, and some rare dose-altering variants in known genes such as VKORC1 may exist in only a subset of populations of European descent [20]. Hence, future research could address ethnic differences in the genetic variants that influence warfarin dose as well as subtle intra-ethnic differences and admixture that may exist in European or other populations.
In a recent study [3], we highlighted the potential benefit of pre-treatment forecasting of required warfarin dose based on patient genotypes at VKORC1 and CYP2C9 together with non-genetic predictors of dose. Indeed, in August 2007, the US Food and Drug Administration (FDA) updated warfarin labeling to recommend initiating lower warfarin dose in some patients based on VKORC1 and CYP2C9 genotypes. However this recommendation is not a requirement due to a lack of large trials demonstrating warfarin patient benefit from dose forecasting (though two small trials [21],[22] do support such benefit; see also [23]–[27] for reviews and other trials). The results of our GWAS provide further impetus for conducting large-scale dose-forecasting trials by identifying CYP4F2 as a third genetic predictor of dose and also by showing that additional major genetic predictors may not exist in Caucasians or may not emerge in the near-term. Hence, large-scale trials of patient benefit from dose forecasting based on VKORC1 and CYP2C9 (with possible inclusion of CYP4F2 as a minor predictor) are likely to provide state-of-the-art clinical benchmarks for warfarin use during the foreseeable future.
The study subjects were 1053 Swedish patients collected for the WARG study [3] (http://www.druggene.org/). This is a multi-centre study of warfarin bleeding complications and response to warfarin treatment [28]. Anticoagulant response is measured by INR, which is the ratio of the time required for a patient's blood to coagulate relative to that of a reference sample. By titrating warfarin dose, physicians aim for a therapeutic INR reading between 2.0 and 3.0; thus the primary quantitative outcome for the GWAS was the mean warfarin dose (mg/week) given to a patient during a minimum series of three consecutive INR measurements between 2 and 3 [3]. As a secondary GWAS outcome, we also catalogued each patient for the occurrence or non-occurrence of “over-anticoagulation” during the first 5 weeks of treatment (defined as an INR reading above 4.0) and tested for genetic association which adjusted for the treatment day (1 to 35) of the over-anticoagulation event (see “Association testing” below). The clinical data collected by the WARG protocol included gender and age since each is a known non-genetic predictor of warfarin dose but did not include bodyweight and dietary information (e.g. vitamin K intake). Regression analysis of prescribed medication which can potentiate or inhibit warfarin action was not a statistically significant predictor of warfarin dose in the 1053 WARG GWAS subjects and hence was not included as a predictor variable in the multivariate regression analyses. The WARG study samples were previously described elsewhere [3],[4],[28],[29] as were the Uppsala followup samples [8]. The WARG and Uppsala studies received ethical approval from the Ethics Committee of the Karolinska Institute and the Research Ethics Committee at Uppsala University, respectively.
From approximately 1500 WARG samples [3] examined for non-degradation and appropriate concentration of DNA (∼50 ng/µl), we randomly selected 1208 subjects for genotyping SNPs and CNV probes using the HumanCNV370 BeadChip array (Illumina). We excluded SNPs with MAF below 1%, call rate below 95%, or if call rate fell below 99% when MAF was below 5%. SNPs that departed from Hardy-Weinberg equilibrium (P<10−6) were also excluded. Subjects with genotyping call rate below 95% were also eliminated. Using iPLEX (Sequenom), subject identity (and associated phenotypic data) was cross-checked by genotyping four gender markers and 47 SNPs also carried on the HumanCNV370 array, enabling us to exclude ∼136 misidentified subjects. Sample quality (contamination) was further assessed by plotting each subject's genome-wide heterozygosity and eliminating outliers (with heterozygosity above or below the range of 0.312–0.372). After these quality control steps, a total of 1053 warfarin patients and 325,997 GWAS SNPs were retained for analysis. The GWAS SNPs included two SNPs not on the HumanCNV370 array but which are highly predictive of warfarin dose [rs9923231 (VKORC1) and rs1799853 (CYP2C9*2)] which we genotyped by TaqMan assay (Applied Biosystems).
Although we retained 325,997 GWAS SNPs for association testing of SNPs, it should be noted that all ∼370,000 probes on the Human CNV370 array were used to define CNVs. Log R ratio values of probes were output from the BeadStudio software [30]. A loess correction was applied to each sample to remove local correlations or genomic wave [31]. The resultant genomic copy number profiles were then segmented using Circular Binary Segmentation [32]. Some samples displayed abnormally high numbers of segments indicating problems in DNA quantity or quality or hybridization. Samples were removed until the number of segments across all samples was approximately normal. Using this technique, 143 (14%) of samples were flagged as problematic. These samples were excluded when CNV regions were defined but included for association testing. Putative CNV were defined from segments by applying a threshold on the segment log R ratio. This threshold was asymmetric allowing for a differing response for deletions and duplications. The central peak of the segment log R ratio distribution was fitted and the threshold values obtained by taking values at ±5 standard deviations from the centre.
In order to define regions for association testing, merging of CNV across samples was performed. This was achieved by merging two putative CNV into a region if there was greater than 40% reciprocal overlap. This procedure defined 2530 CNV regions in total. Of these, most were singletons (54%) or low frequency, <3% (93%), while 820 (70%) of the non-singleton regions overlapped CNVs from the Database of Genomic Variants [33]. We tested all 2530 CNVs for association, because a CNV discovered as a “singleton” might well include multiple copies of a rare CNV allele in the study samples.
At each SNP, genotypes were coded 0, 1 or 2 and the SNP was tested for association with the square root of warfarin dose [8] by either univariate or multivariate linear regression analysis conducted in PLINK [34] (http://pngu.mgh.harvard.edu/~purcell/plink/) or in R software (http://www.r-project.org/). We used the same regression analysis to test association with all HapMap SNPs not on the HumanCNV370 array by imputing ∼2.2 million SNPs using Beagle software [35] trained from genotypes of the 60 HapMap CEU parents [36]. We excluded SNPs whose imputed MAF was below 5% or differed by more than 5% with MAF of the CEU parents.
We also tested haplotypes for association with warfarin dose by two approaches: (1) each subject's warfarin dose residual (difference between observed and predicted dose based on the full multivariate regression model containing CYP4F2) was considered a quantitative trait value and tested for association with haplotypes defined across the genome in sliding windows of 2, 3 or 4 consecutive SNPs as implemented by PLINK software; (2) by scanning GWAS genotypes, Beagle software groups genetically related haplotypes into clusters which it then resolves into diallelic (SNP-like) “pseudo-markers” optimized for detecting phenotypic association. To test haplotypes, we evaluated the pseudo-marker genotypes of warfarin patients at 1.97 million pseudo-markers covering the genome by testing each pseudo-marker in the same multivariate regression framework used to test individual SNPs (as described in the preceding paragraph).
We tested for statistical interaction in modulating warfarin dose for each pair of established dose predictors (VKORC1 rs9923231, CYP2C9*2 and *3, CYP4F2 rs2108622, Age, Sex) using multivariate regression and R software as described above. An interaction term formed by multiplying the pair of predictor variables was added to the multivariate regression equation which contained only main effects of the 6 predictors, and standard ANOVA compared this main-effect model with the enhanced interaction model by testing for a statistically significant increase in explained dose variance. Interaction test p-values were considered statistically significant if below the Bonferroni cutpoint determined by correcting for the 15 interaction tests (i.e. p<0.0033≈0.05/15).
To test for association with over-anticoagulation (INR>4.0) during treatment days 1–35, we performed Cox proportional hazard regression on survival time (day of over-anticoagulation) using the survival library of R software. The GWAS data set of 1053 WARG subjects contained 215 subjects whose INR exceeded 4.0 during days 1–35 while the entire dataset of 1489 WARG subjects contained 312 such subjects.
For each CNV locus, association was tested with square root of warfarin dose by multivariate regression analysis in which subject copy number intensity was the CNV predictor of dose. This analysis differs from association testing with SNP genotypes since the two CNV alleles on homologous chromosomes generate one copy number intensity rather than a separate allele for each chromosome. As a QC strategy, we determined each subject's rank in the dataset for copy number intensity at each CNV on chromosome 17. This enabled us to differentiate the majority of subjects (whose individual distribution of ranks were approximately random and uniform) from 174 obvious outliers due to poor quality DNA (whose ranking distributions were “U-shaped” since their intensities strongly clustered at both high and low ranks). These 174 subjects were excluded from the primary CNV association analysis (with further confirmation of lower quality DNA for these subjects being their rough correspondence to the subjects with lower (<99%) SNP call rates). However, we also crosschecked the primary CNV analysis by conducting association testing on the dataset without excluding the 174 subjects and found no statistically significant association with warfarin dose at any CNV whether the dataset excluded or included the subjects. Association testing of the CNVs was executed using R software [37].
For the replication of CYP4F2 rs2108622, we genotyped a panel of 588 warfarin patients consisting of 410 subjects from the WARG cohort [3] and 178 from the Uppsala cohort [38]. Table 2 shows regression on this pooled sample of 588 subjects. Separate results for each of the two panels are given in Table S4.
To possibly identify SNPs with genuine but weak associations to warfarin dose, we excluded VKORC1, CYP2C9, CYP4F2 and identified 40 other GWAS loci for follow-up genotyping exhibiting multivariate regression p-values below 0.0002, and selected 40 SNPs representing these loci for genotyping. Only genotyped (not imputed) SNPs were chosen for follow-up. We genotyped the same 558 patients as in the CYP4F2 replication using the iPLEX MassARRAY.
Suppose multiple regression analysis is conducted in N total samples by testing a SNP with coefficient of determination (i.e., explained variance) R2test after adjustment for known predictors whose total of coefficient of determination is R2knw. The probability (power) to detect the tested SNP at a significance level α equals:(1)where F′(1, N–2, θ2) is the probability density function for an F distribution with 1, N-2 degrees of freedom and non-centrality parameter θ2 (Section 28.28 in [39], Example 8.4 in [40]). Here the constant c satisfies the equation:(2)where α is the significance level, and F(1, N–2) is the probability density function for a F-distribution of degree of freedom one and N–2.
Association with a quantitative trait (QT) becomes weaker for a marker SNP in LD with a SNP that alters the QT, and hence the association becomes more difficult to detect at the marker than at the QT-altering SNP. Here we quantify the LD attenuation for a QT when testing for association by linear regression (which includes the Cochran-Armitage trend test for dichotomous traits), and we obtain a result analogous to the LD attenuation for the Pearson Chi-square test for allelic association to dichotomous traits as in cases and controls [17]. If a causative QT-altering SNP has a coefficient of determination (i.e., explained variance) and is in pairwise LD of r2 with a marker SNP, then the coefficient of determination for the marker SNP () is approximated by:(3)
In other words, when testing a marker, the proportion of explained variance decreases by a factor of r2.
To begin the proof of Equation 3, let the QT be represented by the random variable “q”, and let “m” and “x” be SNP genotypes (coded 0, 1, or 2) representing the marker and causative (QT-altering) SNP, respectively. The coefficients of determination are equal to the square of two correlation coefficients (denoted by “Corr”) measuring the correlation of m or x with q:(4)(5)
Also note that correlation between genotypes at the marker and causative SNP is given by another correlation coefficient:(6)
It is well known that the partial correlation coefficient of m and q conditioned on x is (equation 16.20, p. 649 in [41]):(7)
However, conditional on genotype at the causative SNP, marker m and the QT q would be uncorrelated (assuming m is not in LD with a second causative polymorphism) and thus the numerator of Equation 7 would be zero implying that:(8)
Based on prior work [42]–[44], we show in Text S1 that the squares of the genotypic correlation coefficient and LD correlation coefficient r2 are approximately equal if the population is in Hardy-Weinberg equilibrium. Therefore, substituting r2 for in Equation 8 gives Equation 3.
|
10.1371/journal.pmed.1002691 | Machine learning to identify pairwise interactions between specific IgE antibodies and their association with asthma: A cross-sectional analysis within a population-based birth cohort | The relationship between allergic sensitisation and asthma is complex; the data about the strength of this association are conflicting. We propose that the discrepancies arise in part because allergic sensitisation may not be a single entity (as considered conventionally) but a collection of several different classes of sensitisation. We hypothesise that pairings between immunoglobulin E (IgE) antibodies to individual allergenic molecules (components), rather than IgE responses to ‘informative’ molecules, are associated with increased risk of asthma.
In a cross-sectional analysis among 461 children aged 11 years participating in a population-based birth cohort, we measured serum-specific IgE responses to 112 allergen components using a multiplex array (ImmunoCAP Immuno‑Solid phase Allergy Chip [ISAC]). We characterised sensitivity to 44 active components (specific immunoglobulin E [sIgE] > 0.30 units in at least 5% of children) among the 213 (46.2%) participants sensitised to at least one of these 44 components. We adopted several machine learning methodologies that offer a powerful framework to investigate the highly complex sIgE–asthma relationship. Firstly, we applied network analysis and hierarchical clustering (HC) to explore the connectivity structure of component-specific IgEs and identify clusters of component-specific sensitisation (‘component clusters’). Of the 44 components included in the model, 33 grouped in seven clusters (C.sIgE-1–7), and the remaining 11 formed singleton clusters. Cluster membership mapped closely to the structural homology of proteins and/or their biological source. Components in the pathogenesis-related (PR)-10 proteins cluster (C.sIgE-5) were central to the network and mediated connections between components from grass (C.sIgE-4), trees (C.sIgE-6), and profilin clusters (C.sIgE-7) with those in mite (C.sIgE-1), lipocalins (C.sIgE-3), and peanut clusters (C.sIgE-2). We then used HC to identify four common ‘sensitisation clusters’ among study participants: (1) multiple sensitisation (sIgE to multiple components across all seven component clusters and singleton components), (2) predominantly dust mite sensitisation (IgE responses mainly to components from C.sIgE-1), (3) predominantly grass and tree sensitisation (sIgE to multiple components across C.sIgE-4–7), and (4) lower-grade sensitisation. We used a bipartite network to explore the relationship between component clusters, sensitisation clusters, and asthma, and the joint density-based nonparametric differential interaction network analysis and classification (JDINAC) to test whether pairwise interactions of component-specific IgEs are associated with asthma. JDINAC with pairwise interactions provided a good balance between sensitivity (0.84) and specificity (0.87), and outperformed penalised logistic regression with individual sIgE components in predicting asthma, with an area under the curve (AUC) of 0.94, compared with 0.73. We then inferred the differential network of pairwise component-specific IgE interactions, which demonstrated that 18 pairs of components predicted asthma. These findings were confirmed in an independent sample of children aged 8 years who participated in the same birth cohort but did not have component-resolved diagnostics (CRD) data at age 11 years. The main limitation of our study was the exclusion of potentially important allergens caused by both the ISAC chip resolution as well as the filtering step. Clustering and the network analyses might have provided different solutions if additional components had been available.
Interactions between pairs of sIgE components are associated with increased risk of asthma and may provide the basis for designing diagnostic tools for asthma.
| The relationship between allergic sensitisation and asthma is complex.
Asthma prediction models based on the IgE responses to the whole allergen extracts exhibit relatively poor performance.
This study examines the relationship between IgE responses to multiple allergen components in component-resolved diagnostics (CRD) and their associations with asthma.
Serum-specific IgE responses to 112 allergen components were measured using a multiplex array among children in a population-based birth cohort.
Researchers applied network analysis and hierarchical clustering (HC) to explore the connectivity structure of component-specific IgEs and identified seven clusters of component-specific sensitisation. Cluster membership mapped closely to the structural homology of proteins and/or their biological source.
HC identified four ‘sensitisation clusters’ among study participants.
The relationship between component clusters, sensitisation clusters, and asthma was explored using a bipartite network.
The differential network of pairwise component-specific IgE interactions was inferred, which demonstrated that interactions among 18 pairs of allergen components predicted asthma with a good balance between sensitivity and specificity. For example, children with IgE antibodies to different allergenic proteins from both dog and cat, or horse and house dust mite, are at higher risk of developing asthma.
IgE responses to multiple allergenic proteins are functionally coordinated and co-regulated.
Pairwise interactions within this complex network predict clinical phenotypes. Interactions between pairs of sIgE components are associated with increased risk of asthma and provide the basis for designing diagnostic tools for asthma.
| Asthma is the most common noncommunicable disease in childhood. Over recent decades, a large body of evidence has demonstrated a close relationship between specific immunoglobulin E (sIgE) antibody responses and asthma [1, 2], but the data about the strength of this association are conflicting [2, 3]. Furthermore, in a clinical situation, confirmation of allergic sensitisation using standard diagnostic tests (skin prick tests [SPTs] and/or measurement of sIgE) does not necessarily indicate that patient's symptoms are caused by an allergic reaction [1]. We have previously proposed that these inconsistencies are in part consequent to ‘allergic sensitisation’ not being a single entity (as considered conventionally) but an umbrella term for a collection of several different classes of sensitisation that differ in their association with asthma and other allergic diseases. To test this, in a previous study we applied a machine learning approach with Bayesian inference to a comprehensive set of skin tests and sIgE data to whole allergen extracts collected from infancy to school age in a population-based birth cohort [4]. Children clustered into four distinct sensitisation classes characterised by different patterns of responses to specific allergens and the time of onset of sensitisation [4]. The risk of asthma was increased almost 30-fold amongst children belonging to one of these classes (assigned as ‘Multiple early sensitisation’, comprising less than one third of children diagnosed as sensitised using conventional definitions). We have replicated these findings in another birth cohort [5] and have shown that diminished lung function in adolescence and early adulthood is associated with ‘Multiple early’, but not other sensitisation classes [6, 7].
In food allergy, there is increasing evidence that sensitisation to some, but not all, allergenic proteins in allergen extracts is important for making a distinction between true allergy and asymptomatic sensitisation [8]. For example, we have shown that immunoglobulin E (IgE) response to peanut protein Ara h 2 is much more predictive of true peanut allergy than standard tests using whole allergen extract [9, 10]. Measuring sensitisation to these individual molecules (referred to as allergen components) using component-resolved diagnostics (CRD) may be more informative than standard tests in respiratory allergy, as well. The developments in molecular diagnostics have led to products such as the multiplex Immuno Solid-phase Allergen Chip (ImmunoCAP ISAC), in which sIgE to more than 100 allergen components can be measured simultaneously [11]. Using a machine learning approach, we have shown that patterns of component-specific IgE responses in this multiplex assay have reasonable discrimination ability for asthma and rhino-conjunctivitis [12]. In a further study using latent variable modelling, we identified several cross-sectional clusters of IgE responses in school age children, and each of these clusters was associated with different clinical symptoms [13]. Our subsequent study using nested latent class probabilistic modelling has indicated that longitudinal trajectories of sensitisation to several grass and house dust mite (HDM) allergens during childhood had different associations with clinical outcomes [14].
Based on these findings, we propose (1) that the impact of allergic sensitisation on asthma is a complex phenomenon that cannot be captured by considering individual allergen sIgE responses separately, or in isolation; and (2) that sIgE responses to multiple allergenic proteins are functionally coordinated and co-regulated, and this complex network of interactions foreshadows asthma development. Specifically, we hypothesise that interaction patterns between component-specific IgE antibodies rather than individual IgE responses to ‘informative’ components are associated with risk of asthma. To address our hypothesis, we measured sIgEs to 112 allergen components using a commercially available multiplex array among participants in a population-based birth cohort, and we used unsupervised machine learning techniques to explore how component-specific IgEs interact with each other and to identify common sensitisation profiles among children. We then used a supervised machine learning approach to explore interactions of component-specific IgEs in relation to asthma.
The Manchester Asthma and Allergy Study is a population-based birth cohort [15]. Participating families were recruited from the maternity catchment area of Wythenshawe and Stepping Hill Hospitals in South Manchester and Cheshire, United Kingdom [15]. All pregnant women were screened for eligibility at antenatal visits (8th–10th week of pregnancy) between 1 October 1995 and 1 July 1997. Of the 1,499 women and their partners who met the inclusion criteria, 288 declined to take part in the study, and 27 were lost to follow-up between recruitment and childbirth. The study was approved by the Research Ethics Committee and parents gave written informed consent.
Children attended review clinics at ages 1, 3, 5, 8, 11, and 16 years. Validated questionnaires were interviewer administered to determine parentally reported history of wheeze, eczema, and rhinitis, and treatments received. SPT was used to ascertain atopic sensitisation to common inhalant and food allergens, and lung function measurements were obtained using spirometry at all visits from age 5 years. A blood sample was collected in children who gave assent for venepuncture [16]. Primary care medical records were examined and data including wheeze episodes, prescriptions of asthma medications and oral corticosteroid, and hospitalisations were extracted.
In this study, we performed a cross-sectional analysis using data collected at age 11 years.
‘Current wheeze’ was defined as a positive answer to the question, ‘Has your child had wheezing or whistling in the chest in the last 12 months?’ [17] ‘Current asthma’ was defined as a positive answer to two out of three of: ‘Has the doctor ever told you that your child had asthma?’; ‘Has your child had wheezing or whistling in the chest in the last 12 months?’; and ‘Has your child had asthma treatment in the last 12 months?’ [18]. Further details of follow-up and definitions of clinical outcomes are presented in the supplementary appendix (S1 Appendix).
We measured sIgE to 112 allergenic molecules using ImmunoCAP ISAC (Thermo Fisher Scientific-Phadia AB, Uppsala, Sweden) at the follow-up at age 11 years. The level of component-specific IgE antibodies was reported in ISAC Standardised Units (ISU). To ascertain co-occurring sensitisations among participants, we dichotomised IgE data according to the manufacturer's guidelines, using a binary threshold (positive>0.30 ISU). To evaluate the differential connectivity structure of component-specific IgEs, we used continuous raw values.
In this cross-sectional analysis, we included all children with available CDR data. We analysed data for components with sIgE>0.30 ISU in at least 5% of children (active components) and among participants with at least one active component sIgE>0.30 ISU (filtering) [19]. A flowchart describing the analysis steps involved in this study is presented in S1 Fig.
We investigated patterns of sIgE co-expression using hierarchical clustering (HC), which transforms a distance matrix into a nested series of partitions that can be represented through a treelike graph (dendogram). By exploring this graph, one can obtain useful information on the hierarchy of the clusters and their similarities. At the lowest level of the hierarchy, each cluster contains a single observation. At the highest level, there is only one cluster containing all of the data. HC algorithms can follow an agglomerative or a divisive approach. Agglomerative strategies start at the bottom and at each level recursively merge a selected pair of clusters into a single cluster. This produces a grouping at the next higher level with one fewer cluster. The pair chosen for merging consist of the two groups with the smallest intergroup dissimilarity. Divisive methods start at the top and at each level recursively split one of the existing clusters at that level into two new clusters. The split is chosen to produce two new groups with the largest between-group dissimilarity. With both paradigms there are N−1 levels in the hierarchy [20]. In our analysis, we used the agglomerative procedure combined with the average linkage method, which defines the distance between two clusters as the average distance between each point in one cluster to every point in the other cluster.
Compared with partitional clustering, HC techniques do not require one to fix the number of clusters a priori, can find different levels of similarity between the sIgE components within the hierarchy of clusters, and, hence, can highlight different patterns of connectivity and biological properties.
Distances between sIgE components were expressed by means of the distance correlation matrix [21]. The advantage of using distance correlation is that it is capable of detecting nonlinear relationships. We then used network analysis to visualise the connectivity structure of sIgEs.
Final partitions can significantly differ according to the chosen clustering approach. Hence, to evaluate the robustness of our findings, we compared the retrieved clusters with partitions obtained through a divisive HC procedure and a partitional clustering technique using the Rand index [22].
To identify patterns of sensitisation among children, we used an HC approach combined with Ward's linkage [23] and the Jaccard distance between binary responses to sIgE profiles. At each iteration of the clustering algorithm, the Ward's method joins the clusters so that the total within-cluster variance is minimised. Ward's linkage is conservative, monotone, correctly infers the hidden structure within the data, and often outperforms the other approaches [24, 25]. We used χ2 and Kruskal–Wallis tests to evaluate the associations between the identified clusters and clinical outcomes.
We used a bipartite network to visually explore the relationship between component clusters, sensitisation clusters, and asthma. We investigated whether sIgE to individual components is associated with the risk of asthma using a penalised logistic regression model. To test the hypothesis that pairwise interactions of component-specific IgEs are associated with asthma, we used joint density-based nonparametric differential interaction network analysis and classification (JDINAC) [26]. We utilised this recently developed nonparametric model to identify differential interaction patterns of network activation of sIgEs that are most closely related to asthma, and to build a classification model using the network biomarkers. JDINAC has the advantage of capturing nonlinear relations between component-specific IgEs without the need for parametric assumption on their probability distribution.
The main assumption of the JDINAC model is that network-level difference between children who have asthma and children who do not have asthma arises from the collective effect of differential pairwise component IgE interactions. Here, the interactions are characterised by the conditional joint density of pairs of component-specific IgEs [26], estimated through a nonparametric kernel method. Formally, let Xn×p be the data matrix of n individuals and p sIgE allergens. Hence, Xl, l = 1,…,n, represents the level of sIgEs in the l-th child. Let Yl denote the binary variable defined as follows:
Yl={0iflisnon−asthmatic1ifotherwise
Let P denote the probability of having asthma, P = Pr(Yl = 1), and Gi denote the i-th sIgE. Then, JDINAC logistic regression-based approach can be exploited to test the model:
logit(P)=α0+∑i=1p∑j>iβijlnfij1(Gi,Gj)fij0(Gi,Gj),s.t.∑i=1p∑j>iβij≤c,c>0
where fij1(Gi,Gj) and fij0(Gi,Gj) denote the class conditional joint density of Gi and Gj for class 1 and class 0, respectively. The conditional joint densities fij1(Gi,Gj) indicate the strength of association between Gi and Gj in class 1, and parameters βij indicate differential dependency patterns between condition-specific groups [26]. The estimation procedure is based on a multiple splitting and prediction averaging procedure, which guarantees robust and accurate results. The data are split in two parts. On the first part, joint kernel density functions, f^ij1 and f^ij0, are estimated, while on the second part, L1 penalised logistic regression is fitted. The procedure is repeated for a predefined number of iterations (for estimation details and algorithm, see [26]).
To ensure robustness of the results, we ran both models with 10-fold cross validation in 50 independent repetitions. To reduce the effect of imbalanced data, we included class weight in both models. sIgE raw values were log-transformed (log(x+1)) prior to these analyses.
To evaluate the robustness of our results and provide external narrow validation [27], we repeated the analysis among cohort participants who had ISAC CRD data at age 8 years, excluding the children whose data were used in the primary analysis at age 11 years. For children in the validation step, both CDR data and clinical outcomes were ascertained at age 8 years.
All statistical analyses were run in the programming language R [28]. Distance correlation was computed with the package energy [29]. JDINAC scripts were made available by the authors [26] at https://github.com/jijiadong/JDINAC. We used igraph package for network visualisations [30], epitools to estimate the odds ratio (OR) [31], clValid to compute internal validity measures for HC [32], and caret to infer the penalised logistic regression model [33].
Among 1,184 children born into the cohort, 822 attended clinical follow-up at age 11 years. CRD data were obtained for 461 (56.1%) children. Demographics of these 461 participants are presented in S1 Table; we have also previously reported that there were no significant differences in demographic characteristics or outcomes between cohort members with and without CRD [13]. Of 461 children with CRD, 221 (47.9%) had positive sIgE to at least one of the 112 allergen components [13], and 94 (20.4%) had current asthma. After filtering [19], 44/112 allergen components were active; 213 (46.2%) children had at least one of the active component IgEs >0.30 ISU, 73 (34.3%) of whom had asthma. The list of components that were inactive [19] and the proportion of children who had positive sIgE to these ‘rare’ components are presented in S2 Table.
There was a significant difference in the total number of positive component-specific IgEs between children who have asthma and children who do not have asthma, with children who have asthma responding to more allergens than children who do not have asthma (median 11 [IQR: 6–18] versus 6 [IQR: 3–10 ], p<0.001, S2 Fig). The responses to individual components stratified by disease status did not show considerable differences between sensitised children with and without asthma (Fig 1). However, we highlight an increase in the positive responses to some allergenic proteins among children who have asthma, particularly group 2 HDM components and furry animal lipocalins (S3 Table).
Of the 44 allergen components included in the model, 33 grouped in seven component clusters (C.sIgE-1–7), while the remaining 11 formed singleton clusters (Table 1). The number of clusters was determined by fixing the threshold for the dissimilarity measure (1−distance correlation) equal to 0.40, which ensured high similarity between the components.
We compared the adopted model with the divisive HC clustering DIANA (Divise Analysis) [34], and the partition around medoids (PAM) [34] algorithm. The Rand index, 0.99 for DIANA and 0.98 for PAM, suggested that the obtained groups were stable and robust. Internal validity indices also showed that cluster membership was very stable (S3 Fig).
C.sIgE-1 was composed exclusively of HDM components (Group 1 and 2 HDM allergens); C.sIgE-2 of peanut components associated with true peanut allergy (2S albumins and 7S globulin) [9]; C.sIgE-3 of lipocalins from cat, dog, horse, and mouse; C.sIgE-4 of grass components; C.sIgE-5 of PR-10 proteins from various sources; C.sIgE-6 of tree allergens; and C.sIgE-7 of profilins. The HC highlighted the structural relationships of the allergen components within protein families.
The co-expression network in Fig 2 shows the interactions and underlying connectivity structure of component-specific IgEs. The connectivity expresses how sIgE components are correlated and co-regulated with each other. Components belonging to the PR-10 (C.sIgE-5) cluster were central to the network, showing higher connectivity than other components; components in this cluster seem to mediate connections between components from grass (C.sIgE-4), tree (C.sIgE-6), and profilin (C.sIgE-7) clusters with components in HDM (C.sIgE-1), lipocalins (C.sIgE-3), and peanut clusters (C.sIgE-2). Alt a 1 and Blo t 5 were weakly connected to other component-specific IgEs. Components in the HDM cluster showed high intraclass connectivity.
The structure of sensitisation profiles among study participants was inferred in a completely unsupervised manner, with the optimal solution suggesting four sensitisation clusters (based on the Calinski-Harabasz criterion [35]). Cluster membership was stable (S4 Fig). In the model comparisons, the Rand index showed moderate agreement with the partition obtained with DIANA (0.53) and good agreement with the partition obtained with PAM (0.79).
After visual inspection of the patterns (Fig 3), we labelled these four sensitisation profiles as (1) Multiple sensitisation, with positive sIgE to multiple components across all seven component clusters (C.sIgE-1–7) and singleton components; (2) Predominantly HDM sensitisation, with IgE responses mainly to components from C.sIgE-1; (3) Predominantly grass and tree sensitisation, with positive sIgE to multiple components across C.sIgE-4–7; and (4) Lower-grade sensitisation.
Association with clinical outcomes (asthma, rhinitis, and atopic dermatitis [AD]) differed for different sensitisation profiles (S4 Table, S5 Table). Children in the HDM cluster were more likely to have asthma (OR: 4.44; 95% CI: 1.72–11.46; p = 0.002) and wheeze (OR: 7.31; 95% CI: 2.74–19.48; p < 0.001), but not rhinitis or AD, while those in the grasses/trees cluster were more likely to have rhinitis (OR: 6.62; 95% CI: 2.84–15.40; p < 0.001). Membership of the Multiple sensitisation cluster was associated with the highest risk of asthma (OR: 4.97; 95% CI: 1.99–12.34; p < 0.001) and a high risk of wheeze (OR: 4.41; 95% CI: 1.70–11.41; p < 0.001) and rhinitis (OR: 6.18; 95% CI: 2.71–14.12; p < 0.001) (S5 Table). No significant associations were found with lung function measurements (S6 Table).
Fig 4 summarises the relationship between sensitisation clusters and asthma, and the connectivity with component-specific IgEs and component clusters. Although a significantly higher proportion of children with asthma was found in the Multiple sensitisation and HDM clusters, the majority of children in each of the sensitisation clusters did not have asthma. All clusters shared similar connection to some component clusters (C.sIgE-3 and C.sIgE-4), but we observed distinct patterns of connectivity between the cluster with a higher proportion of children with asthma compared with those with a higher proportion of children who did not have asthma. Specifically, only children in Multiple sensitisation and Predominantly HDM clusters were strongly connected to the allergens in C.sIgE-1, while children in Predominantly grasses/trees and Lower-grade sensitisation clusters were distinctively connected to C-sIgE-2.
S5 Fig shows examples of bipartite subnetworks of a subset of component clusters. Panel A shows the connectivity between a set of informative components in the lipocalin cluster (C.sIgE-3) with Fel d 1. The analysis has shown that children with connection to only one sIgE were not at higher risk of asthma, but those who were connected to two or more components were at increased risk of having asthma. Similar behaviours are observed for all the other networks, apart from interactions involving the grass IgE cluster (C.sIgE-4).
To investigate whether individual components sIgE or pairwise interactions of component-specific IgEs are stronger associates of asthma, we compared the performances of penalised logistic regression and JDINAC in classifying asthma (Table 2). In the multivariate logistic regression model, we include all the 44 individual components as predictors. To improve comparability between the two models, a penalty on the L1-norm was included in the logistic model.
Penalised logistic regression with individual components had poor performance, with low sensitivity (0.60) and moderate specificity (0.70). It did not provide an efficient classification rule. In contrast, JDINAC provided a good balance between sensitivity (0.84) and specificity (0.87). Results from 10-fold cross validation in 50 independent repetitions on the whole data set showed that JDINAC with pairwise interaction outperformed penalised logistic regression with individual components, with area under the curve (AUC) equal to 0.94, compared with 0.73 (Fig 5).
These results suggest that the interactions between pairs of sIgE are more informative than the individual components in asthma classification.
We then proceeded to infer the differential network of pairwise component-specific IgE interactions that predict asthma by connecting the sIgEs pairs with high differential dependency weights (defined as the number of repetitions in which β^ij≠0). A total of 18 pairs of component-specific IgEs exhibited a significant differential interaction between children who have asthma and children who do not have asthma (Fig 6). The network emphasises multisource connections. HDM and animal components, which were central to the network, showed higher connectivity than other components. The interactions between the grass-related sIgEs (Phl p 2 and Phl p 12) and between Lep d 2 and Fel d 1 were linked to a healthy state. In contrast, the remaining pairwise interactions were linked to asthma. The connections between Fel d 1 and Can f 1, Der p 1 and Equ c 1, and Der f 2 and Der p 1 had a strong impact on the prediction results because of the higher differential weights.
Our study suggests that the relationship between allergic sensitisation and asthma is complex and cannot be fully captured or explained by considering sIgE responses to any individual allergenic molecule(s). In contrast to IgE-mediated food allergy, in which sensitisation to a limited number of ‘informative’ allergenic proteins differentiates between true food allergy and asymptomatic sensitisation (such as Ara h 2 in peanut allergy) [9], we did not identify such ‘informative’ component(s) as a hallmark of an increased risk of asthma. By clustering component-specific IgE responses only (i.e., not the children), we identified seven clusters of component-specific sensitisation, with cluster membership mapped closely to the structural homology of proteins and their biological source. By clustering study participants, we identified four sensitisation clusters that were characterised by unique patterns of sensitisation to allergenic molecules from different component clusters. In this study, the analysis of the relationship between component clusters, sensitisation clusters, and asthma revealed that the key associate of asthma was the interaction between component-specific IgEs, indicating that the important feature of IgE response linked to an increased risk of asthma is not individual IgE to any informative component(s), but the pattern of interactions between component-specific IgEs. Further analyses revealed a differential network of pairwise interactions between a limited number of component-specific IgEs from different component clusters, which predicted asthma with a good balance between sensitivity and specificity. In this study, we found that amongst sensitised children, some of these connectivities were associated with an increased risk of asthma (e.g., between Fel d 1 and Can f 1, Der p 1 and Equ c 1), while others decreased the risk (e.g., between sIgEs to grass components Phl p 1 and Phl p 5).
One of the limitation of our study is that there may be a number of potentially important allergens that are not included on the ISAC chip (e.g., those from fungi), and it is possible that the clustering would provide different solutions if additional components had been available [13]. We acknowledge that our analysis identified only pairwise interactions, and that the relation between asthma and the connectivity structure of sIgE may be more complex. Hence, higher-order interactions will need to be investigated in the future. Furthermore, because of the iterative nature of the JDINAC estimation procedure, we could not estimate the association strength of the differential pairwise interactions. The interpretation is therefore limited to the direction of the association, and further improvements in model design and further validations are needed to fully capitalise on the potential of these findings.
We acknowledge that through our filtering process [19], some potentially important allergens may have been excluded. However, the filtering process was necessary to moderate the effect of measurement errors and noise. Zero-inflated variables can reduce accuracy and usefulness of a cluster analysis, as well as the reliability of the prediction model results. Filtering also increased the confidence of discovering significant association between sIgEs and clinical outcomes of interest. However, we cannot rule out that, despite their rarity, some of the ‘inactive’ components might be associated with asthma and that the inclusion of inactive components might have resulted in different clusters and classification results. We also acknowledge that our findings do not take into account potentially important factors, such as gender and ethnicity, and that they are derived and validated in the same birth cohort (although among different study participants). Therefore, further validations in external populations are needed to ascertain the generalisability of our findings and to evaluate the presence of population-specific characteristics.
In our previous study using machine learning techniques, we identified three patterns of IgE responses to multiple allergens in the same study population, and each of these patterns was associated with different risk for having asthma [13]. In the current study, we identified seven component clusters that mapped closely to the structural homology of proteins and their biological source (PR-10 proteins, profilins, lipocalins, peanut, grass, trees, and mite clusters). These patterns can be explained by the structural relationships of the allergen components within protein families. The current analysis provided considerably finer granularity compared with our previous analysis, which used Expectation Propagation algorithm implemented in Infer.NET [13]. One possible explanation may be that current methodologies were able to uncover nonlinear relations between the components. Our findings of component clusters are consistent with previous observations that sensitised individual may have detectable IgE to multiple members of the same protein family [36]. For example, one previous study has shown a direct relationship between different representative molecules within three 'panallergen' groups (tropomyosins, profilins, and PR-10s) but little evidence of sensitisation to more than one panallergen [36]. In contrast, our study using a machine learning approach has shown that the PR-10 proteins cluster was central to the network of connectivities and mediated connections between components from other clusters.
Using CRD, several studies have shown that sensitisation to component-specific IgEs is an important risk factor for asthma [37–39]. However, most current guidelines do not recommend assessment of allergic sensitisation as an objective test for asthma diagnosis. This is not surprising, given that in respiratory allergy, the interpretation of SPTs and blood tests that measure specific serum IgE to whole allergen extracts traditionally relies on arbitrary cutoffs (e.g., SPTs > 3 mm, sIgE > 0.35 kUA/L), which have a relatively poor ability to distinguish between benign sensitisations and clinically relevant (‘pathologic’) sensitisation [1, 2]. For example, UK National Institute of Health and Care Excellence (NICE) guidance on the diagnosis of childhood asthma proposes a diagnostic algorithm that incorporates the sequential use of four measures of lung function and inflammation (spirometry, bronchodilator reversibility, fractional exhaled nitric oxide, and peak flow variability, https://www.nice.org.uk/guidance/ng80). We have recently tested the NICE algorithm in a cross-sectional analysis amongst children in our birth cohort aged 13–16 years and found poor agreement between the algorithm and asthma diagnosis; adherence to the algorithm resulted in a substantial number of false positive diagnoses, and the majority of children with asthma were not identified as such by adhering to the proposed algorithm [40]. It is clear that no single test exists for the diagnosis of asthma in children, and using any objective test for diagnosing childhood asthma remains challenging [41]. One important question is whether incorporation of better tests or interpretation algorithms for the assessment of allergic sensitisation would improve diagnostic algorithms for asthma, both in terms of confirming asthma diagnosis and for the assessment of future risk (e.g., of asthma exacerbations or disease persistence). The results of our current study support our notion that ‘allergic sensitisation’ is heterogeneous [4], and provide further evidence that there are several distinct subgroups of sensitisation that differ in their association with asthma. In our previous studies, which used machine learning to investigate patterns of skin test and IgE data to whole extracts of eight major allergens collected at multiple time points throughout childhood, we have shown that some, but not all, classes of sensitisation are associated with asthma presence, progression, and severity [4, 5]. However, these subtypes (clusters/classes) of allergic sensitisation have been identified using statistical inference on large amounts of data collected over long periods [4, 5], and their differentiation at any single cross-sectional point was not possible [42, 43]. Therefore, these observations could not be translated into clinical practice, in which a physician sees a patient at a single time point. It is clear that disaggregation of sensitisation, and knowing which subtype a patient belongs to, may help clinicians predict whether a sensitised patient is likely to have asthma. Our current analysis provides evidence that by using machine learning–based methodologies on CRD data, we can develop better diagnostic algorithms to help practicing physicians differentiate between benign and clinically important allergic sensitisation to help asthma diagnosis [44]. It is of note that our previous studies, which used machine learning but incorporated measures of sensitisation using whole allergen extracts (rather than CRD), were markedly inferior in predicting asthma [12, 45]. Furthermore, compared with our previous studies, in which prediction models correctly classified only one state [12, 45], JDINAC correctly distinguished between children who have asthma and children who do not have asthma.
Another important question is whether similar approaches on CRD data can be used for the assessment of future risk (e.g., of asthma exacerbations) and the prediction of asthma persistence and later-life lung function and chronic obstructive pulmonary disease (COPD) outcomes [6, 7]. In two population-based birth cohorts from the UK and Sweden, we have recently shown IgE reactivity to a limited number of components in preschool identified children at high risk of asthma in adolescence [46]. Persistent asthma at age 16 years in Sweden was predicted by IgE reactivity in early life to four risk molecules (peanut Ara h 1, birch Bet v 1, cat Fel d 1, and grass Phl p 1), whilst in the UK, similar association was observed for five allergenic components (dust mite Der p 1 and Der f 2, timothy grass Phl p 1 and Phl p 5, and cat Fel d 1) [46]. We have also shown that different longitudinal trajectories of sensitisation to allergenic molecules from timothy grass and HDM during childhood had different associations with subsequent asthma [14]. These data suggest that understanding developmental pathways of IgE responses to multiple allergenic components may help development of prognostic algorithms for asthma. To address this, we recently applied novel machine learning techniques to CRD sensitisation data throughout childhood to describe the architecture of the evolution of IgE responses to >100 allergen components from infancy to adolescence [19]. This analysis has shown that the timing of onset of specific patterns of sensitisation may be a key indicator of the subsequent risk. The above studies show that better resolution of longitudinal patterns of sensitisation to multiple allergenic components may facilitate the development of prognostic algorithms that can be used for the prediction of future risk of asthma. Based on the current results, we propose that the pattern of interactions between component-specific IgEs may provide additional valuable information.
Our findings suggest that sIgE responses to multiple allergenic proteins are functionally coordinated and co-regulated, and that the patterns of interactions within this complex network may predict clinical phenotypes. In this study, we found that interactions between a limited set of component-specific sIgEs, rather than individual ‘informative’ components, are associated with increased risk of asthma and may provide the basis for designing diagnostic tools. We need to fundamentally rethink the way we interpret data obtained using CRD and move away from the focus on individual component-specific IgEs to a more holistic approach that takes into account the patterns of connectivity between IgEs.
|
10.1371/journal.pcbi.1004965 | How People Use Social Information to Find out What to Want in the Paradigmatic Case of Inter-temporal Preferences | The weight with which a specific outcome feature contributes to preference quantifies a person’s ‘taste’ for that feature. However, far from being fixed personality characteristics, tastes are plastic. They tend to align, for example, with those of others even if such conformity is not rewarded. We hypothesised that people can be uncertain about their tastes. Personal tastes are therefore uncertain beliefs. People can thus learn about them by considering evidence, such as the preferences of relevant others, and then performing Bayesian updating. If a person’s choice variability reflects uncertainty, as in random-preference models, then a signature of Bayesian updating is that the degree of taste change should correlate with that person’s choice variability. Temporal discounting coefficients are an important example of taste–for patience. These coefficients quantify impulsivity, have good psychometric properties and can change upon observing others’ choices. We examined discounting preferences in a novel, large community study of 14–24 year olds. We assessed discounting behaviour, including decision variability, before and after participants observed another person’s choices. We found good evidence for taste uncertainty and for Bayesian taste updating. First, participants displayed decision variability which was better accounted for by a random-taste than by a response-noise model. Second, apparent taste shifts were well described by a Bayesian model taking into account taste uncertainty and the relevance of social information. Our findings have important neuroscientific, clinical and developmental significance.
| People often change their preferences in the light of what others choose. One form of such change is ‘epistemic trust’ for preferences, i.e. preference alignment over and above any direct benefits that may accrue. We sought to explain preference shifting in terms of normative Bayesian inference in which, along with updating beliefs about what the world is like, and what the correct or profitable answers are given one's tastes, subjects also learn about their own personal tastes when these are incompletely certain. In a novel study based on a well-established paradigms, 740 young people expressed their tastes about the degree to which they preferred a smaller but immediate, versus a larger but delayed, reward. They did this both before and after learning about another agent’s choices. We found taste changed between the two assessments to a degree that was correlated with subjects’ choice variability in the absence of social influence. This is consistent with our Bayesian model if, for instance, people make choices by taking random samples from their own uncertain beliefs. Younger people were influenced by others more than older ones, and this observation was explained in the model by the former being less certain about their own preferences.
| People change their choices, usually in the direction of conformity, when they learn what others value [1]. Reasons for this include the mechanistic, such as forms of priming; the instrumental, such as avoiding the dangers of social non-conformity or to seek social approval; and the epistemic, in which people who are unsure about their own preferences use observations of those of others as data. Interpersonal influence, such as choice convergence, has been extensively studied in instrumental settings. First, alignment with others is explicitly sought when conformity is itself rewarded [2]. Second, choices converge if conformity is not rewarded but choices result explicitly from shared information about the state of the world [3,4]. Toelch and Dolan [1] termed these (social-)normative and informational influence respectively. In contrast, here we focus on epistemic preference change where there is no explicit calculation of improved outcomes [5,6] (though this effect may have even contributed to some behaviour change during experiments that examined instrumental conformity).
In this study we use the term ‘taste’ in a strict sense to mean the function directly mapping stimulus attributes to utility [7]. As an example, if I used to choose oranges over apples but, having gathered social information, I now choose apples because I explicitly estimate that oranges don’t sell [4], this is not a preference change in the sense of ‘taste’. Versions of preference (taste) change have been observed in domains as diverse as oenophilia [8] and pain [9], though more typically in contexts where the values of others have to be inferred indirectly from what amounts to price-lists provided in the experiments. Unlike the present work these studies have not examined the computational structure of such changes. Here, we sought to examine epistemic preference change occasioned by the demands of learning about other’s choices [10].
A important domain in which such effects have been shown is temporal discounting [11], which quantifies the extent to which a person prefers a temporally proximal reward over a distal one, even if the latter is larger. Discounting is of economic [11] and psychiatric [11–14] importance. Thus understanding how social influences might lead people to develop or repair maladaptive discounting is of special clinical relevance. By contrast with many other domains of preference, discounting also enjoys extensively tested mathematical formalizations.
In a recent study [11], we showed that when subjects learned to make discounting choices for other individuals, their own tastes apparently changed to become more like those of these partners. Here, we sought to examine a potentially Bayesian basis for this, testing our ideas on a substantial new sample of subjects whose basic discounting preferences and demographics we also present here.
The premise for our account is that subjects are uncertain about their own taste for discounting. This is entirely plausible in the light of the substantial debate as to the rationale for discounting in the first place, as well as of how taste uncertainty may affect other domains of choice [12,13]. We thus proceed in four steps: (1) subjects’ uncertainty would be reflected in the variability of their choices, even in the absence of perturbing influences; (2) the more uncertain subjects are about their preferences, the more they would shift on learning about others; (3) this degree of preference-shifting could be described in terms of relevance, which we operationalise as the width of the distribution of preferences in a notional reference group of people to whom both the index person and the social influencer belong and (4) these effects would dominate over more complex social motives, such as those stemming from mere participation in the experiment (and thus be independent of the direction of social influence), oppositional traits (shifting away from the Other) or competitive traits ('overtaking' the other). We justify and elaborate these steps using theory and experiment.
In a novel study, participants were recruited from North London and Cambridgeshire as part of the Neuroscience in Psychiatry Network (NSPN). They, or their legal guardians if younger than 16, gave informed consent. The study was approved by the Cambridge Central Research Ethics Committee (12/EE/0250). We invited participants so that the final sample was equally distributed between the two genders and between the ages of 14 to 24. Participants were excluded if they currently received help for a mental health issue, if they had moderate or severe learning disability or serious neurological disorders.
We used the 'Delegated Interpersonal Discounting (DID)' task [11,14]. The task was delivered as part of a battery administered to equal numbers of male and female community dwellers between the ages of 14 and 24 in Cambridgeshire and London, as part of the Neuroscience in Psychiatry Network (NSPN) project. At the time of this study 750 participants had been recruited; 5 withdrew consent; in 4 cases, the research assistant conducting the experiment decided not to complete the task for the sake of the wellbeing of the participant (e.g. tired, unhappy). In a further 3 cases, technical problems rendered the data unusable. We therefore present the analysis of 738 cases.
The task involved three phases. In phase 1, subjects made a series of temporal discounting decisions that we used to estimate their initial value K1 in a standard hyperbolic discounting model. The index 1 stands for phase 1 of the experiment, before learning about another individual. According to this model, the value of a reward RD given after a delay D is
VD=RD/(1+KD)
(1)
where K is the hyperbolic discounting parameter [15–17].
In phase 2, they learned to make choices expressed by another, simulated, participant whose K = Ko differed from theirs. Finally, in phase 3, they made more choices for themselves and the other, allowing us to assess whether their K3 ≠ K1 had changed (3 here indexes phase 3, after exposure to the partner). The Ko of the simulated participant was set to be systematically larger or smaller than K1 by a modest amount in order to provide the temptation to change.
In detail, we approximated the behaviour of participants and simulated the ‘other’ using hyperbolic value discounting followed by a softmax rule:
Q0=R0QD=RD1+KDπD=11+eQ0−QDT
(2)
where πD is the policy probability for choosing the delayed option, Q0,QD are the action values for choosing the immediate or the delayed option (of values R0, RD) respectively, and T is the motivational currency or softmax temperature that quantifies how much a unit change in objective outcomes affects choice probability. In assessing K1 during the experiment, in order to determine Ko and realize the other’s choices, we made the assumption that T = 1, since previous work [11] with this method suggested that this would suffice. However the results below are based on fitting T too.
The 60 trials of phase 1 comprised 30 from a standard set covering a wide range of values of K, and an interleaved set of 30 from an adaptive algorithm. The latter calculated a probability distribution over the possible values of Kb characterising the participant under Eq 2; and then chose a pair of options likely to reduce the uncertainty (entropy) of that distribution as much as possible.
In phase 2, we chose Ko based on K1. Previous results [11] and pilot data led us to expect that the population would have an approximately normal distribution of ln(K) with a mean of roughly μ = −4.5 and a standard deviation of roughly σ = 2.3. We therefore chose ko = ln(Ko) (using lower case k = ln(K)) to be shifted from kb by one σ either towards or away from −4.5 with probabilities 2/3 and 1/3 respectively, simulating real-life encounters that were on the whole not unlikely. We presented participants with options much like the ones in phase 1, but now asked participants “What would [name] choose?” [name] was gender-matched to the participant and likely to be encountered among their peers. It was chosen from a selection of typical names given to children born in England in the last 20 years. Once the participant made their choice on behalf of the seeming Other, we simulated the other's choice (using T = 1) and gave the participant veridical feedback as to whether or not they were correct. We presented trials until either the participant got 8 correct answers out of the most recent 10, or 60 learning trials were completed.
In phase 3, we interleaved mini-blocks of 10 trials 'choose for self', which were as in phase 1, and 10 trials 'choose for other', which were as in phase 2. We instructed participants that one of the 'choose for self' trials from the entire task would be chosen at random and the choice they made paid out for real at the appropriate delay. Participants were instructed that the task was about their “true preferences” and there was no financial incentive to make correct choices in the 'choose for other' trials.
The task was thus very similar to that used by Nicolle and co-workers [14], but optimized for delivering to large community samples. We relied on the experimental design but also in the control experiments performed by Garvert and co-workers [11] to guard against explicit instrumental explanations as well as against simple forms of priming accounting for the change (See SI of [11]). For example we made it very clear to the participants that they would be paid according to the preferences they expressed about themselves only; and that there was no “right or wrong answer” regarding what they chose for themselves. Indeed we were “interested in their own preferences”. The task was coded in MATLAB 2012a running on 12' screen laptops with the Cogent graphics toolbox (see Acknowledgments).
We first consider how to model choice variability along with modal preference, as this will play a key role in understanding preference shift. If we faced a participant with just a single delayed option and found that they chose it, say, 60% of the time, we would not be able to tell if this was because of a relatively high variability parameter (T) or because of relatively weak modal preference (K). However over many trials we used a range of triads of R0 and RD and D to disambiguate the two parameters. In Eq 2 for example this is possible as K only affects the components of the delayed choice whereas T affects both (see also supporting information S1 Text and S1 Data).
We first fitted the classic hyperbolic model and the preference-uncertainty model to the data from phase I of the task. We found (see below) that the preference-uncertainty model was of sufficient quality to use as the backbone for the preference-shift Bayesian schema.
The mainstay of our model-fitting was Markov-Chain Monte Carlo (MCMC) with weakly informative priors and the Component-wise Hit-And-Run Metropolis algorithm, implemented in the ‘LaplacesDemon’ software package [29]. All point estimates reported here are the medians of the posterior distributions of the respective variables (once stationarity was achieved). The phase 1 data were fitted with fixed-effects models (KU and KT). A full hierarchical Bayesian, random-effects analysis of the PS model had too high a dimensionality (740x5 = 3700 parameters) to be fit using MCMC. We therefore fitted it in stages. First, we fitted each individual participant separately, using uninformative priors and a Laplace approximation to the maximum-likelihood as initial conditions–a fixed-effects approach. Second, we used the point estimates of the parameters for each participant to construct an estimate of the distribution of each parameter over our sample. To this effect, and in the first instance, we ignored a small minority of participants whose data did not constrain the model well, i.e. where the stationary distribution was not achieved within 2 million un-thinned samples and / or when the effective sample size was less than 100, indicating poor mixing. Third, following the philosophy of type-2 or empirical Bayesian maximum-likelihood fitting [30], we used our estimate of the sample distributions of the parameters as priors for re-estimating individual parameters.
We first present the analysis of phase I of the experiment, as the results crucially informed our modelling choices for all further analyses. The classic KT model yielded a distribution of preferences over the population that was close to the ones we expected. We expected a mean ln(K1) of roughly μ = −4.5 and a standard deviation of roughly σ = 2.3. We obtained -4.67 with SD = 1.82, justifying a posteriori our choice the choice of Ko for the simulated Other being 2.3 ln (+/-)units away from the Self in phases II and III. T had a mean of 1.54 (SD = 1.36).
We unexpectedly found a powerful correlation between K1 and T in the population, as seen in Fig 2A. This hints that the KT formulation is problematic, as there is nothing in the constructs themselves that suggest that, for example, people who prefer not to wait should not exercise their preference as consistently as those who do wait. Such a high correlation raises the possibility that these measures of discounting and behavioural variability may influence each other, either as a neural phenomenon or an analytical artefact.
The KU formulation abolished this correlation (Fig 2B). We therefore performed model comparison to determine whether it sacrificed quality of fit to achieve this, or whether it was as good in this respect. In the event not only did the KU model capture the correlation between preference and noise in a natural manner, but it also fit the data slightly more proficiently, despite having the same number of parameters. 64% of participants had a better log-likelihood over phase 1 choices for the KU model (SEM 1.8%, Wilcoxon p = 1.7e-11, BIC difference over 738 participants = 740, mean KU log-likelihood = -23.8, mean KT LL = -24.3).
In the KU formulation, even if the mean ms and variance us of ln(K) are uncorrelated across the population, the mean of K and the variance of K will in general be correlated. Through the sampling procedure inherent to KU this will also affect the variability in choices, although the precise nature of this effect will depend on the actual options used to probe discounting (See S1 Text). Fig 2B shows that the inferred values of m1 and u1 across the population are indeed uncorrelated. Reassuringly, m1 correlates closely with the inferred ln(K1) (r = 0.99, p < 1e-10) and u1 being very significantly correlated with Ts (and ln(Ts); r (ln(Ts), u1) = 0.61, p < 1e-10). The former relationship is reassuring as option pairs that are indifferent with respect to one model are also indifferent with respect to the other. The latter relationship is also reassuring in terms of face validity.
Having established KU as our preferred parametrisation, we examined the demographic distribution of discounting preferences. There was no significant dependence of m1 or u1 on gender. m1 declined slightly but significantly with age, Pearson r(m1, age) = -0.10, p = 0.0065. The same was true for the amount of preference shifting towards the ‘other’, with older participants shifting slightly less r(|m3 -m1 |, age) = -0.12, p = 0.0021.
Fig 3 shows how participants shifted their preferences in response to learning about the Other's preferences. The parameters plotted here are descriptive, representing the mode of the Laplace approximation to the likelihood in the pre-exposure vs. post-exposure m3−m1; versus the modal preferences of the Other-self estimated from all the choose-for-other trials. We note that the vast majority of participants shifted in the direction of the Other without overtaking them, just as the uncertainty-relevance model would predict.
We then examined how the two key parameters used to describe preference-shifting in the model related to the variance in the data. We found that σr and u were very significantly correlated with the shift m3-m1 over the whole sample, just as expected from the model. In terms of partial correlation coefficients, r (m3-m1, σr; u) = -0.56, p < 1e-30 while r (m3-m1, u; σr) = 0.61, p < 1e-30, and positive shifts being in the direction of the other’s discounting preference (Fig 4).
As noted for speed and convenience, we used a highly approximate procedure to estimate the K1 that was used as the basis of Ko. It is possible that biases in this procedure could lead to incorrect estimates of the key parameters of the shift model (notably the fixed, low temperature T used which is not a good approximation to our final estimates). We explicitly tested for this by exploiting the fact that we randomized whether subjects were asked in phase 2 to learn about a more patient or more impulsive other. Systematic differences in the parameters between these two possibilities would imply procedural problems. There was some modest evidence for this: those who faced a more patient Other were fitted with a slightly larger u (mean 1.27 vs. 1.11; effect size ~ 0.24; Wilcoxon p = 0.00046) and slightly smaller σr (mean 1.06 vs. 1.21; effect size ~ 0.44; Wilcoxon p = 5.7e-8). We were not able to establish a confound in the model that explained the slight overall bias evident in Fig 3 towards becoming more patient.
We also checked whether there were subsets of participants that shifted their preferences in a systematic way, over and above the uncertainty-relevance model. We thus allowed for an arbitrary perturbation in k between phases II and III of the experiment. This would allow the model to produce a high likelihood for any preference shift, as long as preferences were captured as well by the same basic discounting model (here, the KU model) but it would be agnostic as to the mechanism of this. Examples might be participants that overtake the ‘other’, or shift in the wrong direction (i.e. outside the triangles defined by the identity line and x-axis in Fig 3). We then compared the BIC values for the KU vs. perturbed models. The BIC difference in favour of the perturbation model was > 2 in 7.4% of participants and > 6 in 4.2% of participants. We therefore concluded that the overall fraction of participants where there was strong evidence for a process not captured by our main model, according to BIC conventional values, was in fact small.
Finally, we examined whether σr or u explained the age-dependence of preference shifting that we observed. σr was not significantly correlated with age but u declined (r = -0.14, p = 7.7e-5), and this fully mediated the decrease of preference shifting with age (shifting partial r for age: -0.06, p = 0.11; for u: -0.10, p ~ 0.0). The amount of variance in preference malleability explained by age (and mediated by u) in this sample was small.
We used the paradigmatic case of discounting to model how learning about someone else's preferences may lead to a form of learning about one's own. We tested our models in a new empirical study of over 700 young people which allowed us to make a number of novel contributions. First, we provide evidence that in the presence of social information, Bayesian reasoning updates beliefs about preferences, i.e. the personal tastes themselves, as opposed to beliefs regarding profitable decisions given one's tastes. Second, we show that uncertainty about one's own preferences, reflected in behavioural variability in the absence of social influence, is an important basis for a subsequent preference shift. Third, we introduce the notion of ‘reference dispersion’, which relates to epistemic trust [31,32], and which quantifies ‘how likely is it that my taste are similar to those of an other’. It is thus an estimate of similarity, and can be manipulated in future studies to provide further experimental tests of our model. The novel finding here is that 'reference dispersion' is less than the actual dispersion in the study population, quantifying how participants privilege the experimental context. Finally, we report evidence that decreasing uncertainty about one's own preferences, rather than a change in reference dispersion, accounts for a decreasing malleability in preference with increasing age.
Our study was motivated by an observation that discounting preference shifts take place even if there is no obvious, conventional, motive such as direct reward for making choices like another person's, explicit social approval, or direct gains that accrue to others. Further, the original study on which this one builds indicated that simple priming mechanisms, such as repeating previously performed choices, do not account for taste shifts [11]. Previous studies which examined taste change under social influence in domains such as preferences for facial characteristics of the opposite gender [5] addressed similar issues but did not examine their computational basis. Inferring ‘the best discounting factor for me to like’ may entail analogous distal benefits as inferring ‘the right facial characteristics for me to like’–the crucial point being that such distal benefits are not explicitly calculated but absorbed into tastes.
In our account, subjects were modelled as being uncertain about their own tastes and this uncertainty was reflected in the choices they made even before they learned about the preferences of others. We captured these characteristics in the taste-uncertainty (KU) model by assuming that subjects maintained and updated a distribution over their own taste and sampled from it to make a choice on a trial. This overall model fit the subjects’ behaviour better than the classic softmax (KT), and also explained away an otherwise surprising correlation between the hyperbolic discounting parameter K and the temperature T (see also S1 Text). Sampling matched behavioural variability to uncertainty, which is consistent with recent suggestions about the role of sampling in choice [20,33], and goes beyond the view of random preferences describing the distribution of tastes of individuals across a population, or from inevitable imperfections within a neural system [26]. The better fit of the KU model, the dependence of preference-shift on choice variability and the decrease in taste uncertainty with age suggest that choice variability substantially reflects uncertain taste rather than just ‘trembling hand’, taste-independent response noise. Uncertain taste does not by itself necessitate behavioural variability like the one we have observed. For example, people might have estimated their own modal taste (by taking many samples) and acted on that. However in real life the expression of preference uncertainty in matching behaviour may also be beneficial, somewhat analogous to that of resolving the exploration/exploitation dilemma by Thomson sampling [20,21].
Having a model that depends on beliefs about one’s own tastes renders it straightforward to see how such beliefs might normatively be influenced by evidence. However using observations of others as evidence about the self entails some interpretation. The question becomes one of epistemic trust [32], i.e., (a) deciding the extent to which the people whose choices are being observed are part of the same reference group as oneself, and, (b) whether that behaviour is indicative of their true tastes, or rather could be part of a game-theoretic interaction with inefficient or incomplete mechanism design [34]. In our simplified framework, the parameter σr, the variability about the (unknown) mean of the reference population that is assumed for both self and other, captures the degree of epistemic trust; one limitation of our experiment is that we have little independent evidence about the value of σr. We noted that the mean of the fitted σr = 1.13 is a little less than half the actual population dispersion for ms,, ~ 2.7. This could itself come from an implicit assumption by the participants that the other preferences they are learning about are of special relevance to them—an experiment-induced epistemic trust.
We also observed some asymmetry in participants' shifting, with an overall bias for shifting in a more patient direction (Fig 3, green regression line intercept). Our models accounted for this by a smaller σr (and slightly greater u) for those facing more patient partners. It could be that the experimental procedure exerted an influence on preferences over and above the difference between the participant's and the Other's preferences. People may have a slightly skewed belief distribution about their preferences, or perhaps a skewed sense of similarity. They may consider themselves more similar to patient people than impatient ones (perhaps because of some social stigma). Alternatively this effect may be independent of social reasoning, representing for example a slow reversion to the mean or a practice effect. In our conceptualization σr summarises all sources of relevance that influence learning and its fitting may absorb phenomena like the slight overall shift towards more patient choices. This should be understood further. We consider it important for future studies to actively manipulate interpersonal context on the basis of specific hypotheses about factors that determine epistemic trust (e.g. increased relevance induced by experimental context, out-group vs. in-group belonging) and factors best described separately (reversion to the mean, enhanced conformity to patient behaviour due to social stigma against impatience despite explicit instructions).
In such a large community sample individual variation will be more complex than our simple parametrization allowed. For example, 5.4% (40/738) of participants were fitted with very low, almost zero, taste uncertainty parameters–evident in the two clusters of points with very low u or T in Fig 2. They always chose either the larger or the sooner option. To avoid cherry-picking the data, we included all subjects in the statistical analysis. It may be, however that our options did not correctly span their temporal preferences, as they might have been either far too patient or impulsive. Equally, it is possible that, in such a large sample, they did not follow some aspects of the instructions. Most interesting is the possibility that a single preference model (here, the simple hyperbolic) is an approximation that needs to be refined by considering differences in the very structure of preferences across individuals, as beautifully suggested by Hey, Carbone and co-workers [35]. Additional analyses (SI section S3) confirmed that the Bayesian K-shift model accounted rather precisely for the majority of participants who closely followed the hyperbolic model while a further, exploratory analysis provided evidence for a different sort of uncertainty-based updating in those who do not closely adhere to hyperbolic discounting. It would be important for future research to address in more detail the variation of the structure of preference functions across individuals.
In summary, future research should dissect the nature of similarity or relevance (σr) in our theory through hypothesis-based independent manipulations. In addition, individual variability of preference functions could be addressed in more detail (cf. SI section S3).
In terms of further applications, our findings suggest that other preference measures may be subject to uncertain beliefs and a similar inferential process. It would therefore be useful to have a clearer separation of the ‘taste’ vs. the ‘explicit consequence’ components of preferences in other domains; we acknowledge that this is not straightforward: for example, the issue of ‘pure time preference’ is still a matter of debate with respect to temporal discounting. One relevant domain is development, where it would be important to use longitudinal, rather than cross-sectional, studies to test our explanation that preference malleability changed with age because of increased preference certainty. There are also clinical implications–our findings suggest a mechanism by which therapeutic and malign social influence may operate. For example, clinicians use group treatments to ameliorate disorders now thought to be associated with increased discounting, especially alcohol and drug addiction. In group contexts, the presence of members that have already changed their behaviour and are close to 'graduating' is thought to be an important positive influence on new members [36]. Conversely, being a member of a group containing those with societally unfortunate preferences could lead to maladaptive contagion.
|
10.1371/journal.ppat.1007299 | Infection with flaviviruses requires BCLXL for cell survival | BCL2 family proteins including pro-survival proteins, BH3-only proteins and BAX/BAK proteins control mitochondria-mediated apoptosis to maintain cell homeostasis via the removal of damaged cells and pathogen-infected cells. In this study, we examined the roles of BCL2 proteins in the induction of apoptosis in cells upon infection with flaviviruses, such as Japanese encephalitis virus, Dengue virus and Zika virus. We showed that survival of the infected cells depends on BCLXL, a pro-survival BCL2 protein due to suppression of the expression of another pro-survival protein, MCL1. Treatment with BCLXL inhibitors, as well as deficient BCLXL gene expression, induced BAX/BAK-dependent apoptosis upon infection with flaviviruses. Flavivirus infection attenuates cellular protein synthesis, which confers reduction of short-half-life proteins like MCL1. Inhibition of BCLXL increased phagocytosis of virus-infected cells by macrophages, thereby suppressing viral dissemination and chemokine production. Furthermore, we examined the roles of BCLXL in the death of JEV-infected cells during in vivo infection. Haploinsufficiency of the BCLXL gene, as well as administration of BH3 mimetic compounds, increased survival rate after challenge of JEV infection and suppressed inflammation. These results suggest that BCLXL plays a crucial role in the survival of cells infected with flaviviruses, and that BCLXL may provide a novel antiviral target to suppress propagation of the family of Flaviviridae viruses.
| The genus Flavivirus including Japanese encephalitis virus, Dengue virus, and Zika virus all of which are mosquito-borne human pathogen and cause serious diseases in humans. Therefore, the development of effective vaccines and antivirals against several flaviviruses is still needed. BCL2 family proteins control mitochondria-mediated apoptosis to maintain cell homeostasis via the removal of damaged cells and pathogen-infected cells, deregulation of which leads to severe diseases including cancer and autoimmune diseases. Here, we showed that BCLXL is a critical cell survival factor during infection with flaviviruses, and that inhibition of BCLXL by treatment with BH3 mimetics restricts the production of infectious particles and the expression of chemokines in vitro and in vivo. Inhibition of BCLXL induces apoptosis in cells infected with flaviviruses and these cells are quickly removed by engulfment of phagocytes, which leads to inhibition of virus dissemination without any inflammatory reaction. Based on these data, BCLXL would appear to be a suitable target for the development of novel antivirals against a broad range of flavivirus infections.
| The genus Flavivirus belongs to the family Flaviviridae and includes the Japanese encephalitis virus (JEV), Dengue virus (DENV), West Nile virus (WNV) and Zika virus (ZIKV), all of which are mosquito-borne human pathogens [1,2]. The flaviviruses are internalized into dendritic cells, such as Langerhans cells [3], and keratinocytes in skin [4] through mosquito bites. Dendritic cells that are thus infected with flaviviruses are activated and migrate into the lymph nodes, facilitating viral spread into peripheral tissues [5] and the induction of visceral and/or central nervous system diseases. DENV infection may generate dengue haemorrhagic fever and dengue shock syndrome [6], while infection with either WNV or JEV can cause severe encephalitis [7,8]. Recently, an outbreak of ZIKV infection in Brazil was linked to birth defects, including microcephaly, and to Guillain-Barre syndrome in adults [9,10]. Therefore, the development of effective vaccines and antiviral drugs against the family of flaviviruses is an urgent need.
Infection with flaviviruses induces apoptosis in cells by the production of reactive oxidative species, unfolded protein responses, and viral proteins [11], in addition to other types of cell death, such as necrosis and pyroptosis. Receptor-interacting serine/threonine-protein kinase-3 (RIPK3), a central player of necroptosis, has been shown to play an additional role during virus infection [12]. Infection with either DENV or hepatitis C virus (HCV) induces interleukin-1β production and pyroptosis [13]. The biological significance of pyroptosis in the pathogenicity of flavivirus infection remains largely unknown [14].
Apoptosis is a type of programmed cell death that removes excess cells during development, as well as during routine homeostasis [15–17]. B-cell lymphoma-2 (BCL2) family proteins play central roles in mitochondria-mediated apoptosis [18]. BCL2 family proteins have been divided into three groups: BH3-only proteins, BAX/BAK proteins and pro-survival proteins. Once an apoptotic stimulus such as DNA damage occurs, BH3-only proteins, such as BIM, BID, PUMA, BAD and NOXA, are activated and directly or indirectly activate BAX and BAK. Activated BAX/BAK forms homo-oligomers that permeabilize mitochondria and allow the release of pro-apoptotic factors such as cytochrome c, which promotes activation of caspases [19]. Pro-survival proteins, such as BCL2, BCLW, BCLXL, MCL1 and A1, sequester BH3-only proteins to prevent self-activation of BAX/BAK by interaction with the BH3 domains of BAX/BAK [20]. In a notable extension to cancer therapy, small molecule inhibitors have been developed, including ABT-737, ABT-263 (navitoclax, an orally available clinical derivative of ABT-737), ABT-199 (venetoclax) and A-1331852. ABT-199 and A-1331852 target BCL2 and BCLXL, respectively, while ABT-737 and ABT-263 target BCL2, BCLXL and BCLW. ABT-199 was approved by the United States Food and Drug Administration for the treatment of 17-p deleted chronic lymphocytic leukaemia (CLL) [21], supporting the therapeutic use of induction of cell death by mitochondria-mediated apoptosis to eliminate unnecessary cells.
In this study, we showed that cells infected with flaviviruses including JEV, DENV and ZIKV specifically induced apoptosis by inhibition of BCLXL through the suppression of MCL1 expression. Furthermore, we showed that host protein synthesis is impaired in cells upon infection with flaviviruses, leading to the decay of labile proteins such as MCL1. Treatment with inhibitors for BCLXL, as well as deficiency of BCLXL gene, induced BAX/BAK-dependent apoptosis in the infected cells. Finally, engulfment of virus-infected cells by macrophages was significantly enhanced by the inhibition of BCLXL without an increase of the production of pro-inflammatory cytokines, which led to the suppression of viral dissemination.
Firstly, to examine the roles of BAX/BAK-dependent apoptosis during flavivirus-infected cells, we generated BAX and BAK double-knockout (BAX/BAKDKO) Huh7 cell lines (clones #29 and #47) and infected with Japanese encephalitis virus (JEV) (S1A Fig). Although Huh7 cells induced cell death at 4 days post-infection with JEV, BAX/BAKDKO cells exhibited a complete resistant to mitochondria mediated apoptosis [22], suggesting that cells infected with JEV induce mitochondria-mediated apoptosis at 4 days post-infection. However, BAX/BAKDKO cells also exhibited cell death at 5 and 6 days post-infection. The BAX/BAK-independent cell death might be induced by necrosis or pyroptosis rather than apoptosis as previously reported (Fig 1A) [12].
Recent clinical studies showed that survival of chronic lymphocytic leukaemia (CLL) is largely dependent on BCL2 [21], and that loss of a single Bclx allele attenuates MYC-induced lymphoma in vivo [23]. In addition, selective inhibition of MCL1 has been shown to induce apoptosis in acute myelogenous leukaemia cell lines without signs of toxicity to normal human haematopoietic progenitor cells [24]. It had been reported that cellular stress such as ER stress, DNA damage and innate immune response affects on the expression of BCL2 protein family [25–27]. Therefore, we speculated that virus infection alters the expression of BCL2 proteins through the induction of cellular stress. BH3-only proteins are known to exhibit specific interactions with pro-survival BCL2 proteins, including BCL2, BCLXL, BCLW and MCL1 [28]. BIM and PUMA bind to all pro-survival BCL2 proteins; BAD binds selectively to BCL2, BCLXL and BCLW; and NOXA binds specifically to MCL1 (S1B Fig). To investigate the roles of pro-survival BCL2 proteins in cells infected with flaviviruses, we generated Huh7 cell lines stably expressing a BIM-mutant possessing either a binding-defective BH3 domain (BIM-4EBH3), the BH3 domain from BAD (BIM-BADBH3), or the BH3 domain from NOXA (BIM-NOXABH3) (S1C and S1D Fig). Although lentivirus-induced expression of BIM-NOXABH3 or BIM-BADBH3 specifically killed Huh7 cells expressing BIM-BADBH3 or BIM-NOXABH3, respectively, parental and Huh7 cells expressing BIM-4EBH3 were resistant to killing through the expression of BIM-mutants (S1E Fig). Although no cell death was observed in parental Huh7 cells, or in cells expressing either BIM-4EBH3 or BIM-NOXABH3, upon infection with JEV and DENV, cells expressing BIM-BADBH3 showed a significant reduction in viability at 3 days-post infection (Fig 1B).
To assess which pro-survival BCL2 proteins were responsible for the survival of cells infected with flaviviruses, the effects of BH3 mimetics on flavivirus-infected cells were determined. No cytotoxicity was observed in Huh7 cells solely treated with BH3 mimetics at 3days post treatment (S1F Fig). Cells infected with JEV, DENV and ZIKV, and treated with either ABT-737 [29] (targeting BCL2, BCLW and BCLXL) or A-1331852 [30] targeting BCLXL exhibited cell death at 3 days-post infection. In contrast, infected cells treated with ABT-199 [31] (targeting BCL2) showed no cell death (Fig 1C and S1B Fig). Treatment of ABT-737 (1 μM) significantly accelerated JEV, DENV and ZIKV induced cell death (Fig 1D). Next, to examine the roles of BCLXL inhibition on infectious virus production, we quantified viral titres by using a focus-forming assay. Treatment with ABT-737 (1 μM) exhibited no effect on the infectious particle production of JEV, DENV and ZIKV until 2-days post-infection, but significantly reduced the infectious titers at 3 and 4 days post-infection (Fig 1E), suggesting that induction of cell death by the treatment with ABT-737 in viral infected cells reduces infectious virus production. We next determined whether ABT-737 treatment induces BAX/BAK-dependent cell death upon infection with flaviviruses. Although parental Huh7 cells exhibited cell death upon JEV infection and treatment with ABT-737, which activated caspase 3/7, BAX/BAKDKO cells were resistant to cell death under the same conditions, and no activation of caspase 3/7 was detected (Fig 1F and 1G). In addition, release of cytochrome c was observed in parental Huh7 cells infected with JEV and treated with ABT-737, but not in BAX/BAKDKO cells (S1G Fig). Although no increase of infectious particle production was observed in BAX/BAKDKO cells until 2-days post-infection, three to ten times of increase of infectious titer compared to parental Huh7 cells was observed in BAX/BAKDKO cells at 3 and 4 days post-infection (Fig 1H). Collectively, these results suggest that cell death is induced upon infection with JEV and treatment with ABT-737 through a BAX/BAK-dependent pathway.
To confirm the BCLXL-dependent survival of cells upon infection with flaviviruses, we generated Huh7 cell lines that were deficient in either BCLXL or MCL1 (BCLXKO or MCL1KO) (Fig 2A). Expression of BIM-BADBH3 and BIM-NOXABH3 killed MCL1KO and BCLXKO cells, respectively, while expression of BIM-4EBH3 conferred resistance to both parental and KO cells (Fig 2B). Among the pro-survival proteins of BCL2 protein family, BCLX and MCL1 redundantly suppress activation of both of BAX and BAK. On the other hands, BCL2 and BCLW had been reported to suppress BAX but not BAK [32–36]. In addition, deficiency of BCLW in mice developed normally except spermatogenesis in male mice, suggesting that BCLW is important for apoptosis in testis [37]. To examine the redundancy of BCL2 proteins, we treated parental, MCL1KO and BCLXKO Huh7 cells with either A-1331852, BCLX-specific inhibitor or S63845, MCL1 selective inhibitor [24]. MCL1KO Huh7 cells treated with A-1331852 (1 μM) but not with S63845 (1 μM) and BCLXKO Huh7 cells treated with S63845 (1 μM) but not with A-1331852 (1 μM) induced cell death, while parental Huh7 cells are resistant to the treatment with either compounds but induced cell death by the treatment with both compounds (S2A Fig). These data suggest that BCLX and MCL1 redundantly suppress activation of BAX and BAK. The viabilities of BCLXKO cells upon infection with JEV, DENV or ZIKV were consistently lower than those of parental and MCL1KO Huh7 cells (Fig 2C), and we confirmed the induction of cell death upon infection with JEV in other independently established BCLXKO Huh7 cell lines (S2B and S2C Fig). Next, we determined the effect of BCLX inhibition on plaque formation. Furthermore, Huh7 cells did not form any plaque upon infection with JEV, DENV and ZIKV, however, BCLXKO Huh7 cells exhibited formation of clear plaques upon infection with these viruses (Fig 2D). Taken together, these results indicate that apoptosis is induced in BCLX-deficient cells upon infection with flaviviruses.
BAX/BAK activation is suppressed by pro-survival BCL2 proteins [20]. To investigate how BCLXL participates in cell survival upon infection with flaviviruses, expression levels of BCLXL and MCL1 were determined in flavivirus-infected cells. Interestingly, expression of MCL1 decreased upon infection with JEV, DENV and ZIKV, in contrast to the continued stable expression of other BCL2 family proteins (Fig 3A). No MCL1 protein was detected in the insoluble fraction of cell lysates (S3A Fig), suggesting that the total amount of MCL1 protein is decreased by the infection with flaviviruses.
MCL1 has been shown to be a short-lived protein that is degraded by the proteasome [32,38,39]. Northern blot analysis revealed stable expression of mRNA of MCL1 in JEV-infected cells (S3B Fig). Reduction of MCL1 protein by JEV infection was slightly restored by the treatment with a proteasome inhibitor (Ac-Leu-Leu-Nle-Aldehyde, ALLN; (20 μM), however, the expression level of MCL1 was not recoverable by the inhibition of proteasome (Fig 3B). Moreover, lysosome inhibitor treatment did not restore reduction of MCL1 protein (S3C Fig), suggesting that JEV infection suppresses host translation resulting in the reduction of MCL1 protein expression through proteasome-mediated degradation.
A previous study showed that MULE (HUWE1) is a unique BH3-containing E3 ubiquitin ligase that is associated with suppression of MCL1 expression [40]. Suppression of MCL1 expression in JEV-infected MULEKO cells was comparable to MCL1 in JEV-infected parental Huh7 cells (S3D and S3E Fig). Further, the Skp, Cullin and F-box containing complex (SCF complex) has also been shown to play a role in MCL1 degradation [41,42]. Although the overexpression of dominant negative Cullin-1 (DN-CUL1) led to an upregulation of MCL1 in non-infected cells, infection with JEV still suppressed MCL1 expression (S3F Fig). Additionally, overexpression of the BH3-only protein, NOXA, has been shown to induce suppression of MCL1 expression [32]. Suppression of MCL1 expression in JEV-infected NOXAKO cells was comparable with MCL1 in JEV-infected parental Huh7 cells (S3G and S3H Fig). Importantly, expression of MCL1K/R, which is an ubiquitination-defective mutant of MCL1, was still supressed upon JEV infection (S3I and S3J Fig). The ts20 cells are derived from Balb/c 3T3 cells and possesses a temperature-sensitive ubiquitin-activating E1 enzyme [43]. The ts20 cells were infected with JEV, then incubated for 6 h at 37°C (active E1) or 39°C (inactive E1), 2 days post-infection. Notably, the expression of p53, but not of MCL1, was restored by inhibition of E1 (S3K Fig), suggesting that ubiquitination is not involved in the suppression of MCL1 protein expression in cells infected with JEV.
To examine the effects of JEV infection on protein synthesis, JEV-infected cells were labelled with [35S]-methionine and [35S]-cysteine. The total amount of protein synthesis was significantly reduced at both 2 and 3 days post-infection with JEV (Fig 3C, left panel). Additionally, protein synthesis of MCL1 was significantly reduced by infection with JEV (Fig 3C, right panel). In addition, expression of CCND1, another short half-life protein [44,45], was also suppressed by JEV infection (Fig 3D). We hypothesized that expression of short half-life proteins was non-specifically suppressed upon infection with flaviviruses because labile proteins were constitutively degraded under inhibition of newly protein synthesis. To clarify it, we generated Huh7 cell lines stably expressing either GFP or D2GFP, a destabilized form of GFP that contains a PEST sequence [46]. Expression of D2GFP and MCL1 was decreased upon infection with JEV in a dose-dependent manner at 2 days post infection, in contrast to the stable expression of GFP (Fig 3E). These results suggest that infection with flaviviruses induces inhibition of protein translation, leading to a reduction of short half-life proteins such as MCL1.
Next, to examine the effect of viral proteins on MCL1 expression, we transfected the plasmids encoding JEV structural proteins (HA-Core and prM-E), non-structural proteins (HA-NS1, HA-NS2A, HA-NS2B, HA-NS3, HA-NS4A, HA-NS4B and HA-NS5) and control vector into Huh7 cells. Expression of viral proteins did not induce suppression of MCL1 expression (S3L Fig). These results suggest that viral propagtion is required for suppression of MCL1 expression.
To examine the relationship between the suppression of MCL1 expression and a susceptibility to JEV propagation in other human cell lines, we selected cell lines with high and low susceptibility to JEV infection: IGR-OV1, H522 and SF-268 cells (high susceptibility) and OVCAR5 and HT29 cells (low susceptibility) (Fig 4A). Expression of MCL1 was reduced in the highly susceptible cell lines upon infection with JEV; in contrast, MCL1 was stably expressed in cell lines with low susceptibility (Fig 4B). Moreover, cell lines with high susceptibility were killed by the combination of infection with JEV and treatment with the BH3 mimetic ABT-737 (1 μM) (Fig 4C) or A-1331852 (1 μM) (S4A Fig); in contrast, cell lines with low susceptibility remained viable when receiving this treatment. In addition, the suppression of MCL1 expression was enhanced upon infection with JEV (Fig 4D), but suppressed in a dose-dependent manner by treatment with IFNα (Fig 4E). Further, the decrease in cell viability induced by the combination treatment was reversed upon treatment with IFNα (Fig 4F). These data suggest that viral replication is important for suppression of MCL1 expression and suppression of MCL1 expression is independent on IFN signalling pathway. U937 cells, a human lymphoma cell line, exhibited high, moderate and low susceptibility to ZIKV, JEV and DENV, respectively (Fig 4G). Suppression of MCL1 expression was consistent with the degree of susceptibility to each virus at 2 days post infection (Fig 4H). Treatment with the BH3 mimetic ABT-737 (1 μM) or A-1331852 (1 μM) enhanced cell death in U937 cells that were also infected with ZIKV (Fig 4I), but not in cells that were also infected with JEV at 3 days post infection (S4B Fig). Taken together, these data suggest that suppression of MCL1 is required for an efficient viral propagation.
Primary targets of JEV, DENV and ZIKV include haematopoietic and skin cells; therefore, we next examined the effects of BCLX inhibition during flavivirus infection by using host cells closer to human infection. Monocyte-derived dendritic cells (MDDCs) have been shown to be highly susceptible to DENV infection [47]. Treatment with ABT737 (1 μM) significantly increased cell death in MDDCs that were also infected with DENV (S4C Fig). In addition, treatment with ABT-737 (1 μM) or A-1331852 (1 μM) significantly increased cell death in human melanoma cell lines, UACC257 and UACC62, upon infection with JEV (S4D Fig), and in mouse embryonic fibroblasts (MEFs) infected with JEV, DENV or ZIKV at 2 days post infection (S4E Fig). These results suggest that suppression of MCL1 expression and induction of cell death by the treatment with an inhibitor for BCLX are conserved in a variety of human and mouse cells upon infection with flaviviruses.
An arthropod-borne flavivirus primarily infects dendritic cells, such as Langerhans cells, and keratinocytes in skin [3]. These infected cells are stimulated and then migrate into the lymph nodes [48], leading to propagation within the lymphatic system. We hypothesized that the induction of cell death by BCLXL inhibition in flavivirus-infected cells would accelerate exposure of ‘eat-me’ signal on the surface of the cells and engulfment of infected cells by phagocytes, thus inhibiting dissemination of virus infection. To examine the involvement of phagocytosis in JEV-infected cells treated with the BH3 mimetic ABT-737, we performed an in vitro phagocytosis assay using peritoneal exudate cells (PECs) as model phagocytes (Fig 5A). Both of BCLXKO and ABT-737-treated Huh7 cells infected with JEV exhibited a significant increase of engulfment by PECs at 2 and 3 days post-infection, even though no cell death was observed at 2 days post-infection (Fig 5B). In contrast, no phagocytosis was observed in JEV-infected Huh7 cells (Fig 5B).
To examine whether PECs produce chemokines or cytokines upon co-culture with JEV infected cells, BCLXKO and ABT-737-treated Huh7 cells infected with JEV were overlaid onto PECs at 1, 2 and 3 days post-infection. Production of chemokines and cytokines in the supernatants was determined after 24 h of incubation. Significant amounts of chemokines and cytokines were observed in the supernatants of co-cultured PECs with JEV infected but not with mock-infected Huh7 cells (Fig 5C). Inhibition of BCLXL in JEV-infected Huh7 cells suppressed the production of pro-inflammatory cytokines by PECs, including CCL5, CCL2 and CXCL10, while no significant difference in the expression of TNF-α and IL-6 was observed (Fig 5C).
To examine whether infectious particles are released from the co-culture system, the supernatants were also collected and infectious titers were determined after 24 h of co-culture. Infectious titers were significantly reduced in the supernatants of PECs that were co-cultured with either BCLXKO or ABT-737-treated Huh7 cells, compared with PECs that were co-cultured with parental Huh7 cells at 2 and 3 days post-infection with JEV (Fig 5D). Additionally, we confirmed that JEV could propagate in PECs (S5 Fig). These results suggest that the inhibition of BCLXL in viral-infected cells accelerates efficient engulfment by phagocytes, thereby suppressing the production of chemokines and infectious particles.
To further examine the biological significance of BCLXL inhibition in virus-infected cells, mice that were heterozygotic for Bclx were generated using a CRISPR/Cas9 system (S6A Fig), because homozygous mice died around E13 due to massive cell death of immature hemopoietic cells and neurons [49]. We obtained F1 mice and generated Bclx+/- mice (F2) by a cross with wild-type mice. We established three Bclx+/- lines from separate lines of F1 mice (No. 32, No. 40 and No. 44, S6A Fig). Bclx+/- mice expressed less BCLX protein in the spleen compared to wild type mice (S6B Fig). We injected lethal amounts of JEV into the footpad of each mouse (subcutaneous challenge). Bclx+/- mice exhibited higher survival rates against footpad challenge with JEV, compared with wild-type mice (Fig 6A). Expression of viral RNA, as well as Ccl5 and Cxcl10, but not Tnfα, Il6 Ccl2, Ifna and Ifnb, was significantly decreased in the footpads of Bclx+/- mice at 5 days post-infection (Fig 6B and 6C and S6C Fig). In contrast, the expression of Ccl5, Cxcl10, Tnfα and Il6 was comparable in wild-type and Bclx+/- mice upon stimulation with Poly(I:C) (Fig 6D and S6D Fig). In addition, oral administration of a clinically relevant dose of ABT-263 (50mg kg-1), an analogue of the BH3 mimetic ABT-737, conferred prolonged survival against subcutaneous challenge of JEV (Fig 6E). In contrast, Bclx+/- mice exhibited similar survival rates and viral RNA production upon intracerebral challenge of JEV with wild-type mice (Fig 6F and 6G), suggesting that the suppression of virus growth in peripheral tissues participates in the delay of death and higher percentage of survival in Bclx+/- mice against JEV challenge. These results suggest that viral propagation within flavivirus-infected cells can be inhibited by the suppression of BCLXL and subsequent induction of apoptosis.
In this study, we showed that flavivirus infection suppresses expression of labile proteins such as MCL1 without observable cytotoxicity. Clear cell death, through activation of caspase 3/7, occurred only when BCLXL was inhibited. Moreover, we showed that treatment with BCLXL inhibitor (A-1331852) or deficiency of the BCLXL gene induces cell death upon infection with flaviviruses in several cancer cell lines and primary MEFs. Previous report showed that BCLXL and MCL1 predominantly inhibit activation of BAK [50], suggesting that a mechanism to inhibit BCLXL may induce BAK-dependent apoptosis in flavivirus-infected cells, therefore inhibition of BCLX alone could enhanced cell death in virus-infected cells. In addition, viral infection decreased MCL1 expression in a variety of human and murine cell lines. Although the magnitude of suppression of MCL1 expression varied depending on cell line, our data suggest that suppression of MCL1 expression upon propagation of flaviviruses is a common feature. Legionella infection also induces suppression of MCL1 expression [51]. Although the molecular mechanisms of MCL1 protein expression in cells infected with Legionella remain unclear, the suppression of MCL1 through shut off of host translation might be part of the cellular innate response against infection with intracellular pathogens.
MCL1 expression is strictly controlled by transcriptional and post-transcriptional mechanisms [52]. In this study, we showed that flavivirus infection suppresses protein synthesis, thereby reducing the quantity of proteins with short half-lives, such as MCL1. In addition, recent reports have shown that infection with flaviviruses represses translation of host proteins, but does not repress viral proteins, which are independent of PKR and eIF2α [53]. Further, infection with mammalian orthoreovirus has been shown to inhibit host translation by compartmentalization of translation machinery within the viral replication complex [54]. Protein translation is regulated by ribosomal proteins [55]. DENV NS1 protein specifically binds to the ribosomal protein RPL18 to sequester in perinuclear region and to regulate viral translation and replication [56]. RPLP1 and RPLP2 identified by genome-wide RNAi screening were host factors to contribute DENV and yellow fever virus (YFV) infection [57]. We previously reported that heterogeneous nuclear ribonucleoprotein A2/B1 (HNRNPA2B1) interact with JEV core protein and facilitate JEV replication through the translocation of core protein from nucleus to cytoplasm [58]. In addition, infection with HCV has been shown to induce accumulation of adenosine 5′-triphosphate in the viral replication complex [59]. Because expression of JEV proteins exhibited no effect on host translation (S3L Fig), viral propagation might induce shut-off of host translation by unknown mechanism rather than PKR and eIF2α. On the other hand, viruses counteract host responses through the rearrangement of membrane structures to restrict the components of translation machinery into the viral replication complex.
Our data indicate that inhibition of BCLXL in viral-infected cells activates phagocytosis by macrophages, leading to the suppression of infectious particle production and suggest that inhibition of BCLXL function as an anti-viral effect. Although the survival of mice that were intracerebrally inoculated with JEV was not significantly different between wild-type and Bclx+/- mice, Bclx+/- mice showed higher survival rate against subcutaneous challenge by JEV, compared with their wild-type counterparts. Our data suggest that phagocytosis plays an important role in removing apoptotic.
JEV-infected cells after BCLX inhibition. Microglial cells, a resident subset of macrophages in the brain, and hematogenous macrophages play a role in phagocytosis in the brain and peripheral tissues, respectively. These two-types of macrophage have phagocytotic activity but have some different functions [60]. The functional difference of the microglia and hematogenous macrophages might be involved in the phagocytosis-mediated removals of virus-infected cells by BCLX inhibition in intracerebral or subcutaneous infection. Collectively, these data suggest that apoptosis induced by inhibition of BCLXL attenuates viral virulence in peripheral tissues.
The production of cytokines and chemokines during viral infection is strongly associated with viral pathogenicity [61], especially, deficiency of several chemokine receptors such as CCR5 and CXCR3 promotes severity of pathogenicity against WNV infection [62]. We showed that the inhibition of BCLXL specifically induces apoptosis and impaired production of CCL5 and CXCL10, which are ligands for CCR5 and CXCR3, respectively, during in vitro and in vivo viral infections. Impairment of these chemokines may contribute attenuation of pathogenicity by inhibition of BCLXL. In addition, administration of the BH3 mimetic, ABT-263, into wild-type mice increased survival rate after subcutaneous JEV challenge. We also showed that production of viral RNA and chemokines was impaired in Bclx+/- mice (Fig 6B and 6C). In addition, treatment of wild type and Bclx+/- mice with polyIC exhibited no effect on chemokine induction (Fig 6D). These data suggest that enhanced apoptosis in Bclx+/- and ABT-263-treated mice upon infection with JEV promotes engulfment by phagocytes, which leads to reduces viral production and the number of leukocytes migrating to the infection site due to lower production of chemokines.
Treatment with inhibitors of BCLX such as ABT-263 was demonstrated to cause thrombocytopenia due to their neutralization of BCLXL in circulating platelets without bone marrow toxicity [30,63,64]. It was also demonstrated that thrombocytopenia induced by ABT-263 is dose-dependent and reversible effect [64], therefore it is possibility to control thrombocytopenia by appropriate terms of treatment or amounts of BCLX inhibitors [65,66]. It is also another possibility to develop the drug delivery system to virus-infected cells to avoid the unexpected side effects. Further studies are needed to use inhibitors of BCLX for antiviral therapy.
Based on these results, we propose the model as shown in Fig 7 to explain the regulation of BCL2 family proteins upon infection with flaviviruses. In this study, we showed that infection with flaviviruses reduces expression of MCL1 but remaining BCLXL delayed apoptosis in infected cells. This delay of apoptosis is crucial for viral dissemination to neighbouring cells, leading to high pathogenicity. On the other hand, impairment of BCLXL accelerates apoptosis in cells infected with flavivirus and dissemination is reduced by the removal of infected cells through phagocytosis, leading to local infection and low pathogenicity in vivo.
In conclusion, we have demonstrated that BCLXL is a critical cell survival factor during infection with flaviviruses, and that inhibition of BCLXL by treatment with BH3 mimetics restricts the production of infectious particles in vitro and in vivo. We showed that expression of pro-survival MCL1 is reduced upon infection with flaviviruses. Although further studies are needed to clarify the molecular mechanisms, the translational suppression induced by viral infection is suggested to participate in the instability of unstable proteins including MCL1. Bclx+/- mice exhibited impairment of pathogenicity to JEV infection, together with the enhanced viral clearance by phagocytosis of infected cells through the stimulation of apoptosis by BCLXL inhibition. In addition, Bclx+/- mice, and wild-type mice that received ABT-263, exhibited prolonged survival rates after JEV challenge compared with untreated wild-type mice. These data suggest that BCLXL may serve as a novel antiviral target that facilitates suppression of the propagation of a broad range of the family Flaviviridae viruses.
Huh7, 293T, U937, and Vero cells were obtained from the National Institute of Infectious Diseases and cultured in Dulbecco’s Modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 100 U/ml penicillin and 100 μg/ml streptomycin. IGR-OV1, SF268, H522, OVCAR5, HT29, UACC257, UACC62 and ts20 were provided by the Walter & Eliza Hall Institute and maintained in RPMI1640 supplemented with 10% fetal bovine serum (FBS), 100 U/ml penicillin and 100 μg/ml streptomycin. The MEFs were obtained from E14.5 embryos and maintained in DMEM supplemented with 10% fetal bovine serum (FBS), 100 U/ml penicillin and 100 μg/ml streptomycin. JEV (AT31 strain), DENV (H241 strain) and ZIKV (MR766-NIID strain) were propagated in C6/36 cells. Infectious titers of JEV, DENV and ZIKV were determined by focus-forming assay. Mouse monoclonal antibodies against JEV/DENV/ZIKV NS3 (#578, #1791);[58] were used to visualize the focuses.
BIM, BIM-BADBH3, BIM-NOXABH3 and BIM-4EBH3 [28] were amplified by Tks Gflex DNA Polymerase (Takara-Bio) and cloned into an FUIPW lentiviral transfer vector [67] by using an In-Fusion HD cloning Kit (Takara-Bio). The MCL1K/R gene was synthesized by Thermo Fisher Scientific (GeneArt Strings), amplified and cloned into an FUIPW vector or a pCAGGS vector [68]. The cDNA of D2GFP [46] was synthesized (Integrated DNA Technologies (IDT)) and cloned into an FUIPW vector. pCAG-HA-Ubiquitin was previously described [69]. Plasmids encoding HA-tagged JEV Core and NS proteins (pCAGPM-HA-NS) were previously described [58,70]. Plasmids encoding JEV prM-E protein was previously described [71]. pcDNA3-DN-hCUL1-FLAG was obtained from Addgene (#15818). The pSPT18 was obtained from Dr. Kamitani. For the CRISPR/Cas9 system, the plasmids of pX330 (#42230) and pCAG EGxxFP (#50716) were obtained from Addgene. All sgRNAs are listed in Appendix S1 Table. The lentivirus packaging vectors, pMDLg/pRRE, pRSV-Rev and pCMV-VSV-G, were obtained from Addgene. The sequences of all plasmids used in this study were confirmed by using an ABI Prism 3130 genetic analyzer (Applied Biosystems).
To generate lentiviruses, 293T cells were transfected with transfer vectors, pMDLg/pRRE, pRSV-Rev and pCMV-VSV-G, using polyethylenimine (PEI; MW: 25,000; Polysciences Inc.), and the culture supernatants collected at 3 days post-transfection were passed through a 0.45-mm filter. Culture supernatants containing the lentiviruses were inoculated into target cells seeded on 6-well plates using Polybrene (Sigma) and were centrifuged at 2,500 rpm for 45 min at 32 °C.
The following antibodies were used: anti-JEV/DENV/ZIKV NS3 monoclonal antibody (#578, #1791, 1:1000 dilution) [58], anti-ACTIN mouse monoclonal antibody (A2228; Sigma, 1:5000 dilution), horseradish peroxidase-conjugated anti-FLAG mouse monoclonal antibody (clone M2; Sigma, 1:1000 dilution), anti-HA rat monoclonal antibody (clone 3F10; Roche, 1:1000 dilution), mouse anti-cytochrome c (BD 556433, 1:1000 dilution), rabbit polyclonal antibodies to MCL1 (sc-819; Santa Cruz, 1:1000 dilution, 600-401-394; Rockland, 1:5000 dilution), mouse monoclonal antibodies to BCLXL (BD 610747, 1:1000 dilution, BD 610212, 1:1000 dilution), BAK (8F8, 1:1000 dilution), BAX (21C10-23-8-3, 1:1000 dilution), CYCLIN D1 (BD 556470, 1:1000 dilution), FKBP8 [72] (1:1000 dilution), MULE (NB100-652; Novus Biologicals, 1:1000 dilution), p62 (M162-3; MBL, 1:1000 dilution), p53 (1C12; Cell Signaling Technology, 1:1000 dilution), ABT-737, ABT-199 and A-1331852 were synthesized by Walter & Eliza Hall Institute. ABT-263 was obtained from Synkinase. S63845 was obtained from Active Biochem. ALLN (A6185) and bafilomycin A1 (B1793) were obtained from Sigma, QVD-OPH (sc-222230) was obtained from Santa Cruz Biotechnology and Digitonin (#300410) was obtained from Calbiochem, E64d (4321-v), and pepstatin A (4397-v) were obtained from Peptide Institute, Inc. human IL-4 (200–04) and GM-CSF (300–03) were obtained from peprotech. Poly(I:C) (tlrl-pic) was obtained from invivogen.
Knockout Huh7 cell lines were established by using CRISPR/Cas9 as described previously [73]. The plasmids of pX330 (#42230) and pCAG EGxxFP (#50716) were obtained from Addgene. The target sequences of each gene are summarized in S1 Table. Genomic DNAs of Huh7 were amplified and cloned into pCAG EGxxFP. Huh7 cells were transfected with pX330 and pCAG EGxxFP and incubated for 1 week. GFP-positive Huh7 cells were sorted by FACS and formed a single colony. Gene deficiency was confirmed by sequencing and Western blotting.
Cell lysates were prepared by incubation in Onyx lysis buffer consisting of 20 mM Tris-HCl (pH 7.4), 135 mM NaCl, 1% Triton X-100, 1% glycerol and protease inhibitor cocktail tablets (Roche Molecular Biochemicals) for 15 min at 4°C. Cell lysates were centrifuged at 13,000 rpm for 5 min at 4°C after sonication. Protein concentrations of supernatants of cell lysates were determined by using Bio-Rad Protein Assay Dye Reagent Concentrate (Bio-Rad). The supernatants were incubated with sample buffer at 95°C for 5 min. The samples (50 μg of proteins) were resolved by SDS-PAGE (Novex, Life Technologies) and transferred onto nitrocellulose membranes (iBlot, Life Technologies). These membranes were blocked with PBS containing 5% skim milk, and then incubated with primary antibody at 4°C overnight. After washing, the membrane was reacted with HRP-conjugated secondary antibody at room temperature for 2 h. The immune complexes were visualized with Super Signal West Femto substrate (Pierce) and detected by an LAS-3000 image analyzer system (Fujifilm).
Total RNA was subjected to electrophoresis in 1.2% agarose-formaldehyde gel and morpholinepropanesulfonic acid buffer. rRNA was visualized by ethidium bromide staining, and electrophoresed RNA was transferred onto a positively charged nylon membrane (Cat No. 11 417 240 001; Roche). Full-length MCL1 cDNA was amplified and cloned into pSPT18 digested by EcoRI using an In-Fusion HD cloning kit (Clontech), an RNA probe was synthesized by using a digoxigenin (DIG) RNA labeling kit (Roche), and hybridization and detection were performed by using a DIG Northern Starter kit (Roche) according to the manufacturer’s protocols. The signals were detected by an LAS-3000 image analyzer system (Fujifilm).
Huh7 cells (2x104 cells) were seeded on 12-well plates (Greiner) and incubated for 1 day. JEV, DENV or ZIKV were infected for 2 h and the culture medium was changed. At each time point, supernatants containing cells and adherent cells were collected and stained with 5 μg/mL of propidium iodide (PI, P4170; Sigma). Cell viability was determined by flow cytometry analyses (BD) and FlowJo software (FlowJo, LLC).
Huh7 and BAX/BAKDKO (#29 and #47) cells were infected with JEV (MOI = 5) and incubated at 37°C for 2 days. Caspase 3/7 activity was determined by using Caspase 3/7Glo (Promega) according to the manufacturer’s protocol.
Cells (2x105) seeded on 6-well plates (Greiner) were infected with lentiviruses expressing BIM, BIM-BADBH3, BIM-NOXABH3 and BIM-4E at one day after incubation, expanded into 10 cm dishes (Greiner) at 2 days post-infection, and cultured with medium containing 1 μg/ml of puromycin for 1 week with a change of culture medium every 2 days. The remaining cells were washed with PBS, incubated with 5 mL of Giemsa’s azur-eosin-methylene blue solution (Merck Milipore: 109204) for 1 h at room temperature, and washed in water.
293T cells (2x106) seeded on 10 cm dishes were transfected with 1.0 μg of either pCAGGS MCL1 or pCAGGS MCL1K/R and 0.05 μg of HA-ubiquitin plasmid by using PEI after 12 h of incubation. After 2 days, cells were treated with proteasome inhibitor (20 μM, ALLN) for 6 h and lysed with buffer A (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 1% sodium deoxycholate, 1% SDS) supplemented with 1 mM sodium fluoride and protease inhibitor cocktail tablets (Roche) on ice for 30 min and sonicated. Cell lysates (200 μl) were diluted with buffer A lacking SDS, incubated with anti-HA antibody (HA.11, 2 μl per sample; Covance) at 4 °C for 90 min and then further incubated with Protein G Sepharose 4B (GE Healthcare) at 4 °C for 90 min. The beads were washed five times in buffer A lacking SDS and eluted in sample buffer after incubation at 95°C for 5 min. Denatured samples were resolved by SDS-PAGE and immunoblotting.
Huh7 and BAX/BAKDKO (#47) cells seeded on 10 cm dishes were infected with JEV (MOI = 5) and treated with ABT-737 (1μM) for 1 h at 37°C. Cells (1.5 x106) were pelleted at 2 days post-infection, resuspended in 100 μl of permeabilization buffer consisting of 250 mM sucrose, 1 mM EDTA, 1 mM EGTA, 5 mM MgCl2, 100 mM KCl, 20 mM Hepes, and 0.025% of digitomin (#300410; Calbiochem) supplemented with protease inhibitor cocktail tablets, and incubated on ice for 5 min. Cell lysates were centrifuged by 13,000 rpm at 4°C for 5 min and the supernatants were collected as a soluble fraction. Pellets were resuspended with Onyx lysis buffer and then incubated on ice for 30 min, and the supernatants were collected as a pellet fraction. Both lysates were incubated with sample buffer at 95°C for 5 min, and denatured samples were resolved by SDS-PAGE and immunoblotting.
Huh7 and BCLXKO Huh7 cells seeded onto 12-well or 24-well plates were inoculated with serially diluted virus and incubated for 2 h. The supernatants were removed and culture medium containing 1% methylcellulose was overlaid onto the cells. After 2 days of incubation, cells were fixed with 4% paraformaldehyde and visualized by staining with methylene blue (Sigma) or Giemsa’s azur-eosin-methylene blue solution (Merck Milipore: 109204) for 1 h at room temperature, and washed in water.
Huh7 cells or JEV-infected cells were seeded on 12-well plates and washed twice with DMEM (deficient in methionine and cysteine; Gibco 21013024), then incubated with [35S] methionine and [35S] cysteine (EasyTag Express35S Protein labeling mix, Perkin Elmer) in DMEM without methionine and supplemented with 10% dialyzed FBS, 100 U/ml penicillin and 100 μg/ml streptomycin for 15 min at 1, 2 and 3 days post-infection. Labeled cells were washed three times with cold PBS, then lysed with 200 μl/well radioimmunoprecipitation (RIPA) buffer, and the cell lysates were centrifuged at 13,000 rpm for 10 min at 4°C. Protein concentrations were determined by using Bio-Rad Protein Assay Dye Reagent Concentrate (Bio-Rad). Clear-sol (900 μl; Nakalai Tesque) was mixed with each supernatant (100 μl), and measured with a liquid scintillation counter (LS6500; Beckman Coulter). For detection of synthesis of MCL1 protein, Huh7 cells were seeded on 10 cm dishes, infected with JEV (MOI = 5) and incubated for 2days. These cells were washed twice with DMEM (deficient in methionine and cysteine; Gibco 21013024), then incubated with [35S] methionine and [35S] cysteine (EasyTag Express35S Protein labeling mix, Perkin Elmer) in DMEM without methionine and supplemented with 10% dialyzed FBS, 100 U/ml penicillin and 100 μg/ml streptomycin for 15 min at 37°C. Cell lysates were prepared by incubation in Onyx lysis buffer for 15 min at 4°C. Cell lysates were centrifuged at 13,000 rpm for 5 min at 4°C. Protein concentrations were determined by using Bio-Rad Protein Assay Dye Reagent Concentrate (Bio-Rad). Same amounts of cell lysates were incubated with 2μL of MCL1 antibody at 4°C for 90 min and then further incubated with Protein G Sepharose 4B (GE Healthcare) at 4°C for 90 min. The beads were washed three times with Onyx lysis buffer, boiled at 95°C for 5 min, and subjected to SDS-PAGE. The 35S-labeled proteins were visualized by using a Fujifilm bioimaging analyzer (FLA-7000, Fujifilm).
RNA was extracted by using ISOGEN II (Nippon Gene), and cDNAs were synthesized from the total RNA using a High Capacity RNA-to-cDNA kit (Applied Biosystems) according to the manufacturer’s instructions. JEV RNA were quantified using a TaqMan RNA-to-Ct 1-Step Kit and the ViiA7 real-time PCR system (Life Technologies). The primers of JEV were 5′-GGGTCAAAGTCATTTCTGGTCCATA’ and 5′-TCCACGCTGCTCGAA’. The following probe were used: for JEV, 50-6-FAM/ATGACCTCG/ZEN/CTCTCCC/-30IABkFQ;. mRNAs of Il6, Tnfα, Ccl5, Cxcl10, Ccl2, Ifna, Ifnb and β-actin were used for quantification of mRNA expression by using a Power SYBR green RNA-to-Ct 1-Step kit and theViiA7 real-time PCR system (Life Technologies). The following primers were used: for Il6 5′-CCACTTCACAAGTCGGAGGCTTA-3′ and 5′-GCAAGTGCATCATCGTTGTTCATAC-3′ Tnfα 5′-CAGGAGGGAGAACAGAAACTCCA-3′ and 5′-CCTGGTTGGCTGCTTGCTT-3′, Ccl5 5′-AGATCTCTGCAGCTGCCCTCA-3′ and 5′-GGAGCACTTGCTGCTGGTGTAG-3′, Cxcl10 5′-ACACCAGCCTGGCTTCCATC-3′ and 5′-TTGGAGCTGGAGCTGCTTATAGTTG-3′, Ccl2 (5′-GCATCCACGTGTTGGCTCA-3′ and 5′-CTCCAGCCTACTCATTGGGATCA-3′, Ifna1 5′-AGCCTTGACACTCCTGGTACAAATG-3′ and 5′-TGGGTCAGCTCACTCAGGACA-3′ Ifnb1 5′-ACACCAGCCTGGCTTCCATC-3′ and 5′-TTGGAGCTGGAGCTGCTTATAGTTG-3′, β-actin 5′-TTGCTGACAGGATGCAGAAG-3′ and 5′-GTACTTGCGCTCAGGAGGAG- 3′.
Human monocyte-derived DCs were generated from healthy human blood donors using Lymphocyte Separation Solution (nacalai tesque). CD14+ cells were isolated after Ficoll-Hypaque gradient centrifugation using MACS CD14+ magnetic beads with FcR Blocking Reagent (Miltenyi Biotech) as manufacturer’s instructions. These human monocytes (5x105 cells) were seeded on 24 well plates and differentiated into naive DCs by culturing for 4 days in differentiated medium, Iscove’s Modified Dulbecco’s Media containing 10% fetal bovine serum (FBS), 100 U/ml penicillin, 100 μg/ml streptomycin 10 μg/ml human IL-4 (PeproTech) and 50 μg/ml human granulocyte-macrophage colony-stimulating factor (GM-CSF) (PeproTech). MDDCs (4 x 105 cells) were infected with DENV (MOI = 3) for 60 min at 37°C. Cells were then pelleted at 2,000 rpm for 5 min and resuspended in 1000 μl of differentiated medium.
Bclx+/- mice were generated as previously described [74] using a BDF1 genetic background. The pX330 containing an sgRNA against the mouse Bclx gene (S1 Table) was injected into mouse zygotes and transplanted into pseudopregnant female mice. The obtained mice (F1) were crossed with wild type mice and F2 mice were analyzed by sequencing. We obtained three Bclx+/- lines from different F1 mice and kept them by crossing with wild type mice.
Peritoneal exudate cells (PECs) were isolated from C57BL/6 (Japan SLC, Inc.) mice intraperitoneally injected with 2 ml of 3% thioglycolate solution by rinsing the peritoneal cavity with 5 ml of PBS. Cells were pelleted by 1,500 rpm at 4 °C for 5 min and suspended in RPMI supplemented with 10% FBS, 100 U/ml penicillin and 100 μg/ml streptomycin, and cultured on 10 cm culture dishes. After 2 h of incubation, adherent cells were collected and seeded in culture medium containing 10 ng/ml human IL-4 (PeproTech), then incubated overnight at 37 °C. Huh7 and BCLXKOHuh7 cells infected with JEV were incubated with pHrodo green (10 μM; Life Technologies) at 37 °C for 30 min and washed twice with culture medium. Labeled-cells were overlaid onto PECs and incubated at 37 °C for 45 min. PECs were washed with culture medium three times to remove non-adherent cells. The amounts of engulfed PECs were determined by FACS (Becton Dickinson). Labeled cells overlaid on PEC were incubated at 37 °C for 2 h and replaced with fresh medium. Infectious titers and cytokine production in the supernatants were determined after 24 h of incubation.
The productions of mouse IL-6 (KMC0061, Novex), TNF-α (KMC3010, Novex), CCL5 (NMR00, R&D Systems), CCL2 (MJE00, R&D Systems) and CXCL10 (BMS6018, Thermo Fisher Scientific) were quantified by ELISA according to the manufacturer’s protocol.
Three-week-old mice were anaesthetized by intraperitoneal administration of 0.75 mg kg-1 medetomidine (Meiji Seika), 4 mg kg-1 midazolam (Sandoz) and 5 mg kg-1 butorphanol tartrate (Meiji Seika). JEV (2x106 FFU/50 μl) was injected subcutaneously into both rear footpads (4x106 FFU in total). JEV (1x105 FFU) was also intracerebrally injected. The survival of the mice was monitored for 25 days. To collect the brain tissues, mice were sacrificed at the end-point and then recorded as dead.
Dosing of ABT-263 was performed as previously described [51,64]. Briefly, ABT-263 was formulated in 10% ethanol, 30% polyethylene glycol 400, and 60% Phosal 50 PG for dosing. The mice were dosed orally once a day for 2 to 8 days with vehicle or 50mg kg-1 ABT-263.
All animal experiments conformed to the Guidelines for the Care and Use of Laboratory Animals, and were approved by the Institutional Committee of Laboratory Animal Experimentation (Research Institute for Microbial Diseases, Osaka University. Project number: H27-06-0). All healthy donors provided oral consent and were properly informed of the risks of the study.
The statistical analyses were performed using GraphPad Prism (GraphPad Prism Software, Inc.). Significant differences were determined using Student’s t-test and are indicated with asterisks (*P<0.05) and double asterisks (**P<0.01) in each figure. Significant differences of in vivo survival data were determined using log-rank test and are indicated with asterisks (*P<0.05) and double asterisks (**P<0.01) in each figure.
|
10.1371/journal.pntd.0001309 | Kinetics of Viremia and NS1 Antigenemia Are Shaped by Immune Status and Virus Serotype in Adults with Dengue | Dengue is a major public health problem in tropical and subtropical countries. Exploring the relationships between virological features of infection with patient immune status and outcome may help to identify predictors of disease severity and enable rational therapeutic strategies.
Clinical features, antibody responses and virological markers were characterized in Vietnamese adults participating in a randomised controlled treatment trial of chloroquine.
Of the 248 patients with laboratory-confirmed dengue and defined serological and clinical classifications 29 (11.7%) had primary DF, 150 (60.5%) had secondary DF, 4 (1.6%) had primary DHF and 65 (26.2%) had secondary DHF. DENV-1 was the commonest serotype (57.3%), then DENV-2 (20.6%), DENV-3 (15.7%) and DENV-4 (2.8%). DHF was associated with secondary infection (Odds ratio = 3.13, 95% CI 1.04–12.75). DENV-1 infections resulted in significantly higher viremia levels than DENV-2 infections. Early viremia levels were higher in DENV-1 patients with DHF than with DF, even if the peak viremia level was often not observed because it occurred prior to enrolment. Peak viremias were significantly less often observed during secondary infections than primary for all disease severity grades (P = 0.001). The clearance of DENV viremia and NS1 antigenemia occurs earlier and faster in patients with secondary dengue (P<0.0001). The maximum daily rate of viremia clearance was significantly higher in patients with secondary infections than primary (P<0.00001).
Collectively, our findings suggest that the early magnitude of viremia is positively associated with disease severity. The clearance of DENV is associated with immune status, and there are serotype dependent differences in infection kinetics. These findings are relevant for the rational design of randomized controlled trials of therapeutic interventions, especially antivirals.
| Dengue is an acute viral disease that affects tens of millions of people annually in tropical and sub-tropical countries. In some cases, this infection happens to be severe and even life threatening. Severe cases have been associated with higher levels of virus in the blood. Several hypotheses have been proposed to explain the occurrence of these cases notably by involving the patient's history of previous DEN virus infection(s). Little is known about the relationships between the evolution over time of virus levels in the blood, the clinical outcome and the previous infection(s) history—a better understanding of these features could help in anti-viral drug development. To analyze these relationships, we studied well characterized patients who participated in a clinical trial. The majority of these patients were infected by DENV-1 serotype and had higher levels of virus than those infected by DENV-2 and sometimes DENV-3 serotypes. We also found that patients with more severe symptoms had higher levels of virus in the first days of their illness. We found as well that the virus was cleared faster and earlier from the blood of patients previously infected. These findings are of major importance for further anti-viral drug testing.
| Dengue viruses (DENVs) are members of the Flavivirus genus and are the most important arboviral pathogens of humans. The four DENVs are antigenically-related and have single-stranded, positive-sense RNA genomes that share 60–70% sequence identity between each others [1]. There are no licensed vaccines to prevent dengue and vector control remains the cornerstone of public health interventions.
The clinical outcome from DENV infection ranges from the asymptomatic to an acute, often debilitating illness called dengue fever (DF), to the severe and potentially life-threatening dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS). The cardinal feature of DHF/DSS is a capillary permeability syndrome characterised by plasma leaking from the vasculature into interstitial spaces. Thrombocytopenia, a coagulopathy and a hemorrhagic diathesis are also common findings. DSS manifests when capillary permeability is severe enough to result in an inadequate intravascular volume that then leads to poor tissue perfusion. DSS is managed, and possibly prevented, by careful restoration and maintenance of the intravascular volume by use of parenteral fluids. Viral strain and host immune status have been suggested as major risk factors for DHF/DSS. In particular, two sequential infections, with the second infection caused by a DENV serotype different from the first, is a risk factor for severe disease in children and adults [2]–[5]. A process called antibody dependent enhancement of infection (ADE), coupled with strong anamnestic cellular immune responses is the leading hypothesis to mechanistically explain more severe disease in secondary infections [6]–[9]. Severe dengue can also occur in primary infection of infants born to dengue-immune mothers with, indicating anamnestic immune responses are not absolutely critical for eliciting the capillary permeability syndrome in all patients. Viral traits may also be important in pathogenesis, with strong evidence that some viral genotypes are fitter than others [10], [11].
The literature describing the overall relationship between plasma/serum viral burden, disease severity and immune status generally supports the hypothesis that there is a positive correlation between markers of viral burden in the first 2–3 days of fever and the severity of clinical outcomes. For example, during the febrile and early convalescent periods, Taiwanese adults with secondary DENV-3 infections and DHF had higher plasma levels of viral RNA than did patients with DF [12]. An association between higher peak viremia and increased disease severity was observed in Thai children with acute DENV-1 and -2 infections [13]. Similarly, DHF was associated with higher plasma viremia early in illness in Thai children with secondary DENV-3 infections [14]. Duyen et al recently showed that DENV-1 infections were associated with higher viremia and NS1 antigenemia than DENV-2 infections in ambulatory Vietnamese paediatric patients [15]. In the same patients, viremia and NS1 antigenemia persisted for longer in patients with primary infections.
In adults, where the risk of clinically apparent disease occurring in primary infection is possibly greater [16], [17], there is less evidence relating virological features of infection to immune status or clinical outcome. Kuberski et al reported in 1977 that the magnitude of viremia in young adult patients was higher in primary than secondary DENV-1 infections [18]. More recently, DENV viremia levels from Taiwanese adult patients were reported to be lower in secondary than in primary DENV-2 infections [19].
The dynamics of virus clearance might also be relevant to clinical outcome. In Thai children, Vaughn and others have shown that the slope of the descending portion of the viremia curve was steeper for patients with secondary infection versus those with primary infection and viremia decreased more quickly for patients with DHF than for patients with DF at defervescence [13]. The accelerated clearance of viremia in secondary infection most likely reflects the contribution of anamnestic humoral and cellular immune responses, which themselves have been implicated in the pathogenesis of capillary leakage. Conversely however, Wang et al suggested clearance of the virus and virus-containing immune complexes was slower in adult DHF patients [20].
A better understanding of the relationship between biomarkers of virus infection, the immune response and disease evolution is critical for the rational use of intervention therapies in dengue, e.g. anti-viral drugs or immune-modulating therapies. To this end, this study describes the kinetics of viremia and NS1 in an intensively investigated cohort of Vietnamese adults with dengue and less than 72 hrs of fever enrolled in a randomized placebo-controlled trial (RCT).
A double blind RCT of chloroquine (CQ) in 307 adults hospitalized for suspected DENV infection was conducted at the Hospital for Tropical Diseases (Ho Chi Minh City, Vietnam) between May 2007 and July 2008. Information on recruitment, inclusion criteria, randomization, treatment and investigations have been published previously [21]. The Scientific and Ethical committee of the HTD and the Oxford Tropical Research Ethical Committee approved the study protocol and all patients gave written informed consent. The trial was registered with the ISRCTN Register (ISRCTN38002730). Herein we describe the clinical and virological features of the 257 patients with laboratory confirmed dengue enrolled into this trial, which hitherto have not been described in detail.
A diagnosis of laboratory confirmed dengue was reached using serological, antigen detection and molecular methods [22]. In brief, RT-PCR detection of DENV RNA in plasma was performed using an internally controlled, serotype-specific, real-time RT-PCR TaqMan assay that has been described previously [23]. RNA extraction from plasma samples was automated (NucliSens easyMAG, BioMerieux, Lyon, France). Results were expressed as cDNA equivalents per mL of plasma. A capture IgM and IgG ELISA (MAC and GAC ELISA) using DENV/JEV antigens and mAb reagents provided by Venture Technologies (Sarawak, Malaysia), was performed as previously described [24]. NS1 was detected by using the NS1 Platelia ELISA assay from BioRad (Hercules, CA) according to the manufacturer's instructions. Samples defined as equivocal in the NS1 Platelia ELISA assay were repeated and if they were still equivocal they were regarded as being negative.
The interpretation of primary and secondary serological responses was based on the magnitude of IgG ELISA units in early convalescent plasma samples taking into account the illness day. The cut-off in IgG ELISA units for distinguishing primary from secondary dengue by illness day was calibrated using a panel of acute and early convalescent sera from Vietnamese dengue patients that were assayed at the Centre for Vaccine Development, Mahidol University, Bangkok, Thailand using a reference IgM and IgG antibody capture ELISA described previously [25].
Clinical history and examination findings were recorded daily into case record forms. An ultrasound was performed in all patients within 24 hrs of defervescence. Venous blood samples were collected at hospital admission, then twice daily (around 9am and 3pm) for a minimum of 5 days after hospital admission and again 10–14 days after discharge from the hospital. A complete blood count, including hematocrit (Hct) and platelet measurements, was performed daily for all patients. Hct measurements were performed more frequently if clinically indicated. The extent of hemoconcentration during symptomatic illness was determined by comparing the maximum Hct recorded during hospitalization with either the value recorded at follow-up when available (191/248 i.e. 77% of the patients) or against a sex- and age-matched population value. Plasma was stored frozen in multiple aliquots at −80°C until use in the real-time RT-PCR and NS1 ELISA. The day of fever onset was self-reported by the patient and was designated illness day 1.
DF and DHF were diagnosed according to 1997 World Health Organization (WHO) classification criteria and was applied to each case after review of study notes [26]. The 1997 definitions were used for this study because at the time of clinical assessment the 2009 WHO Guidelines and revised classification scheme was not available. DF was defined as a laboratory confirmed dengue case with no evidence of capillary permeability as defined for a DHF case. DHF was defined as laboratory confirmed dengue case with thrombocytopenia (<100,000 platelets/mm3), any hemorrhagic manifestation, and evidence of plasma leakage (as denoted by a >20% increase in the Hct from the baseline value or by the presence of pleural or abdominal effusions).
The data used in this analysis was taken from a randomised controlled treatment trial of dengue. Since the intervention (CQ) had no measurable impact on virological or immunological outcomes, for the purposes of this analysis we did not distinguish between patients in the CQ or placebo arms of the study. All statistical analysis was performed and figures designed using the software R (version 2.10.1). Significance was assigned at P<0.05 and were two-sided unless otherwise indicated. Uncertainty was expressed by 95% confidence intervals. The Kruskal-Wallis rank sum test was used for continuous variables and the Fisher's exact test for categorical variables. For the viremia kinetics analysis, when the RT-PCR signal was below the assay limit of detection (defined as the last dilution of standard that gave a specific signal), a value equal to concentration of the last dilution of standard that gave a specific signal divided by 10 was assigned. The maximum viremia level was defined as the highest plasma viremia level measured during illness. The maximum viremia level was considered to be a peak viremia level only in cases in which viremia rose after the enrolment specimen. To compare kinetics of viremia between patients with different serological status, disease severity and serotype, the means of log-transformed viremia measurements made on the same illness day were used as a summary measure of the viremia on that day. To estimate the maximum daily rate of DENV clearance, the slope of the viremia curve was calculated for each illness day as the change in the means of log-transformed viremia measurements made on the same illness day. Only the maximum decreasing daily rate of each patient was used for analysis. Survival analysis using the Kaplan-Meier method and log-rank test was used for all time-to-event outcomes. Time to resolution of viremia or NS1 antigenemia was defined as the time from the start of symptoms until the first of two consecutive plasma samples below the RT-PCR limit of detection or NS1 ELISA negative. The fever clearance time (FCT) was defined as the time from the start of symptoms to the start of the first 48 hours period during which axillary temperature remained below 37.5°C.
Of the 307 adults with suspected dengue enrolled in the CQ RCT between May 2007 and July 2008, 257 had laboratory-confirmed dengue including 248 patients with a defined serological and clinical classification and 9 patients with ambiguous or unknown clinical outcomes or serology (mainly because they left the study prematurely). The characteristics of the study population are summarized in Table 1 (and Table S1). DENV-1 (57.3%) was the commonest serotype detected in this population of patients, then DENV-2 (20.6%), DENV-3 (15.7%) and DENV-4 (2.8%). DHF was significantly associated with secondary infection compared with primary infection (65/215 vs 4/33 i.e. 30.2% vs 12.1%, P = 0.04, Odds ratio (OR) = 3.13, 95% CI 1.04–12.75). DHF resulting from secondary infection was more commonly associated with DENV-2 (21/45 (46.7%)) than for other serotypes (DENV-1: 33/124 (26.6%), DENV-3: 10/33 (30.3%) and DENV-4: 1/7 (14.3%) (DENV-2 vs DENV-1, -3 and -4 P = 0.02, OR = 2.38, 95% CI 1.14–4.96) (Table S1).
Median viremia levels by illness day for DENV-1, -2 and -3 are shown in Figure 1 (and Table S2). In DF patients with primary infection, DENV-1 viremia levels were significantly higher than DENV-2 or DENV-3 levels at multiple time-points during the acute illness (Figure 1A). In DF patients with secondary infection, the most common serological and clinical grouping, and DHF patients with secondary infection, DENV-1 levels were significantly higher than DENV-2 levels and there was also a non-significant trend towards higher DENV-1 levels than DENV-3 levels (Figure 1B and C). Collectively, and despite small sample sizes for some subgroups, these data suggest that DENV-1 infections were associated with higher viremias (as measured by qRT-PCR) than DENV-2, irrespective of disease severity and immune status.
DENV-1 was the commonest serotype detected in this patient population and therefore there was sufficient data to enable direct comparisons of viremia kinetics across serological states and clinical severity whilst controlling for the infecting serotype (Table S2). These data show that in the early acute phase (illness day 3) patients with DENV-1 infection and DHF had significantly higher viremia levels than DENV-1 patients with DF, irrespective of the patient immune status (Figure 2). These data show also that later in the acute phase (from day 4 of illness) patients with primary DENV-1 infections had significantly higher viremia levels than patients with secondary DENV-1 infections, irrespective of the disease severity (Figure 2).
A limitation of these analyses is that in the majority of patients with secondary infections the viremia was already declining at the time of enrolment i.e. we did not observe an obvious peak viremia (Table S3). Overall, a peak viremia was significantly less often observed in secondary infections than in primary infections for all disease severity grades (P = 0.001, OR = 3.64, 95% CI 1.55–8.74). However, there were no significant differences in the duration of illness prior to enrolment between patients in different categories of serological status or disease severity (P between 0.11 and 0.96 if all the serotypes are considered, and 0.16 and 0.39 if and only DENV-1), suggesting this snapshot of viremia levels is unbiased by differences in duration of illness at study enrolment.
Peak viremia levels were identified in 72 patients. Peak viremia occurred significantly earlier in secondary DF than in primary DF (P = 0.008) and in secondary DHF than in primary DHF (P = 0.04) but there were no significant differences between primary DF and primary DHF (P = 0.73) and between secondary DF and secondary DHF (P = 0.13) for the peak viremia time (Table S3). Amongst DENV-1 infected patients, peak viremia levels, when observed (in 51 of 142 DENV-1 infected patients), happened significantly earlier in secondary DF than in primary DF (P = 0.0006) but, possibly because of small sample size, not in secondary DHF compared to primary DHF (P = 0.31). There was no significant difference between secondary DF and secondary DHF (p = 0.48) but there was a non-significant trend towards later viremia peaks in primary DF than in primary DHF (P = 0.052).
There were sufficient observations of the magnitude of peak viremia in DENV-1 infections to look for associations with clinical outcome in this subgroup (Table S3). Peaks of viremia were observed in 9 primary DF, 29 secondary DF, 3 primary DHF and 10 secondary DHF DENV-1 infected patients. Peak viremia levels were not significantly different between DF and DHF patients (DF vs DHF log10 median peak levels 9.89 vs 10.27, P = 0.28) but there was a non-significant trend towards higher peak viremia during secondary infections than primary infections (primary vs secondary P = 0.096 and primary DF vs secondary DF P = 0.086). If considering the highest viremia levels (as distinct from peak viremia levels) in DENV-1 patients, these were significantly higher in DHF than in DF (log10 median levels 9.84 vs 9.19, P = 0.03). Because most DHF cases were associated with secondary infections, for which peak viremia had already past by the time of enrolment, this difference is probably underestimated. These results suggest secondary infections are generally associated with earlier peak viremia but do not provide any conclusive evidence of higher peak viremia levels in DHF and/or secondary infections.
In the 239 patients with detectable viremia, the median of maximum daily rates of DENV clearance (estimated as the slope of the steepest descending daily portion of the viremia curve) was 2.2 log10 per day in primary DF, 2.8 log10 per day in secondary DF, 2.1 log10 per day in primary DHF and 3.0 log10 per day in secondary DHF. The maximum daily rate of clearance was significantly higher in patients with secondary infections (median of the maximum daily loss 2.9 logs per day) versus those who experienced primary infections (median of maximum daily losses = 2.1 logs per day, P<0.00001) (primary DF vs secondary DF P = 0.00004 and primary DHF vs secondary DHF P = 0.025). The results were very similar when considering only DENV-1 patients for analysis (data not shown). These data suggest secondary infection is associated with steeper declines in viremia.
Amongst all viremic patients (n = 239) time to resolution of viremia was significantly longer in primary infections than in secondary infections (hazard ratio (HR) = 2.88, 95% CI 1.79–4.63, log rank test P = 0.000005), in primary DF than in secondary DF (HR = 2.60, 95% CI 1.57–4.32, log rank test P = 0.0001) and in primary DHF than in secondary DHF (HR = 4.92, 95% CI 1.19–20.32, log rank test P = 0.015) (Figure 3A). Median times to resolution of dengue viremia were 148 hrs (IQR 140–173 hrs) in primary DF, 162 hrs (134–>171 hrs) in primary DHF, 120 hrs (97–141.5 hrs) hrs in secondary DF and 123 hrs (113–138 hrs) in secondary DHF.
Amongst DENV-1 infected patients only (n = 142), times to resolution of viremia were also significantly longer in primary infections than in secondary infections (HR = 4.21, 95% CI 2.12–8.35, log rank test P = 0.000009), in primary DF than in secondary DF (HR = 3.67, 95% CI 1.76–7.62, log rank test P = 0.0002) and in primary DHF than in secondary DHF (HR = 7.00, 95% CI 0.94–52.17, log rank test P = 0.03) (Figure 3B). Median times to resolution of DENV-1 viremia were 162 hrs (IQR 144–>176 hrs) in primary DF, >171.1 hrs (134–>179 hrs) in primary DHF (since less than 50% of primary DHF had cleared viremia before discharge), 125 hrs (99–150 hrs) in secondary DF and 127 hrs (107–143.5) in secondary DHF.
Of the 248 patients with defined serological and clinical classifications, there were 214 patients NS1 positive at the time of study enrolment (plus 2 patients negative at enrolment but NS1 positive 24 and 42 hrs later). Consistent with the viremia findings, times to resolution of NS1 antigenemia were significantly longer in primary infections than in secondary infections (HR = 4.57, 95% CI 2.01–10.40, log rank test P = 0.00007), in primary DF than in secondary DF (HR = 3.66, 95% CI 1.47–9.07, log rank test P = 0.003), in primary DHF than in secondary DHF (HR = 7.04, 95% CI 0.97–51.15, log rank test P = 0.02) but also in secondary DF than in secondary DHF (HR = 1.86, 95% CI 1.29–2.67, log rank test P = 0.0007) (Figure 4A). Interestingly, only 5 of 25 patients (i.e. 20%) with primary DF and 1 of 4 patients (i.e. 25%) with primary DHF had cleared NS1 when discharged from hospital. In patients with secondary dengue, 69 of 127 with secondary DF (i.e. 54.3%) and 51 of 60 with secondary DHF (i.e. 85%) had cleared NS1 when discharged from hospital. Median times to resolution of NS1 antigenemia since illness onset were >166 hrs (>146–>178 hrs) in primary DF and >158 hrs (138–>171 hrs) in primary DHF (since less than 50% of primary DF and primary DHF had cleared NS1 before discharge), and 137 hrs (105–>174 hrs) in secondary DF and 121 hrs (100–153 hrs) in secondary DHF.
Amongst DENV-1 infected patients (n = 142), 134 were NS1 positive at the time of study enrolment. Times to resolution of NS1 antigenemia were significantly longer in DENV-1 primary infections than in DENV-1 secondary infections (HR = not applicable, log rank test P = 0.00008), in DENV-1 primary DF than in DENV-1 secondary DF (HR = 3.66, 95% CI 1.47–9.07, log rank test P = 0.003), in DENV-1 primary DHF than in DENV-1 secondary DHF (HR = 7.04, 95% CI 0.97–51.15, log rank test P = 0.004) and in DENV-1 secondary DF than in DENV-1 secondary DHF (HR = 1.86, 95% CI 1.29–2.67, log rank test P = 0.00001) (Figure 4B). Strikingly, none of the DENV-1 infected patients with primary DF (n = 15) or primary DHF (n = 3) had cleared NS1 when they were discharged from hospital. In contrast, 36 of 84 (42.9%) secondary DF and 29 of 31 (93.5%) secondary DHF patients had cleared NS1 when they were discharged. Median times to resolution of NS1 antigenemia since illness onset were >172.5 hrs in primary DF, >171 hrs in primary DHF, >174 hrs in secondary DF (since less than 50% of primary and secondary DF, and primary DHF had cleared NS1 before discharge) and 121 hrs (103–144 hrs) in secondary DHF.
Collectively, these results suggest that DENV infection is cleared earlier and faster in secondary infections than in primary infections.
There were 240 patients febrile at enrolment (plus 2 afebrile patients who developed fever soon after). Overall, FCT were significantly longer in primary infections than in secondary infections (log rank test P = 0.037) but there was no significant difference between primary DF and secondary DF (HR = 1.44, 95% CI 0.93–2.23, log rank test P = 0.096), and between primary DHF and secondary DHF (HR = 1.85, 95% CI 0.67–5.14, log rank test P = 0.23) (Figure 5). Median FCT since illness onset was 131 hrs (IQR 95.5–151.4 hrs) in primary DF, 141 hrs (135.5–160.5 hrs) in primary DHF, 118 hrs (93.1–140.2 hrs) in secondary DF and 120 hrs (105–142 hrs) in secondary DHF. Consistent with the viremia and NS1 findings, these data indicate primary infection was associated with a longer-lived febrile period.
For descriptive purposes, the evolution over time of DENV-1 viremia in the context of white blood cell and platelet counts and percentage hemoconcentration was plotted (Figure 6) and summarised in Table S4. The highest levels of hemoconcentration and the lowest platelet counts occur when the viremia is close to resolution and when the patient is already or very nearly afebrile. The time to platelet nadir and maximum hemoconcentration was shorter in secondary DENV-1 infections (P<0.01). The data also suggest that leucopenia lasts longer in primary infections than in secondary infections.
The interplay between DEN virus infection and host immune status is postulated to play a central role in the pathophysiology of severe dengue. In this current study, we observed important features of this dynamic. First, early viremia levels were higher in patients with DHF, even if the peak viremia level was often not observed because it occurred prior to enrolment. Second, DENV-1 infections manifested as higher and longer-lived viremias, suggesting serotype dependent differences in infection kinetics. Third, the clearance of DENV viremia and NS1 antigenemia occurs earlier and faster in patients with secondary dengue and is also consistent with a faster time to defervescence.
Our findings are in agreement with previous studies that found higher viremias associated with more severe disease [12]–[14]. Our data also suggests that quantitative differences exist between DENV serotypes with respect to the kinetics of viremia and NS1 antigenemia. In particular, DENV-1 infections were associated with higher and frequently longer-lived viremia levels than infections with either DENV-2 or DENV-3. This is in agreement with recent observations in Vietnamese children and adults [15], [27]. DENV-2 was associated with secondary infection and severe disease in our study; this is also in accordance with previous studies [2], [27], [28]. The mechanisms that facilitate relatively higher viremia in DENV-1 infections relative to DENV-2 infections in our study population, and also recently in Vietnamese children [15], are unknown. Plausibly, DENV-1 has an intrinsically faster rate of replication in this patient population and thereby attains a higher infected cell mass in vivo than DENV-2. Clearly, further studies will be needed to explore this.
The duration of NS1 antigenemia was shorter in patients with secondary infections and this is consistent with previous studies that have suggested reduced sensitivity of NS1-based diagnostic tests in patients with secondary infections [22], [29]–[32]. One explanation is that NS1 is less likely to be available for detection when a sufficient level of DENV-reactive IgG (including anti-NS1 IgG) develops during secondary infections. This may serve to mask the antigen from detection in the immunoassay, and/or result in rapid clearance of NS1 in the form of immune-complexes.
Secondary dengue is associated with faster resolution of viremia infection and shorter duration of fever. Interestingly, the daily rates of virus clearance observed in our study were very compatible with those found by Vaughn et al in Thai children [13]. The early adaptive immune response during secondary infection is dominated by populations of memory B and T cells (and possibly memory-like NK cells [33]) and at least some components of this response mediate a strong anti-viral action, as evidence by faster clearance rates of the viremia and the NS1 antigenemia. Clearly however, aspects of this rapid host immune response are clinically deleterious given the epidemiological association between secondary dengue and more severe outcomes, and also the timing of when clinical manifestations of capillary permeability occur. This poses the intriguing question of whether modulating the host immune response (e.g. through early corticosteroid therapy) could achieve both a more gradual clearance of the virus and a host immune response that elicits less pathology.
Assuming blood viremia is a reasonable surrogate of the whole-body virus burden, then the rapid decline of viremia 48–72 hrs after illness onset, especially in secondary infections that carry higher risk for severe outcomes, has implications for rational design of therapeutic pharmacological interventions. An efficacious anti-viral will need favourable pharmacokinetic properties and potency if it is to impact on the viral burden in a rapid and clinically significant manner. Pharmacological targeting (e.g. with corticosteroids) of the host immune response, which accounts for the rapid decline of viremia but which also likely contributes to the capillary permeability syndrome, may equally need to be administered early on in illness e.g. in a prophylactic way, to prevent clinical complications such as DSS. The rapid decline in viremia in secondary infections also highlights the importance of early diagnosis, since early diagnosis will provide the greatest opportunity for an intervention (e.g. an anti-viral), to have an impact. Point of care NS1 diagnostics are available but more can be done to improve their sensitivity [34]. More clinical research is also needed to understand if the sensitivity and specificity of early clinical diagnoses (and prognosis) can be improved, particularly in primary health care settings. Recent literature suggests this is feasible [35], [36].
Current animal models of DENV infection are able to provide for in vivo measurements of anti-viral activity [37], [38]. However these models do not reproduce the temporal changes in virological biomarkers and clinical manifestations seen in naturally infected dengue patients. The lack of concordance between virological and clinical events in small animal models, and what occurs in patients, needs to be carefully considered when evaluating candidate anti-viral drugs for dengue. For instance, the onset of vascular permeability does not follow the disappearance of virus during enhanced DENV infection in mouse models [38], [39].
There are several limitations to our study. Our results are derived from hospitalized patients who are not necessarily representative of patients being seen at the primary health care level. Interestingly however, the same themes identified in this study in hospitalized Vietnamese adults were also observed in Vietnamese children presenting to primary health care level clinics in Ho Chi Minh City [15]. Of the 248 patients in this study with a defined serological and clinical classification, 126 received a 3 day course of CQ and, whilst no virological or immunological effects were detected, an effect of CQ on the clinical phenotype cannot be excluded because vomiting was more frequent in this treatment arm, possibly leading to more dehydration [21]. The majority of the patients in this study were infected by DENV-1. Very few had primary DHF (4 of 248). RT-PCR measurements of viremia in plasma may not be an entirely accurate surrogate of the infected cell mass in vivo, although it is certainly a better surrogate than NS1 antigenemia, which persists well after the febrile period and is heavily influenced by the immune status of the host (i.e. primary versus secondary). Viremia measurements assessed by RT-PCR encompass both infectious and non-infectious viral particles, and the relative proportions may vary between the different serological responses and/or serotypes [40]. Assessment of plasma virus titers based on plaque titration would have been a valid alternative approach to measuring virus concentrations in plasma, however this biological assay is more difficult to standardize and validate relative to a RT-PCR assay. Another general limitation of measuring virus concentrations in plasma is that viruses might be sequestered in other tissues but inaccessible to measurement while still playing a role in disease pathogenesis.
Our study emphasizes the importance of the period before and just after the onset of fever. This and other studies have established that early viremia levels are associated with disease severity, although they are very clearly not the only determinant of outcome. Very little is known of the virological events in the hours preceding and shortly after fever onset, mainly because this is very difficult to investigate without a good experimental model. An interesting insight was provided by clinical trials of DENV-1, -3 and -4 monovalent live attenuated vaccines in the 1980s [41]–[43]. These vaccines were not sufficiently attenuated and some volunteers developed dengue fever. These studies suggest that viremic period starts several days before the onset of symptoms This presymptomatic viremic period should not be underestimated because of its possible contribution to DENV transmission to uninfected mosquitoes. In our study, we also observed prolonged times of virus clearance in primary dengue, and long-lived higher viremia levels in DENV-1 infections. These might lead to a higher possibility of human to mosquito virus transmission by maintaining viremia above the infectious level over a longer period of time.
Collectively, our findings reveal important patterns of the viremia and NS1 antigenemia kinetics according to the patient immune response, disease severity and virus serotype, and may help for the rational design of clinical trials of therapeutic interventions, especially antivirals.
|
10.1371/journal.pgen.1007414 | The complex geography of domestication of the African rice Oryza glaberrima | While the domestication history of Asian rice has been extensively studied, details of the evolution of African rice remain elusive. The inner Niger delta has been suggested as the center of origin but molecular data to support this hypothesis is lacking. Here, we present a comprehensive analysis of the evolutionary and domestication history of African rice. By analyzing whole genome re-sequencing data from 282 individuals of domesticated African rice Oryza glaberrima and its progenitor O. barthii, we hypothesize a non-centric (i.e. multiregional) domestication origin for African rice. Our analyses showed genetic structure within O. glaberrima that has a geographical association. Furthermore, we have evidence that the previously hypothesized O. barthii progenitor populations in West Africa have evolutionary signatures similar to domesticated rice and carried causal domestication mutations, suggesting those progenitors were either mislabeled or may actually represent feral wild-domesticated hybrids. Phylogeographic analysis of genes involved in the core domestication process suggests that the origins of causal domestication mutations could be traced to wild progenitors in multiple different locations in West and Central Africa. In addition, measurements of panicle threshability, a key early domestication trait for seed shattering, were consistent with the gene phylogeographic results. We suggest seed non-shattering was selected from multiple genotypes, possibly arising from different geographical regions. Based on our evidence, O. glaberrima was not domesticated from a single centric location but was a result of diffuse process where multiple regions contributed key alleles for different domestication traits.
| For many crops it is not clear how they were domesticated from their wild progenitors. Transition from a wild to domesticated state required a series of genetic changes, and studying the evolutionary origin of these domestication-causing mutations are key to understanding the domestication origins of a crop. Moreover, population comparisons provide insight into the relationship between wild and cultivated populations and the evolutionary history of domestication. In this study, we investigated the domestication history of Oryza glaberrima, a rice species that was domesticated in West Africa independent from the Asian rice species O. sativa. Using genome-wide data from a large sample of domesticated and wild African rice samples we did not find evidence that supported the established domestication model for O. glaberrima—a single domestication origin. Rather, our evidence suggests the domestication process for African rice was initiated in multiple regions of West Africa, caused potentially by the local environments and cultivation preference of people. Hence domestication of African rice was a multi-regional process.
| Domestication of crop species represents a key co-evolutionary transition, in which wild plant species were cultivated by humans and eventually gave rise to new species whose propagation were dependent on human action [1–3]. The evolutionary origin(s) of various crop species have been the subject of considerable interest. Studying it has broadened our understanding of the early dynamics associated with crop species origins and divergence, the nature of human/plant interactions, and the genetic basis of domestication. Moreover, an understanding of the evolutionary history of crop species aids genetic mapping approaches, as well as informs plant breeding strategies.
Within the genus Oryza, crop domestication has occurred at least twice—once in Asia and separately in Africa. In Asia, the wild rice O. rufipogon was domesticated into the Asian rice O. sativa approximately 9,000 years ago [4]. In West Africa, the wild rice O. barthii was independently domesticated into the African rice O. glaberrima about 3,000 years ago [4]. Recent archaeological studies have also suggested that a third independent domestication event occurred in South America during pre-Columbian times, but this crop species is no longer cultivated [5].
The domestication history of Asian rice has been extensively studied both from the standpoint of archaeology [6] and genetics [7]. In contrast, much less is known about the domestication of O. glaberrima. Based on the morphology of rice grown in West Africa, the ethnobotanist Portères was the first to postulate an O. glaberrima domestication scenario [8,9], in which the inner Niger delta region in Mali as the center of domestication (Fig 1). He based this hypothesis on O. glaberrima in this area predominantly having wild rice-like traits (termed “genetically dominant characteristics” by Portères), observing loosely attached spikelets, reddish brown pericarps, and anthocyanic pigmentation. In contrast, O. glaberrima with domesticated rice-like traits (termed “genetically recessive characteristics” by Portères) were found in two geographically separated regions: (i) the Senegambia region bordering the river Sine to the north and river Casamance to the south, and (ii) the mountainous region of Guinea. Portères hypothesized the derived traits observed in O. glaberrima from Senegambia and Guinea were due to those regions being secondary centers of diversification, but the inner Niger delta region remained as the primary center of diversity for African rice. Initial archaeological excavations found ceramic impressions of rice grains in north-east Nigeria dating ~3,000 years ago, but first evidence of documented O. glaberrima has been found in the inland Niger delta at Jenne-Jeno, Mail dating ~2,000 years ago [10]. A more recent find at an excavation at the lower Niger basin north of Benin found O. glaberrima dating to ~1,600 to ~1,100 years ago, suggesting domesticated African rice had spread down the Niger river by this time from the inland Niger delta region [11].
Few population genetic studies have attempted to understand the evolutionary history and geographic structure of O. glaberrima. Microsatellite-based analysis showed genetic structure within O. glaberrima [12], suggesting the phenotypic differences observed by Portères may have stemmed from this population structure. With high-throughput sequencing technology, population genomic analysis indicated O. barthii, the wild progenitor of O. glaberrima, had evidence of population structure as well, dividing it into 5 major genetic groups (designated as OB-I to OB-V) [13]. The OB-V group from West Africa was most closely related to O. glaberrima, which caused previous researchers to suggest that this O. barthii group from West Africa was likely to be the direct progenitor of African rice [13]. Genome-wide polymorphism data also indicated that O. glaberrima had a population bottleneck spanning a period of >10,000 years, indicating a protracted period of pre-domestication related management during its domestication [14]. A recent study based on simulations detected population expansion after the bottlenecking event, to originate from the inland Niger delta suggesting the origin of O. glaberrima to be in this region [15].
While previous genome-wide variation studies have given valuable insights into the evolutionary history of O. glaberrima, they have not necessarily examined how O. glaberrima was domesticated from O. barthii. This is because the domestication history of a crop is best examined from the pattern of variation observed in genes underlying key domestication phenotypes [2]. Crop domestication accompanies a suite of traits, often called the domestication syndrome [16], which modified the wild progenitor into a domesticated plant dependent on humans for survival and dispersal [1]. In rice, these traits include the loss of seed shattering [17,18], plant architecture change for erect growth [19,20], closed panicle [21], reduction of awn length [22,23], seed hull and pericarp color changes [24,25], change in seed dormancy [26], and change in flowering time [27]. During the domestication process, it is likely that these traits were not selected at the same time and selection would have occurred in subsequent stages. Traits such as loss of seed shattering and plant erect growth would have been among the initial phenotypes humans have selected to distinguish domesticates from their wild progenitors. On the other hand, traits that improved taste and appearance of the crop, or adaptation to the local environment would likely have been favored in later diversification/improvement stages of crop evolution [3,28].
Genes involved in the early stage domestication process are key to understanding the domestication process of a crop. Sequence variation from these early stage domestication genes can indicate whether a specific domestication trait had single or multiple causal mutations, revealing whether domestication has a single or multiple origin. The geographic origin and spread of domestication traits can be inferred from sequence variation in domestication loci within contemporary wild and domesticate populations [17,29–32]. In Asian rice, for example, genome-wide single nucleotide polymorphisms (SNPs) have suggested that each rice subpopulation had independent wild rice populations/species as their progenitors [33–37], but the domestication genes revealed a single common origin of these loci [35], suggesting a single de novo domestication model for Asian rice [37–42]. On the other hand, the domestication gene for the non-brittle phenotype (btr1 and btr2) in barley had at least two independent origins [43,44], likely from multiple wild or proto-domesticated individuals [45]. This suggests barley follows a multiple domestication model [46–48] originating from multiple ancestral population [45,49,50].
To better understand the domestication of O. glaberrima, we have re-sequenced whole genomes of O. glaberrima landraces and its wild progenitor O. barthii from the hypothesized center of origin in the inner Niger delta, the middle and lower Niger basin that includes the countries Niger and Nigeria, and from Central Africa which includes Chad and Cameroon. The latter two regions were not heavily sampled in previous genomic studies. Together with published O. barthii samples from West Africa [13] and O. glaberrima samples from the Senegambia and Guinea region [14], we conducted a population genomic analysis to examine the domestication history of O. glaberrima. The domestication history were further examined from the evolutionary analysis of genes involved in the early stage domestication process, mainly in the traits involving loss of shattering and erect plant growth. To complement the inferred domestication history, we measured panicle threshability, an important early domestication trait associated with seed non-shattering, from our O. glaberrima samples to further elucidate the domestication history of O. glaberrima. With our data we examine the evolutionary and population relationships between O. glaberrima and O. barthii, the demographic history, and the geographic origin(s) of domestication of the African rice O. glaberrima.
We re-sequenced the genomes of 80 O. glaberrima landraces from a geographic region that spanned the inner Niger delta and lower Niger basin region (S1A Fig). Together with 92 O. glaberrima genomes that were previously re-sequenced [14], which originated mostly from the coastal region (S1B Fig), the 172 O. glaberrima genomes analyzed in this study represent a wide geographical range from West and Central Africa. We also re-sequenced the genomes of 16 O. barthii samples randomly selected from this area, which includes the areas from coastal west Africa, inner Niger delta, and the lower Niger basin (S1C Fig). These were analyzed together with the 94 O. barthii genomes that were previously re-sequenced [13].
The average genome coverage in the data set we gathered for this study was ~16.5× for both domesticated and wild African rice samples, and is comparable to the sequencing depth (~16.1×) in our previous study. The Wang et al. [13] study sequenced a subset of their samples to a higher depth (~19.4×), although the majority of their samples had relatively low coverage (~3.9×) (see S1 Table for genome coverage of all samples in this study). To avoid potential biases in genotyping that arises from differences in genome coverage [51,52], we conducted our population genetic analysis using a complete probabilistic model to account for the uncertainty in genotypes for each individual [53,54]. For the subset of our analysis that required genotype information for each sample, we employed SNPs called from individuals with greater than 10x genome-wide coverage. After quality control filtering, we identified a total of 634,418 and 1,568,868 post-filtered SNPs from the non-repetitive regions of the O. glaberrima and O. barthii genomes, respectively.
The genetic structure across domesticated and wild African rice was examined by estimating the ancestry proportions for each individual in our dataset. We employed the program NGSadmix [55], which uses genotype likelihoods from each individual for ancestry estimation and is based on the ADMIXTURE method [56]. Ancestry proportions were estimated by varying the assumed ancestral populations (K) from 2 to 9 groups (S2 Fig).
With K = 2 NGSadmix divided the data set into O. glaberrima and O. barthii species, with several O. glaberrima samples having varying degrees of O. barthii ancestry (ranging from 4.5 to 40.5%). Interestingly, there were a number of O. barthii samples that had high proportions of O. glaberrima ancestry. All these wild rice with discernible O. glaberrima admixture corresponded to the designated OB-V O. barthii group and hypothesized progenitor of O. glaberrima from Wang et al. [13]. However, our ancestry analysis suggests this wild O. barthii group could also be a result of either wild-domesticated rice hybridization or mislabeling of O. glaberrima as O. barthii (see below) [57].
Increasing K further subdivided O. glaberrima into subpopulations that had a geographical basis (see S2 Fig for all K and their geographic distribution). For instance, K = 3 divided the O. glaberrima into two major subpopulations, first a coastal population that includes the Senegambia and Guinea highland region, and second an inland population that includes the inland Niger delta and lower Niger river basin region (Fig 2).
At K = 5, there were three major genetic groups within O. glaberrima and two within O. barthii. The two O. barthii genetic groups corresponded to OB-I and OB-II group identified in Wang et al. [13]. For O. glaberrima, the ancestry proportions showed structuring into 3 major geographic regions: coastal, inner Niger delta, and lower Niger basin populations (Fig 2).
At K = 7, O. glaberrima were divided into 5 genetic groups where the coastal and inner Niger delta population were further divided into northern and southern genetic groups. It is also at K = 7 where O. glaberrima divided into genetic groups that are consistent with Portères observation—that the coastal population is divided into a Senegambia or Guinea highland genetic cluster, while the samples closest to the inner Niger delta forms a unique genetic cluster (Fig 2).
At K = 9, O. barthii is separated into the three genetic groups OB-I, OB-II, and OB-III that were previously identified from Wang et al. [13]. For O. glaberrima the lower Niger basin population divided into two geographic regions, where the samples closer to the inner Niger delta formed its own genetic cluster (Fig 2). Importantly, what is noticeable with increasing K is that populations appeared to be separating into smaller, and more highly localized geographical clusters.
We then conducted phylogenomic and principal component analysis (PCA) to verify our ancestry proportion results. Phylogenomic analysis were conducted using genotype likelihoods to estimate the pairwise genetic distances [58] and build a neighbor-joining tree (Fig 3A). O. glaberrima formed a paraphyletic group relative to several O. barthii individuals. We noticed that O. glaberrima landraces could be divided into 5 phylogenetic groups sharing a common ancestral node. Although the bootstrap support on the five ancestral nodes were weak, the geographic distribution of these 5 phylogenetic groups (Fig 3B) were concordant with the geographic distribution of the ancestry components in the O. glaberrima subpopulations identified at K = 7 (Fig 2, note at K = 7 O. glaberrima forms five major genetic clusters while O. barthii forms two major genetic clusters). The 5 phylogenetic groups clustered into five geographic locations: north and south coastal population, north and south inland Niger delta population, and a lower Niger basin population.
The O. glaberrima population genotype likelihoods were also used for principal component analysis (PCA), which visualized the population relationships [54]. For the PCA plot, individuals were color coded according to the grouping status determined from the phylogenomic results (Fig 3A and Fig 3B). When color-coded according to the phylogenomic tree grouping, PCA results showed 5 independent clusters for O. glaberrima (Fig 3C). In addition, the distribution of individuals along the two principal components showed striking similarity with their geographic distribution (Fig 3B versus Fig 3C).
Together, our analyses of ancestry components, phylogenomics, and PCA suggest O. glaberrima has a geographically based population structuring with at least 5 subpopulations (Fig 3A). Consistent with the hypothesized Guinea highland and Senegambia populations, the coastal populations were divided into OG-A1 and OG-A2 genetic groups (collectively the OG-A supergroup). The lower Niger basin and central African individuals formed as a single OG-B group. Finally, for the inner Niger Delta region, landraces closest to the delta formed the OG-C1 group while the others formed the OG-C2 group; collectively these represent the OG-C supergroup.
We note there are several methods of testing and choosing the most appropriate number of genetic clusters (K) for a population sample [56,59–62]. However, these statistical tests can be misleading [63,64], often prompting overconfidence in a single K value that may or may not be biologically relevant. Thus, we emphasize our choice of dividing O. glaberrima into five major groups represent the minimum possible grouping based on historical observations [8,9] and geography (Figs 2 and 3). We also find that these 5 groups had significant correlations with the geographical distributions of domestication gene mutations and phenotypes (see below domestication gene analysis for more detail), further suggesting they represent biologically relevant groupings.
At K = 7 the majority of the newly sequenced O. barthii from this study belonged to either OB-I or OB-II subpopulations designated by Wang et al. [13] (S3 Fig). The ancestry proportion for the OB-III and OB-IV groups suggested these individuals were an admixed group, with OB-III an admixture of OB-I and OB-II, and the OB-IV group possessing a mix of ancestry from both wild and domesticated rice. Note that at higher K values, OB-III formed its own genetic cluster while OB-IV showed ancestry with large proportions from wild and domesticated rice. Unlike the OB-V group of O. barthii, which also had several individuals of mixed wild and domesticated rice ancestry, the OB-IV group did not phylogenetically cluster with O. glaberrima (Fig 3 and S3 Fig). This suggests that the OB-IV subpopulation may be an evolutionary distinct population, and the ancestry proportions were possibly mis-specified [64]. Hence, we considered individuals that were monophyletic with the OB-I or OB-II subpopulations as the wild O. barthii subpopulation and henceforth designated it as OB-W [= wild] (Fig 3A). O. barthii that were paraphyletic with O. glaberrima were considered as a separate O. barthii group and designated as OB-G [= glaberrima-like] (Fig 3A). Geographically, OB-G was found throughout West Africa but OB-W was found mostly in inland West African countries such as in Mali, Cameroon, and Chad (S2 Table).
Before examining the relationships between our 5 inferred genetic clusters for O. glaberrima, we filtered individuals with spurious classification.
First, we find that there were 3 O. glaberrima individuals (IRGC104883, IRGC105038, and IRGC75618) that did not group with any of the 5 population groups (Fig 3A white arrow), but rather formed as a sister group to all O. glaberrima or sister group to both OG-A and OG-B group. Because they were most closely related to OB-G samples we considered them as OB-G as well. Interestingly, there were also two O. glaberrima individuals (IRGC103631 and IRGC103638) that phylogenetically clustered with O. barthii (Fig 3A filled arrows). Ancestry estimates for the two samples showed high proportions of both O. glaberrima and O. barthii ancestry (S4 Fig). These two O. glaberrima samples were not used in subsequent analyses. Moreover, all O. barthii samples not grouped as OB-W or OB-G, as discussed above, were excluded from downstream analysis.
Second, we examined other potentially spuriously grouped individuals by calculating silhouette scores for each individual [65], which measures similarity with members of its own group compared to members of other groups (see Materials and Method for details). Initially, 174 O. glaberrima samples with greater then 10× coverage were used for genotyping. A multidimensional scaling (MDS) plot of the population and their grouping status showed that even before the silhouette score-based filtering, there were clear separation among the OG-A, OG-B, and OG-C groups (S5A Fig). But there were also several individuals whose status was questionable, as they overlapped in coordinate space with other groups. Individuals with negative silhouette scores (i.e. potential mis-grouping) or scores lower then 0.12 (i.e. individuals with significant portions of ancestry coming from a different genetic group) were filtered out (S5B Fig) to remove individuals with questionable grouping status and thus specify genetically unique populations [66]. We note some individuals that were filtered from the silhouette score-based method were likely filtered because they are admixed individuals. Omission of those individuals would lead to an underestimation of the recent admixture history of O. glaberrima. Here, our interest is in determining the long-term population histories that shaped each O. glaberrima population; hence, removal of those recently admixed individuals are necessary.
This last filtering process resulted in 94 individuals, which we refer to as the core set population (see S3 Table for list of accessions). MDS plot of this core set population showed clear separation among each other (S5C Fig), suggesting these are genetically distinct populations (S5D Fig). This core set population was used to infer the population relationship within O. glaberrima. To determine the population relationships, we also included polymorphism data from the OB-W group individuals with greater then 10× coverage. Because our grouping is based on K = 7 ancestry (Fig 2), which had two population groups for OB-W, we divided the OB-W group into two (OB-W1 and OB-W2) based on the common ancestor they shared in the phylogenomic tree (Fig 3A). For an outgroup population, polymorphism data from six O. rufipogon individuals with greater then 10× coverage were used [35]. An MDS plot of the nucleotide variation showed clear separation among the three species and separation within species depending on the population grouping status (S6 Fig).
Using the core set population, we inferred the population relationships between the five genetic groups of O. glaberrima with Treemix [67]. For the graph rooted with O. rufipogon population, without modeling any migration events the OB-W1 group were sister to all O. glaberrima (Fig 4). This model without any migration events was able to explain 99.4% of the variance, suggesting most of the allele frequency variability in the data can be explained without evoking migration between groups. Nevertheless, residuals from the covariance matrix suggested several population pairs could be more closely related (population pairs with positive residuals) compared to the best-fitting tree.
Fitting models with 1, 2, and 3 migration events brought marginal increase in the variance explained for each migration model (variance increases as 99.8%, 99.9%, and 99.94% respectively). Fitting 1 and 2 migration events suggested an admixture event between a population ancestral to OB-W1 and modern OG-A1 or OG-B (Fig 4). This suggests an unsampled O. barthii population may have admixed with OG-A1 and/or OG-B population. The first within-O. glaberrima admixture, specifically between OG-A1 and OG-B, was observed in the model fitting 3 migration events (S7 Fig). But the f4 test [68] indicated no significant deviation from non-admixed topology for the tree [[wild rice, OG-B],[OG-A1,OG-A2]], suggesting the 3 migration model is an overfitted model (see S4 Table for f4 test result).
Collectively, our analysis suggests the O. glaberrima population could be modeled as a bifurcating tree-like population, with small ancient admixture events from O. barthii genetic groups. Here then, it is tempting to interpret the O. glaberrima population topology, specifically the order of splitting of each genetic group, as the order of the domestication/diversification events. However, we should note that the topology changes with and without modeling migration, and in higher migration models several population pairs (e.g. based on the residuals between OG-C1 and OB-W2) are still not well fitted, while bootstrap support for several internal branches are low. Thus, while this analysis provides an initial framework for depicting population relationships, one should exercise caution in over-interpreting the inferred trees.
Previous molecular studies have argued the close genetic affinities of some west African O. barthii (namely the OB-G group in this study) to O. glaberrima, as evidence of the former being the progenitor population of African rice [13,69]. We thus examined the properties of the OB-G group in relation to OB-W and O. glaberrima.
First, we found that the level of population differentiation between OB-G and O. glaberrima was low (~ 0.06) (Fig 5A), almost comparable to the level seen between O. glaberrima genetic groups (~0.09). In contrast, there is a higher level of differentiation between each O. glaberrima genetic group and OB-W (Fst ~ 0.26). Similarly, OB-G group also had high levels of differentiation to OB-W (Fst ~ 0.21).
Second, we examined levels of linkage disequilibrium (LD) decay, as wild and domesticated populations have different LD profiles, due to the latter undergoing domestication-related bottlenecks and selective sweeps [70]. In the African rice group, as expected, all O. glaberrima genetic groups had higher levels of LD compared to the OB-W group (Fig 5B). The OB-G group also had high levels of LD that was comparable to those observed in O. glaberrima, although compared to other OG groups, the OB-G group had longer tracts of LD.
Finally, genome-wide polymorphism levels for the OB-G group were also comparable between OB-G and O. glaberrima. Specifically, compared to the OB-W group, SNP levels and Tajima’s D [71] were significantly lower in both OB-G and O. glaberrima (S8 Fig).
Together, the levels of genetic differentiation, linkage disequilibrium, SNP levels and patterns, all suggest that the OB-G genomes resemble O. glaberrima more than O. barthii. Furthermore, the majority of the OB-G samples carried at least one domestication mutation (see domestication gene haplotype analysis section for detail), calling into question its status as the wild progenitor. In contrast all OB-W individuals do not carry the causal mutation/deletion at known domestication genes. All in all, this suggests the OB-G population is actually O. glaberrima that was mislabeled as O. barthii. It is also possible that this population may represent feral weedy rice [72], resulting from the hybridization of domesticated and wild African rice; this is certainly consistent with the increased LD structure within OB-G [73]. While the different demographic histories between the source populations can generate an overall negative Tajima’s D for the resulting admixutre population [74]. Together, our results suggest that OB-G may have formed after the domestication event and supports a de-domestication (endoferality) origin for that group [57].
To further identify the domestication origin(s) of O. glaberrima, we examined the haplotypes for the domestication genes involved in erect plant growth (PROG1) and the non-shattering phenotype (sh4 and sh1) in both wild and domesticated African rice. We took an approach we term functional phylogeography, where we examined the haplotype structure surrounding the domestication gene of interest [29], inferred a haplotype phylogenetic network, and determined the geographic origin and spread of the functional mutation by comparing the geographic distributions of haplotypes in wild and domesticated African rice in a phylogenetic context. Because we focused on the non-recombining region surrounding a domestication gene, there were only a few sites being analyzed between O. glaberrima and O. barthii. However, we were specifically interested in those few mutations that differ between the domestication gene haplotype and the progenitor gene haplotype, and used those differences to build the haplotype phylogenetic network.
The PROG1 gene was first identified as a domestication gene in the Asian rice O. sativa. A mutation in this gene causes the plant to grow erect in both Asian and African rice, increasing growing density and enhancing photosynthesis efficiency for higher grain yields [19,20,75,76]. Our analysis of O. sativa PROG1 gene orthologs in O. glaberrima and O. barthii indicates that this gene is missing only in O. glaberrima (S5 Table). We expanded the analysis to our population dataset, and a sequencing read depth analysis found PROG1 was missing in all O. glaberrima landraces (Table 1). None of the OB-W individuals had the PROG1 deletion and all but two of the OB-G individuals had the PROG1 deletion. Synteny of the genes immediately surrounding O. sativa PROG1 was maintained in both O. glaberrima and O. barthii, suggesting the PROG1 gene is deleted specifically in O. glaberrima. Because of its importance in early domestication and lack of gene structure in O. glaberrima, we considered PROG1 as a candidate domestication gene in African rice and examined the population genetics of the PROG1 gene in O. glaberrima and O. barthii. We note this is the first candidate domestication gene that has been identified where the causal mutation is fixed in all O. glaberrima population.
We first examined whether the PROG1 locus showed evidence for positive selection in O. glaberrima, using genome-wide sliding window analysis of the ratio of polymorphism between the wild OB-W group to all domesticated O. glaberrima (πw/ πD). A domestication-mediated selective sweep would lead to a reduction in nucleotide variation around the target domestication gene, but only within the domesticated group. Because PROG1 is deleted in O. glaberrima, the selection signal will only persist around the candidate deletion region. Spanning 10 kbp of the candidate deletion region, πw/ πD is within the top 1% value, and this is observed regardless of whether the O. glaberrima or O. barthii genome was used as the reference genome in SNP calling (S9 Fig). This is consistent with the PROG1 region having gone through a selective sweep during O. glaberrima domestication. Cubry et al. [15] has also independently found evidence of a selective sweep in the PROG1 region of O. glaberrima, supporting our finding that this region has been a target of domestication-related selection.
Polymorphisms surrounding the PROG1 deletion comprised a single unique haplotype segregating across all O. glaberrima samples and most of the OB-G samples (Fig 6A). A haplotype network of a non-recombining 5 kbp region immediately upstream of the deletion showed that all individuals with the deletion belonged to the same major haplotype group, with the dominant haplotype I, as well as peripheral haplotypes III, VII, and VIII (Fig 6D). Maximum-likelihood tree of the region surrounding PROG1 collapsed all O. glaberrima into a single phylogenetic group (S10A Fig), which suggests a single origin for the deletion. We tabulated the geographic distributions of PROG1 haplotypes (S6 Table). PROG1 haplotype VII is the earliest haplotype with the deletion and is found in an OB-G individual from Cameroon. The ancestral non-deleted PROG1 haplotype was carried by haplogroup IV (Fig 6D), which was most closely related to all haplotypes with the PROG1 deletion, and was made up of three OB-W individuals: IRGC103912 (Tanzania), IRGC105988 (Cameroon), and WAB0028882 (Cameroon). The downstream region of the deletion was consistent with what is observed in the upstream region (S11 Fig). Twenty-two polymorphic sites from a non-recombining 7 kbp downstream region show the same OB-W individuals (IRGC105988 and WAB0028882), both from Cameroon, were the most closely related haplotype to the PROG1 deletion haplotype. Maximum-likelihood trees of both the upstream and downstream regions also showed these two individuals to be the sister group to all O. glaberrima samples (S10A Fig).
Together, the geographic distribution of the PROG1 region haplotypes suggest that the PROG1 deletion may have occurred in a wild progenitor closely related to those found in Cameroon, Central Africa, and spread throughout West Africa to the different O. glaberrima genetic groups (Fig 6G). The PROG1 conclusion must be tempered, however, by an acknowledgment that the sample size of ancestral haplotypes is small (n = 3). Interestingly, a similar observation has been made in O. sativa where all Asian rice subpopulations are monophyletic in the PROG1 region, but genome-wide the different Asian rice variety groups/subspecies do not share immediate common ancestors [35,77].
Evidence for a selective sweep around the causal domestication mutation, a C-to-T nonsense mutation at position 25,152,034 leading to a loss-of-function allele (Fig 6B arrow), has been previously shown [78,79] for the sh4 gene (O. glaberrima chromosome 4:25,150,788–25,152,622). The haplotype structure around the sh4 gene showed most of the O. glaberrima landraces carried the causal domestication mutation (Fig 6B). Several individuals within OG-A1 group, including the reference genome, did not carry the causal mutation but still had long tracks of homozygosity at the sh4 locus (Fig 6B star).
A four-gamete test [80] of the 4 kbp upstream and 2 kbp downstream region spanning the sh4 gene detected evidence of recombination, within the O. barthii population (both OB-G and OB-W) and but not within O. glaberrima. A maximum-likelihood tree of the region surrounding sh4 showed all O. glaberrima populations were divided into two major phylogenetic groups, but with weak bootstrap support (S10B Fig). O. glaberrima individuals without the causal mutation (Fig 6B star) formed their own phylogenetic group (S10B Fig star).
To determine the origin of the non-shattering trait, we reconstructed the haplotype network of the non-recombining region of the sh4 gene in all O. glaberrima and O. barthii genetic groups (Fig 6E). Majority of the O. glaberrima and OB-G group sh4 haplotypes belonged to haplotypes II, VI, and XIII and they all shared the nonsense mutation. The two main haplotypes II and VI corresponds to the difference observed in the upstream region of the sh4 gene (Fig 6B), with haplotype II arising prior to haplotype VI. The closest haplotype to II was haplotype I, which was separated by two mutations (position 25,146,871 and the causal domestication mutation 25,152,034).
We tabulated the geographic distributions of O. glaberrima haplotypes II and VI/XIII, and haplotype I from the O. barthii OB-W group (Fig 7). The ancestral haplotype I is found in 13 O. barthii individuals (4 OB-G group and 9 OB-W group), and these individuals originated over a wide geographic region of West Africa that includes both coastal and inland areas (See S8 Table for full list of members of each haplogroup and their country of origin). Of those in OB-W, 2 are from Mali, 2 from Nigeria and 5 are from Cameroon. Among the O. glaberrima that have the sh4 mutation, the older haplotype II is found mostly in Mali, Burkina Faso and also Guinea.
Here, we made the assumption that the areas of overlap between the ancestral haplotype (without the causal mutation) and the derived haplotype (with the causal mutation) is likely the place of origin of the domestication allele. For sh4, the distribution of haplotype II overlaps with haplotype I in Mali, pointing to Mali as being a likely place of origin for the sh4 nonsense mutation (Fig 6H). The haplotypes VI and XIII thus subsequently evolved from haplotype II, which expanded over a much wider area, particularly in the Senegambia, and also to Nigeria, Cameroon and Chad. It should be noted that the sample size for haplotype I among OB-W is relatively small (n = 9) leading to disjoint geographic ranges for its distribution (Fig 6H). Localizing the origin of the sh4 causal mutation to Mali may be revised as more O. barthii samples are analyzed. However, haplotype II is found at highest frequency in Mali as well (~46%, see Fig 7), which provides further support for a Malian origin of the sh4 mutations.
Wu et al. [79] had first noticed that several O. glaberrima individuals in the coastal region of West Africa did not have the causal domestication mutation in the sh4 gene (Fig 6B star). Our data shows that all inland O. glaberrima carries the haplotype with the nonsense mutation, and the haplotype without the nonsense mutation was indeed limited to the coastal region, specifically in the OG-A1 genetic group. The haplotype network and neighbor-joining tree suggests these individuals had distinct evolutionary histories for the sh4 gene (Fig 6E and S10B Fig); they carry haplotype VII which is confined to Guinea. The non-fixed status of the nonsense mutation suggests a role of independent mutation(s) in domestication for non-shattering in haplotype VII carriers.
Our results showed the causal domestication mutations for the shattering genes sh1 and sh4 were not fixed in several O. glaberrima varities, suggesting their seed non-shattering may be incomplete. Thus we examined the phenotypic consequence of the domestication-related selection process of non-shattering by measuring panicle threshability for O. glaberrima.
We measured the degree of non-shattering in 149 O. glaberrima accessions according to the Standard Evaluation System for Rice (SES) [83]. We report our measurement of panicle threshability, which is directly related to seed shattering, on a scale of 1, 3, 5, 7, and 9, which indicates a percent shattering of less than 1%, 1–5%, 6–15%, 26–50%, and 51–100% respectively (see S10 Table for each O. glaberrima individuals’ shattering score).
The geographic distribution of the panicle threshability score showed an east to west gradient, where inland O. glaberrima varieties were more likely to have samples with higher threshability score values (Fig 8A). Specifically, the OG-B and OG-C1 group had a mix of individuals with varying degree of shattering, while the groups closer to the coastal area, namely the OG-A1, OG-A2, and OG-C2 group, had predominantly individuals with panicles that were non-shattering (Fig 8B). We compared the shattering scores for each genetic group by conducting Mann-Whitney U test for all pairwise combinations (S11 Table). Results showed significant difference in shattering scores between the coastal and inland genetic groups (OG-B and OG-C1 vs. OG-A1, OG-A2, and OG-C2).
Noticeably the threshability scores were consistent with the shattering mutation results (Table 1). In the coastal region, most individuals (with the exception of the OG-A1 group and see below for detail) had both the sh1 and sh4 mutations and were non-shattering. On the other hand, OG-B and OG-C1 were the only groups that were fixed for the sh4 casual domestication mutation while the sh1 gene was wildtype, and many individuals had higher proportions of shattering seeds. This indicates mutations in at least two shattering genes (in the case of the OG-A2, OG-B, and OG-C group the genes sh1 and sh4) were required for complete non-shattering in O. glaberrima. In the case of OG-B and OG-C1 group, selection for non-shattering was incomplete either because the group represents an ancestral population or is a result from the cultural preference on the degree of seed shattering.
Samples closer to the coast and belonging to the groups OG-A1, OG-A2, and OG-C2 were predominantly non-shattering rice. Interestingly for the OG-A1 group, the casual domestication mutation was polymorphic in both sh1 and sh4 genes (Table 1) but all varieties in OG-A1 had non-shattering seeds (Fig 8B). There were 27 OG-A1 individuals with the same sh1 and sh4 allelic status (i.e. sh1 wildtype and sh4 mutant) as the OG-B and OG-C1 group (S12 Table). However, unlike the inland group, all individuals had non-shattering seeds suggesting there may be a third shattering gene, and/or different mutations in sh1 involved in the non-shattering phenotype. In addition, all seed non-shattering OG-A1 individuals without the casual sh4 mutation (Fig 6B star) had the sh1 deletion (S12 Table). This suggests the casual mutations for sh1 and sh4 were independently selected, possibly from different genetic backgrounds.
In the end, our panicle threshability results are consistent with the population genetics result of sh1 and sh4. Specifically, the selection process for non-shattering was either incomplete (i.e. OG-B and OG-C1 population) or heterogeneous (i.e. OG-A and OG-C2 population), where two individuals with the same degree of threshability did not share the same casual domestication mutations in their shattering genes (i.e. OG-A1 population). This opens up the possibility that domestication, at least involving seed non-shattering, does not have a single origin in O. glaberrima, but may have occurred in multiple genetic backgrounds and/or geographical regions.
Our analysis of whole genome re-sequencing data in the African rice O. glaberrima and its wild ancestor O. barthii provides key insights into the geographic structure and nature of domestication in crop species. Our analysis suggests that O. glaberrima is comprised of at least 5 distinct genetic groups, which are found in different geographic areas in West and Central Africa. We find that many individuals that have been identified as O. barthii (and which in the past have been thought to be the immediate ancestor of the domesticated crop) form a distinct genetic group that behaves almost identically to O. glaberrima. These include similarities in LD decay, polymorphism levels, and low genetic differentiation with domesticated African rice. Moreover, several of these O. barthii individuals carried causal mutations in the key domestication genes sh4, sh1 and PROG1. Together this suggests these O. barthii individuals, which we collectively refer to as the OB-G group, may represent a feral O. glaberrima or may have been misidentified as the crop species.
Portères hypothesized that western inland Africa near the inner Niger delta of Mali as the center of origin for O. glaberrima [8,9], and this has been the commonly accepted domestication model for O. glaberrima [84]. Under this single center of origin model, O. glaberrima from the OG-C1 genetic group (closest to the inner Niger delta) would have acquired key domestication mutations before spreading throughout West Africa. Here, we suggest that the domestication of O. glaberrima may be more complex. Phylogeographic analysis of three domestication loci indicates that the causal mutations associated with the origin of O. glaberrima may have arisen in three different areas. Phenotype assay of panicle threshability, a core early plant domestication trait [28], showed that the selection for seed non-shattering was incomplete in several inland O. glaberrima samples. Within the coastal O. glaberrima samples, almost all individuals have non-shattering seeds but the casual domestication mutations in two key shattering genes (sh1 and sh4) are not fixed.
Our results support a view, in which domestication has largely been a long protracted process, often involving thousands of years of transitioning a wild plants into a domesticated state [82,85,86]. If this indeed happened for O. glaberrima, our study suggests this protracted period of domestication had no clear single center of domestication in African rice. Instead domestication of African rice was likely a diffuse process involving multiple centers [86–88]. In this model, cultivation may have started at a location and proto-domesticates spread across the region with some (but maybe not all) domestication alleles. Across the multiple regions, the differing environmental conditions and cultural preferences of the people domesticating this proto-glaberrima resulted in differentiation into distinct genetic groups. Temporal and spatial variation in the domestication genes resulted in causal mutations for domestication traits appearing at different parts of the species range. The genetic and geographic structure in this domesticated species suggests that admixture might have allowed local domestication alleles to spread into other proto-domesticated O. glaberrima genetic groups in different parts of West and Central Africa. This would have facilitated the development of modern domesticated crop species, which contain multiple domestication alleles sourced from different areas. In the end, these gradual changes occurring across multiple regions provided different mutations at key domestication genes, which ultimately spread and came together to form modern O. glaberrima.
There has been intense debate on the nature of domestication, and recently (with particular emphasis on early Fertile Crescent domestication) discussion on whether this process proceeds in localized (centric) vs. a diffuse manner across a wider geographic area (non-centric) [82,87,89]. As we begin to use more population genomic data and whole genome sequences, as well as identify causal mutations associated with key domestication traits, we can begin to study the interplay between geography, population structure and the evolutionary history of specific domestication genes and reconstruct the evolutionary processes that led to the origin and domestication of crop species. Moreover, a functional phylogeographic approach, as demonstrated here, could provide geographic insights into key traits that underlie species characteristics, and may allow us to understand how functional traits originate and spread across a landscape.
O. glaberrima and O. barthii samples were ordered from the International Rice Research Institute and their accession numbers can be found in S1 Table. DNA was extracted from a seedling stage leaf using the Qiagen DNeasy Plant Mini Kit. Extracted DNA from each sample was prepared for Illumina genome sequencing using the Illumina Nextera DNA Library Preparation Kit. Sequencing was done on the Illumina HiSeq 2500 –HighOutput Mode v3 with 2×100 bp read configuration, at the New York University Genomics Core Facility. Raw FASTQ reads are available from NCBI biproject ID PRJNA453903.
Raw FASTQ reads from the study Wang et al. [13] and Meyer et al. [14] were downloaded from the sequence read archive (SRA) website with identifiers SRP037996 and SRP071857 respectively.
FASTQ reads were preprocessed using BBTools (https://jgi.doe.gov/data-and-tools/bbtools/) bbduk program version 37.66 for read quality control and adapter trimming. For bbduk we used the option: minlen = 25 qtrim = rl trimq = 10 ktrim = r k = 25 mink = 11 hdist = 1 tpe tbo; which trimmed reads below a phred score of 10 on both sides of the reads to a minimum length of 25 bps, 3' adapter trimming using a kmer size 25 and using a kmer size of 11 for ends of read, allowing one hamming distance mismatch, trim adapters based on overlapping regions of the paired end reads, and trim reads to equal lengths if one of them was adapter trimmed.
FASTQ reads were aligned to the reference O. glaberrima genome downloaded from EnsemblPlants release 36 (ftp://ftp.ensemblgenomes.org/pub/plants/). Read alignment was done using the program bwa-mem version 0.7.16a-r1181 [90]. Only the 12 pseudomolecules were used as the reference genome and the unassembled scaffolds were not used. PCR duplicates during the library preparation step were determined computationally and removed using the program picard version 2.9.0 (http://broadinstitute.github.io/picard/).
Using the BAM files generated from the previous step, genome-wide read depth for each sample was determined using GATK version 3.8–0 (https://software.broadinstitute.org/gatk/).
Because of the differing genome coverage between samples generated from different studies, depending on the population genetic method we used different approaches to analyze the polymorphic sites. A complete probabilistic framework without hard-calling genotypes, was implemented to analyze levels of polymorphism (including estimating θ, Tajima’s D, and FST), population relationships (ancestry proportion estimation and phylogenetic relationship), and admixture testing (ABBA-BABA test). For methods that require genotype calls, we analyzed samples that had greater then 10× genome coverage. Details are shown below.
We used ANGSD version 0.913 [53] and ngsTools [54] which uses genotype likelihoods to analyze the polymorphic sites in a probabilistic framework. ngsTools uses the site frequency spectrum as a prior to calculate allele frequencies per site. To polarize the variants the O. rufipogon genome sequence [36] was used. Using the O. glaberrima genome as the reference, the O. rufipogon genome was aligned using a procedure detailed in Choi et al. [37]. For every O. glaberrima genome sequence position, the aligned O. rufipogon genome sequence was checked, and changed to the O. rufipogon sequence to create an O. rufipogon-ized O. glaberrima genome. Gaps, missing sequence, and repetitive DNA were noted as ‘N’. For all analysis we required the minimum base and mapping quality score per site to be 30. We excluded repetitive regions in the reference genome from being analyzed, as read mapping to these regions can be ambiguous and leading to false genotypes.
The site frequency spectrum was then estimated using ANGSD with the command:
For each genetic group a separate site frequency spectrum was estimated and the options–minInd, -setMinDepth, and–setMaxDepth were changed accordingly. Parameter minInd represent the minimum number of individuals per site to be analyzed, setMinDepth represent minimum total sequencing depth per site to be analyzed, and setMaxDepth represent maximum total sequencing depth per site to be analyzed. We required–minInd to be 80% of the sample size, -setMinDepth to be one-third the average genome-wide coverage, and–setMaxDepth to be 2.5 times the average genome-wide coverage. Using the site frequency spectrum, θ was calculated with the command:
The θ estimates from the previous command was used to compute sliding window values for Tajima’s θ and D [71] with the command:
thetaStat do_stat $Theta–nChr $Indv–win 10000 –step 10000
The option nChr is used for the total number of samples in the group being analyzed. Window size was set as 10,000 bp and was incremented in non-overlapping 10,000 bp.
FST values between pairs of population were also calculated using a probabilistic framework. Initially, we calculated the joint site frequency spectrum (2D-SFS) between the two populations of interest with the command:
Each population’s site frequency spectrum estimated from previous step is used to estimate the 2D-SFS. With the 2D-SFS FST values were calculated with the command:
FST values were calculated in non-overlapping 10,000 bp sliding windows. For the sliding windows calculated for θ, Tajima’s D, and FST values, we required each window to have at least 30% of the sites with data or else the window was discarded from being analyzed.
Ancestry proportions were estimated using NGSadmix [55]. Initially, genotype likelihoods were calculated using ANGSD with the command:
To reduce the impact of LD would have on the ancestry proportion estimation, we randomly picked a polymorphic site in non-overlapping 50 kbp windows. In addition we made sure that the distance between polymorphic sites were at least 25 kbp apart. We then used NGSadmix to estimate the ancestry proportions for K = 2 to 9. For each K the analysis was repeated 100 times and the ancestry proportion with the highest log-likelihood was selected to represent that K.
Phylogenetic relationships between samples were reconstructed using the genetic distance between individuals. Distances were estimated using genotype posterior probabilities from ANGSD command:
Genotype posterior probability was used by NGSdist [58] to estimate genetic distances between individuals, which was then used by FastME ver. 2.1.5 [91] to reconstruct a neighbor-joining tree. Tree was visualized using the website iTOL ver. 3.4.3 (http://itol.embl.de/) [92].
Principal component analysis were also conducted using genotype likelihoods. Genotype posterior probabilities from ANGSD command:
The genotype posterior probability was then used by the program ngsCovar [54] to conduct the principal component analysis.
Since several methods require genotype calls for analysis SNP calling was also performed. Samples with greater than or equal to 10× genome coverage (GE10 dataset) was considered to ensure sufficient read coverage for each site at the cost of excluding individuals from genotype calling. These were 174 individuals that belonged to the genetic grouping designated by this study, and full list of individuals can be found in S13 Table.
For each sample, genotype calls for each site was conducted using the GATK HaplotypeCaller engine under the option `-ERC GVCF`mode to output as the genomic variant call format (gVCF). The gVCFs from each sample were merged together to conduct a multi-sample joint genotyping using the GATK GenotypeGVCFs engine. Genotypes were divided into SNP or INDEL variants and filtered using the GATK bestpractice hard filter pipeline [93]. For SNP variants we excluded regions that overlapped repetitive regions and variants that were within 5 bps of an INDEL variant. We then used vcftools version 0.1.15 [94] to select SNPs that had at least 80% of the sites with a genotype call, and exclude SNPs with minor allele frequency <2% to remove potential false positive SNP calls arising from sequencing errors or false genotype calls.
The GE10 dataset was used for this analysis, as it requires hard-called genotypes. The O. glaberrima samples were grouped according to the grouping scheme designated in this study (Fig 3), and any members that were more similar to other grouping then its own were examined by estimating their silhouette scores [65]. Using the program PLINK version 1.9 [95] for calculating genetic distances from all pairwise comparisons, silhouette scores were calculated using the formula:
s(i)=[b(i)−a(i)]/max{a(i),b(i)}
(1)
where i represents an individual, s the silhouette score, a the average genetic distance to members of own group, and b the average genetic distance to members of foreign group. Individuals with negative silhouette scores were filtered out. After filtering, using the remaining individuals the silhouette score based filtering method was iteratively conducted until all individuals had silhouette scores higher then 0.1.
To obtain the outgroup nucleotide variants we downloaded raw sequencing data for six O. rufipogon species corresponding to the Or-C and Or-D clade, which were shown to contain the least amount of domesticated Asian rice admixture from feralization [96]. These samples have identifiers W0137, W1739, W1807, W0170, W0630, and W2263 with SRA run accession IDs of DRR088674, ERR224552, DRR088680, ERR2245549, DRR001185, and DRR088691. O. rufipogon raw FASTQ reads were aligned to the O. glaberrima reference genome as outlined in our previous steps. GATK HaplotypeCaller engine was used for calling genotypes but the multi-sample joint genotyping step for the six O. rufipogon samples were limited to polymorphic sites that overlapped the SNP positions analyzed in the silhouette score analysis.
Population relationships were examined as admixture graphs using Treemix version 1.13 [67]. SNP calls from the core set population was used to calculate the allele frequencies for each genetic group. One hundred SNPs were analyzed together as a block to account for the effects of LD between SNPs. The O. rufipogon variation was used as the outgroup and a Treemix model assuming 0–3 migration events were fitted. The four-population test [68] was conducted using the fourpop program from the Treemix package.
Genome-wide levels of LD (r2) was estimated with the GE10 dataset and using the program PLINK. LD was calculated for each genetic group separately across a non-overlapping 1Mbp window and between variants that are at most 99,999 SNPs apart. LD data was summarized by calculating the mean LD between a pair of SNPs in 1,000 bp bins. A LOESS curve fitting was applied for a line of best fit and to visualize the LD decay.
We downloaded protein coding sequences for O. sativa, O. glaberrima, and O. barthii from EnsemblPlants release 36. An all-vs-all reciprocal BLAST hit approach was used to determine orthologs between species and paralogs within species. We used the program Orthofinder ver. 1.19 [97] to compare the proteomes between and within species for ortholog assignment. Orthofinder used the program DIAMOND ver. 0.8.37 [98] for sequence comparisons.
Synteny based on the O. sativa sh1 gene (Ossh1; O. sativa cv. japonica chromosome 3:25197057–25206948) indicated orthologs surrounding Ossh1 was found in chromosome 3 of O. barthii and on an unassembled scaffold named Oglab03_unplaced035 in O. glaberrima (S14 Table). The sh1 gene was missing in O. glaberrima suggesting the gene deletion may have led to complex rearrangements that prevented correct assembly of the region in the final genome assembly. Because of this we used the O. barthii genome sequence to align raw reads and call polymorphic sites for downstream analysis.
The approximate region of the deletion in the O. barthii genome coordinate was examined by looking at the polymorphic sites, since our quality control filter removed polymorphic sites if it had less than 80% of the individuals with a genotype call. Between the genomic positions at O. barthii chromosome 3 position 23,100,000–23,130,000, no polymorphic sites passed the quality control filter (S13 Fig) and contained the gene Obsh1. Between the region at O. barthii chromosome 7 position 2,655,000–2,675,000 there was also no polymorphisms passing the filter and contained the gene ObPROG1 (S14 Fig).
Gene deletion was inferred from comparing the read depth of a genic region inside and outside a candidate deletion region. Read depth was measured using bedtools ver. 2.25.0 [99] genomecov program. Individuals with and without the deletion were determined by comparing the median read coverage of the domestication gene within the candidate deletion region, to a gene that is outside the deletion region. We checked the orthologs to make sure the gene outside the deletion region existed in O. barthii, O. glaberrima, and O. sativa. To determine the sh1 deletion status we examined its read depth and compared it to the O. barthii gene OBART03G27620 that was upstream and outside the candidate deletion region. Ortholog of OBART03G27620 is found in both O. sativa (Os03g0648500) and O. glaberrima (ORGLA03G0257300). To determine the deletion status of PROG1 gene we examined its read depth and compared to O. barthii gene OBART07G03440. Ortholog of OBART07G03440 is found in both O. sativa (Os07g0153400) and O. glaberrima (ORGLA07G0029300).
Because some individuals had low genome-wide coverage (S1 Table) there is the possibility that some of those individuals had been detected as false positive deletion events. There are two main reasons we believe the deletions are likely to be present even for low coverage individuals. For example for the sh1 deletion, (i) all individuals had at least a median coverage of ~1× in the OBART03G27620 gene (S15 Table) suggesting read coverage may be low but if the gene is not deleted it is evenly distributed across a gene, and (ii) even comparing individuals with and without the sh1 deletion that had a ~1× median coverage in the non-deleted OBART03G27620 gene, there were clear differences in the sh1 gene coverage (S15 Fig) where the individuals with the deletion always had a median coverage of zero.
Gene names for the non-shattering phenotype have unfortunately varied between different Oryza studies. Genetic studies comparing Asian rice O. sativa cv. Japonica and its wild progenitor O. rufipogon had identified a single dominant allele responsible for non-shattering and named the locus as Sh3 [100,101]. The causal gene was later identified on chromosome 4 and was given a new name as sh4 [17]. Studies have used the names Sh3 and sh4 synonymously as the common gene name for the gene with locus ID Os04g0670900 [78].
Lv et al. [81] had found an O. glaberrima specific gene deletion in chromosome 3 that caused a non-shattering phenotype and named this gene as SH3. SH3 belongs to a YABBY protein family transcription factor. Using the SH3 coding sequence in O. barthii (ObSH3), which the gene is not deleted, orthologs were found in maize (B4FY22), barley (M0YM09), and Brachypodium (I1GPY5) [81]. We discovered this group of proteins belonged to a group identified in Plant Transcription Factor Database ver 4.0 [102] under the ID OGMP1394. The O. sativa gene member of this group was gene ID Os03g0650000, which has previously been identified as a gene involved in non-shattering [103]. Thus, ObSH3 and Os03g0650000 are orthologs of each other and Os03g0650000 has been named as sh1. Here, we followed the guideline recommended by Committee on Gene Symbolization Nomenclature and Linkage (CGSNL) [104] to designate SH3 from Lv et al. [81] as sh1 to avoid using the overlapping gene name sh3.
To investigate the haplotype structure around the domestication genes we used all individuals from O. glaberrima, OB-G, and OB-W population regardless of the genome coverage. The O. glaberrima and O. barthii genome were used as reference to align the raw reads and call polymorphisms as outlined above. Missing genotypes were then imputed and phased using Beagle version 4.1 [105].
We used vcftools to extract polymorphic sites around a region of interest. The region was checked for evidence of recombination using a four-gamete test [80], to limit the edges connecting haplotypes as mutation distances during the haplotype network reconstruction. To minimize false positive four-gamete test results caused from technical errors such as genotype error and sequencing error, if the observed frequency of the fourth haplotype was below 1% we considered the haplotype an error and did not consider it as evidence of recombination. If a region had evidence of recombination we checked if the recombination was limited to the wild or domesticated African rice. If recombination was only detected in the wild population then we determined the pair of SNPs that failed the four-gamete test. Here, because the four-gamete test did not detect any evidence of recombination in the O. glaberrima population, the fourth haplotype observed in the wild population is only limited to O. barthii and do not provide any information with regard to the direct origin of the O. glaberrima haplotypes. Hence, we removed individuals with the fourth haplotype and estimated the haplotype network of the region.
Haplotype network was reconstructed using the R pegas [106] and VcfR [107] package, using the hamming distance between haplotypes to construct a minimum spanning tree. For each domestication gene and its surrounding region, a phylogenetic tree was reconstructed by sampling a single haplotype for each individual. Bootstrap replicated phylogenetic trees were built using RAxML [108] and plotted with iTOL.
O. glaberrima landraces were grown during the 2018 dry season at the International Rice Research Institute (IRRI) block L4 (14°09'34.6"N 121°15'42.4"E) experimental field. At maturity, when at least 85% of the grains on a panicle are matured [109,110], panicles were harvested and evaluated for threshability using a established method by IRRI [111]. In brief, a total of 6 plants for each landrace from three plot replicates were sampled. During panicle threshability measurement, each panicle was grasped to apply slight pressure. Grains detached from the panicle and panicles intact with grains were collected. The numbers of grains that detached and remained attached were counted separately to obtain the percentage of shattered grains [83]. Percent shattering were converted to panicle threshability scores according to the Standard Evaluation System for Rice [83].
|
10.1371/journal.pntd.0003854 | Autochthonous Chikungunya Transmission and Extreme Climate Events in Southern France | Extreme precipitation events are increasing as a result of ongoing global warming, but controversy surrounds the relationship between flooding and mosquito-borne diseases. A common view among the scientific community and public health officers is that heavy rainfalls have a flushing effect on breeding sites, which negatively affects vector populations, thereby diminishing disease transmission. During 2014 in Montpellier, France, there were at least 11 autochthonous cases of chikungunya caused by the invasive tiger mosquito Aedes albopictus in the vicinity of an imported case. We show that an extreme rainfall event increased and extended the abundance of the disease vector Ae. albopictus, hence the period of autochthonous transmission of chikungunya.
We report results from close monitoring of the adult and egg population of the chikungunya vector Ae. albopictus through weekly sampling over the entire mosquito breeding season, which revealed an unexpected pattern. Statistical analysis of the seasonal dynamics of female abundance in relation to climatic factors showed that these relationships changed after the heavy rainfall event. Before the inundations, accumulated temperatures are the most important variable predicting Ae. albopictus seasonal dynamics. However, after the inundations, accumulated rainfall over the 4 weeks prior to capture predicts the seasonal dynamics of this species and extension of the transmission period.
Our empirical data suggests that heavy rainfall events did increase the risk of arbovirus transmission in Southern France in 2014 by favouring a rapid rise in abundance of vector mosquitoes. Further studies should now confirm these results in different ecological contexts, so that the impact of global change and extreme climatic events on mosquito population dynamics and the risk of disease transmission can be adequately understood.
| During last years, we have seen an astonishing expansion of Chikungunya virus and an increase in dengue cases worldwide, together with the worldwide expansion of the Asian tiger mosquito Aedes albopictus. In addition, extreme rainfall events are envisaged to become increasingly likely as a result of ongoing climate change, but controversy surrounds the relationship between extreme rainfall events and mosquito-borne diseases. The common view in most works on climate and mosquito-borne diseases is that heavy rainfalls produce a flushing effect of immature mosquitoes in breeding containers, diminishing the mosquito abundance and in turn diminishing disease transmission. We analysed the relationships between the autochthonous chikungunya transmission in Montpellier (Southern France) in 2014, an extreme rainfall event that flooded the city, and a close monitoring of the vector Ae. albopictus, revealing an unexpected pattern. This extreme rainfall event did not, in fact, decrease but instead had increased the global risk of chikungunya transmission by sustaining high abundance of the disease vector Ae. albopictus, hence extending the transmission period. We propose that an effort on source reduction campaigns must be implemented after heavy rainfall events. These results are relevant to those involved in the surveillance and control of chikungunya and dengue transmission in temperate as well as tropical areas.
| Extreme precipitation events are envisaged to increase as a result of ongoing global warming [1]. Heavy rains have consequences for infectious diseases, with, in particular, some vector-borne disease outbreaks being associated with flooding [2,3]. However, the relationships between flooding and vector-borne diseases are mired in controversy and need to be clarified [4].
A common belief among the scientific community and public health officers is that “heavy rainfalls produce a flushing effect of immature mosquitoes (larvae and pupae) in breeding containers, diminishing the mosquito abundance” [3,5], in turn diminishing disease transmission. The common view in most works on climate and mosquito-borne diseases is that “Intense rainfall may wash out breeding sites and thus have a negative effect on vector populations” [6].
During September-November 2014, French health authorities reported a cluster of 11 autochthonous cases of chikungunya in the city of Montpellier in the vicinity of a recently imported case [7, 8]. This was the first report of locally transmitted chikungunya in France since 2010, adding to the 4 cases of dengue reported earlier that year in the neighbouring PACA region, and dengue cases in 2010 and 2013 [7–9]. Abundances of the tiger mosquito Aedes albopictus, the competent disease vector, in an increasing number of places and the large number of imported cases of chikungunya (443 cases), dengue (163 cases) and co-infections (6 cases) in France [7], did indeed concur to increase the risk of autochthonous transmission in Southern France.
The objective of this study was to analyse the influence of climatic variables, including an extreme rainfall event, on Ae. albopictus abundance in Montpellier in 2014, and its impact on the risk of autochthonous chikungunya transmission.
Mosquitoes were sampled in the municipalities of Montpellier and Castelnau-le-Lez in the Province of Hérault, region Languedoc-Roussillon, southern France, with a Mediterranean climate and a resident human population of 400.470 inhabitants spread over 14.62 Km2.
Adult mosquitoes were collected using 24 BG-Sentinel traps (BioGents, Regensburg, Germany), with the attractant BG lure, a synthetic lure developed to mimic human odors, consisting of lactic acid, ammonia, and caproic acid on a long-lasting lure. Traps were located in 8 sampling sites separated by a mean geographic distance of 2.2 Km. Each sampling site consisted of a private house with garden, and three trap replicates set 5–10 m apart were operated in each site. Each trap was connected to a battery (12 V), and the captures were conducted for 24 H, once a week, from the 21th week (2nd week of May) to the 50th week (2nd week of December) of 2014. Sampled mosquitoes were transported live to the laboratory in an insulated thermal bag filled with ice and frozen at -20°C. The species and sex of each individual were identified in the laboratory with a stereomicroscope using the appropriate taxonomic keys [10]. Although several species were captured, including Ae. albopictus, Culex pipiens, Culiseta longiareolata, Aedes caspius and Anopheles maculipennis s.l., only Ae. albopictus females were included in the analysis. The mean weekly abundance of Ae. albopictus females captured in the 24 traps was used as the dependent variable.
In parallell, eggs of Ae. albopictus were collected using oviposition traps (ovitraps) placed in each of the 24 locations where the adult population was monitored. Bacilus thuringiensis var. israeliensis granules (VectoBac, ValentBioSciences) were added to the water to prevent larval development and traps were checked once a week. Eggs were counted under a stereomicroscope and averaged across traps and over collection sites.
Daily mean, minimum and maximum air temperatures, daily total precipitation and insolation time for the study period were obtained from the Montpellier Fréjorgues meteorological station (http://www.meteociel.fr/climatologie/villes.php?code=7643&mois=6&annee=2014), located 6 Km from the city centre. For the purposes of our study, these data were pooled by week. Then, for temperature, weekly means and weekly range were computed, and for rainfall, total weekly precipitation was calculated. Environmental variables may have a strong influence on mosquito populations not only at the time of adult capture but also, and mainly, at egg-laying and larval development which occur 1–4 weeks earlier. Therefore, accumulated temperature and rainfall were calculated over the 4 weeks preceding the week of sampling. Growing Degree Days (GDD) were calculated with a baseline temperature of 11°C to compute weekly accumulated Growing Degree Days (GDD). Based on our previous work, we also assessed a ‘bounded’ estimate of the accumulated GDD (Bounded accumulated GDD), with a maximum threshold of 1350 accumulated GDD above which any further increase in GDD is considered detrimental and counted negatively [11, 12].
We developed a Generalized Linear Model with negative binomial distribution (as the data were over-dispersed) to investigate whether Ae. albopictus adult female abundance was influenced by temperature and rainfall before and after the extreme rainfall event. The response variable was the total weekly Ae. albopictus adult female abundance, and the selected explanatory variables were the average weekly temperatures (minimum, mean and maximum), the total weekly precipitation, all the accumulated temperature and rainfall variables (see Climatic data section for details), Weekly Growing Degree Days, Accumulated Growing Degree Days and Bounded Accumulated Growing Degree Days. Due to the highly expected co-linearity between the explanatory variables, different univariate models were built and the best model was selected on the basis of AIC (Akaike Information Criterion), ΔAIC and Akaike weights. We calculated the explained deviance as: (Null deviance-Residual deviance)/ Null deviance. Statistical analysis was performed in R version 2.14.2 [13] with the packages MASS [14] and MuMin[15].
We developed a close monitoring of the adult population of the chikungunya vector Ae. albopictus in Montpellier through weekly sampling over the entire mosquito breeding season (i.e., May to November 2014). Although mosquito densities steadily declined after peaking in late August, extreme rainfall events flooding the area at the end of September and beginning of October (Week 39), with up to 252 mm of rain falling in just 3 hours (recorded on 29th September in Montpellier), resulted in an explosive mosquito population growth extending into October and surpassing the abundance peak recorded earlier in August (Fig 1). Statistical analysis (Table 1; Fig 2) revealed that, before the inundations, temperature (Accumulated Growing Degree Days) was the most important variable predicting the seasonal dynamics of Ae. albopictus, with 69.3% of the variance explained. However, after the inundations, accumulated rainfall over the 4 weeks before capture predicted Ae. albopictus seasonal dynamics, explaining 92.3% of the variance (Table 2; Fig 2). The seasonal pattern of eggs abundance in the ovitraps closely matched that of female abundance, with a lag of several days (Fig 1).
We have shown that the extreme precipitation event clearly contributed to increasing and extending the abundance of the disease vector Ae. albopictus, and hence to extending the period of autochthonous transmission of chikungunya in Montpellier in 2014. Female density increased rapidly after the extreme event, soon followed by a rise in the number of eggs collected in ovitraps. Therefore, and contrary to common belief [3,4,5,6], our empirical data suggest that heavy rainfall events do not, in fact, decrease but instead may increase the global risk of chikungunya (and other arboviruses) transmission, by extending the transmission season. This is the first evidence to support the relationship between heavy rainfall and chikungunya transmission. Our observations before and after the extreme precipitation event, suggest that heavy rains after a dry period with low precipitations have filled all the peridomestic containers where desiccated eggs of Ae. albopictus were to be found, and that it was this situation which gave rise to the increase in mosquito numbers several weeks later. Indeed, the majority of breeding sites colonized by Ae. albopictus larvae in Mediterranean Europe are small peridomestic containers of less than 10 L. such as scuppers, flowerpot saucers, drums, buckets, solid waste and others, whereas the productivity of catch basins is generally much lower [16, 17, 18]. In Southern France, more than 80% of breeding sites are situated in the private domain and only 12% of the immature stages were detected in catch basins [16]. The outcome of such extreme climatic events might therefore be different in different ecological contexts, depending on the larval ecology of the species. For example, West Nile virus (WNV) outbreaks have been associated with flooding, in Romania, the Czech Republic and Russia [19, 20, 21]. However, these outbreaks were related to flooded building basements, resulting in an increase in populations of the WNV vector Culex pipiens, which thrives in large water collections that are not suitable for Ae. albopictus development. Further studies are needed to explore the general relevance of our findings and their implications for diseases transmission by Aedes albopictus in Europe, as well as elsewhere where the species has established.
Because our data showed that heavy rains in Mediterranean cities impacted on Ae. albopictus abundances, we propose that an effort on source reduction campaigns must be implemented after such heavy rainfall event. In fact, these high-volume rainfall events, referred to as Cévenol episodes, are relatively frequent in the south of France.
We further encourage tropical countries that are endemic for Chikungunya and Dengue to develop similar studies in order to explore the relevance of enforcing source reduction approaches following extreme climatic events and we believe that our results should be brought to the attention of those involved in the surveillance and control of vectors and vector-borne diseases in the context of global change in temperate as well as tropical areas of the world.
|
10.1371/journal.pcbi.1003055 | Connection between Oligomeric State and Gating Characteristics of Mechanosensitive Ion Channels | The mechanosensitive channel of large conductance (MscL) is capable of transducing mechanical stimuli such as membrane tension into an electrochemical response. MscL provides a widely-studied model system for mechanotransduction and, more generally, for how bilayer mechanical properties regulate protein conformational changes. Much effort has been expended on the detailed experimental characterization of the molecular structure and biological function of MscL. However, despite its central significance, even basic issues such as the physiologically relevant oligomeric states and molecular structures of MscL remain a matter of debate. In particular, tetrameric, pentameric, and hexameric oligomeric states of MscL have been proposed, together with a range of detailed molecular structures of MscL in the closed and open channel states. Previous theoretical work has shown that the basic phenomenology of MscL gating can be understood using an elastic model describing the energetic cost of the thickness deformations induced by MscL in the surrounding lipid bilayer. Here, we generalize this elastic model to account for the proposed oligomeric states and hydrophobic shapes of MscL. We find that the oligomeric state and hydrophobic shape of MscL are reflected in the energetic cost of lipid bilayer deformations. We make quantitative predictions pertaining to the gating characteristics associated with various structural models of MscL and, in particular, show that different oligomeric states and hydrophobic shapes of MscL yield distinct membrane contributions to the gating energy and gating tension. Thus, the functional properties of MscL provide a signature of the oligomeric state and hydrophobic shape of MscL. Our results suggest that, in addition to the hydrophobic mismatch between membrane proteins and the surrounding lipid bilayer, the symmetry and shape of the hydrophobic surfaces of membrane proteins play an important role in the regulation of protein function by bilayer membranes.
| A fundamental property of living cells is their ability to detect mechanical stimuli. Microbes, in particular, often transition between different chemical environments, leading to osmotic shock and concurrent changes in membrane tension. The tension of microbial cell membranes is detected and controlled by membrane molecules such as the widely-studied mechanosensitive channels which, depending on the tension exerted by the surrounding lipid bilayer, switch between closed and open states. Thus, the biological function of mechanosensitive channels relies on an interplay between bilayer mechanical properties and protein structure. Using a physical model of cell membranes it was shown previously that the basic phenomenology of mechanosensitive gating can be understood in terms of the bilayer deformations induced by mechanosensitive channels. We have generalized this physical model to allow for the molecular structures of mechanosensitive channels reported in recent experiments. Our methodology allows the calculation of protein-induced membrane deformations for arbitrary oligomeric states of membrane proteins. We predict that distinct oligomeric states and hydrophobic shapes of mechanosensitive channels lead to distinct functional responses to membrane tension. Our results suggest that the shape of membrane proteins, and resulting structure of membrane deformations, plays a crucial role in the regulation of protein function by bilayer membranes.
| The biological function of membrane proteins is determined by a complex interplay between protein structure and the properties of the surrounding lipid bilayer [1]–[6]. In particular, the bilayer hydrophobic core couples to the hydrophobic regions of membrane proteins [7]–[10]. The resulting deformations in the lipid bilayer membrane from its unperturbed state can be described quantitatively [11]–[17] using the continuum elasticity theory of membranes [18]–[20]. The energetic cost of protein-induced membrane deformations depends on the protein conformational state as well as on the bilayer material properties, which allows [11]–[17] the lipid bilayer to act as a regulator of protein function. A widely-studied model system for the coupling between membrane protein function and the elastic deformation of lipid bilayers is provided by mechanosensitive ion channels. Mechanosensitive channels are capable of transducing membrane tension into an electrochemical response [21]–[23] by switching from a closed to an open conformational state with increasing membrane tension, allowing cells to sense touch, sound, and pressure.
A paradigm of mechanosensation is the prokaryotic mechanosensitive channel of large conductance (MscL) [24]–[26]. In particular, biophysical approaches such as patch-clamp experiments and reconstitution of MscL in artificial lipid bilayer vesicles have allowed [24]–[34] a systematic analysis of the relation between lipid material properties and the gating probability of MscL with increasing membrane tension. However, despite its central significance, even basic issues such as the physiologically relevant oligomeric states and molecular structures of MscL remain a matter of debate [26], [35]–[37]. So far, the oligomeric state and molecular structure of MscL have mainly been studied [24]–[27], [33], [35]–[47] using crystallographic, biochemical, and computational approaches. This has led to the identification of a number of possible oligomeric states and molecular structures of MscL. In particular, early low-resolution electron microscopy studies suggested that MscL is a hexamer [39], while more recent high-resolution x-ray crystallographic studies demonstrated pentameric [40] and tetrameric [46] MscL structures. Do the various reported stoichiometries of MscL induce distinct membrane deformations, yielding distinct functional responses to membrane tension? More generally, theoretical studies of the energetic cost of protein-induced membrane deformations [11]–[15] have mostly focused on membrane inclusions with a cylindrical or conical hydrophobic shape. But experimental surveys of the protein content in the membranes of, for instance, synaptic vesicles [48] and Acinetobacter baumannii [49] suggest [50] that membrane proteins exhibit great diversity in their oligomeric state and transmembrane shape. What is the relationship between the oligomeric state and hydrophobic shape of a membrane protein and the elastic energy required to accommodate the membrane protein within the lipid bilayer?
In this article we address the above questions on the basis of the continuum elasticity theory of lipid bilayer membranes [18]–[20]. In particular, we generalize the standard framework for calculating the energetic cost of protein-induced membrane deformations [11]–[15], which was employed previously to understand the basic phenomenology of MscL gating [51]–[54], to account for non-circular cross sections of membrane proteins. Our methodology establishes a quantitative relationship between the oligomeric state and hydrophobic shape of a membrane protein and the elastic energy required to accommodate the membrane protein within the lipid bilayer membrane. We make quantitative predictions pertaining to the gating characteristics associated with various structural models of MscL and, in particular, show that different oligomeric states and hydrophobic shapes of MscL yield distinct membrane contributions to the gating energy and gating tension. Generally we find that the oligomeric state and hydrophobic shape of a membrane protein are reflected in the energetic cost of the lipid bilayer deformations necessary to accommodate the protein within the membrane. Our results suggest that, in addition to the hydrophobic mismatch between membrane proteins and the surrounding lipid bilayer [11]–[15], the symmetry and shape of the hydrophobic surfaces of membrane proteins play an important role in the regulation of protein function by bilayer membranes. The results and predictions of our model calculations are described in the Results and Discussion sections. The Models and Methods section provides a detailed mathematical formulation of our analytic methodology linking the hydrophobic shape of membrane proteins to the elastic deformations in the surrounding lipid bilayer membrane.
The basic experimental phenomenology of mechanosensitive gating is captured by a two-state Boltzmann model [27]–[33] describing the competition between the closed and open states of MscL. The central quantity in this model is the channel opening probability(1)where , in which is Boltzmann's constant and is the temperature, is the total free energy difference between the open and closed states of MscL, is the membrane tension, and is the area difference between the open and closed channel states. Equation (1) implies that, for a fixed , a given channel is more likely to be in the open state for larger values of the membrane tension and provides a simple description of experimental data on MscL gating [14], [25], [27]–[33], although a more detailed description of MscL gating would need to take into account the existence of multiple conductance states [31], [33], [43], [44].
A deeper understanding of Eq. (1) in terms of the physical mechanisms underlying MscL gating hinges on a quantitative description of the various contributions to . To this end it is useful [51]–[54] to write as the sum of protein and lipid bilayer contributions,(2)where denotes the difference in internal protein free energy between the open and closed channel states, and denotes the difference in membrane deformation energy between the open and closed states. In general, depends on the oligomeric state and hydrophobic shape of MscL in the closed and open channel states, as well as on bilayer material properties such as the bilayer hydrophobic thickness and bending rigidity. In the remainder of this article we focus on the membrane deformations induced by MscL. To simplify our notation we therefore drop the subscript in and denote by the membrane deformation energy associated with MscL.
The continuum elasticity theory of membranes [18]–[20] provides a general framework for evaluating bilayer-protein interactions [11]–[15], [55]–[63] and, hence, the membrane contribution in Eq. (2). On this basis, the elastic membrane deformations required to accommodate MscL within the bilayer membrane were estimated previously [51]–[54] under the assumption that the transmembrane region of MscL is cylindrical in the closed and open channel states. In particular, it was found that thickness deformations , where and are spatial coordinates along the bilayer membrane, are the dominant elastic membrane deformations induced by MscL. The quantitative details of this previous model of MscL gating, which forms the foundation for the work presented here, are summarized in the Models and Methods section. The overall conclusion of the cylinder model of MscL [51]–[54] is that can be of the same order of magnitude as the measured values of [27], [29], [30], [33] in Eq. (1), with both and being (much) larger than the thermal energy. This suggests that membrane mechanics plays a central role in mechanotransduction and the biological function of MscL. This conclusion is also consistent with experiments measuring the dependence of MscL gating on membrane composition [27], [34]. We emphasize, however, that in general the protein contribution to the free energy difference in Eq. (2) must be considered, and may very well dominate over the membrane contribution. The calculation of the membrane contribution to the gating energy merely represents one step in drawing up a general energy budget of gating.
As mentioned above, the determination of the oligomeric state and, more generally, molecular structure of MscL in different conformational states is a problem of intense experimental interest [24]–[27], [33], [35]–[47]. How do the observed discrepancies in the oligomeric state and molecular structure of MscL relate to the mechanosensitive gating characteristics relevant for the biological function of MscL? In order to address this question from the perspective of membrane mechanics we formally divide into two contributions,(3)where corresponds to the membrane deformation energy associated with the idealized cylinder model of MscL [14], [51]–[54], which we employ as our point of reference when estimating the membrane deformations induced by different oligomeric states of MscL, and corresponds to the modification of due to deviations of the hydrophobic cross section of MscL from the circle. In particular, depends on the oligomeric state (symmetry) of MscL. We have obtained the analytic solution of the general elastic equations describing bilayer deformations induced by MscL in the limit of weak perturbations about the cylindrical reference shape, thus providing a general framework for estimating for arbitrary oligomeric states. The mathematical details of these calculations are described in the Models and Methods section. As discussed below, we find that the oligomeric state and hydrophobic shape of MscL can have a considerable effect on the membrane deformation energy. Thus, based on the membrane deformation energy, distinct boundary shapes and, in particular, distinct oligomeric states of MscL are predicted to yield distinct mechanosensitive gating curves.
A variety of different approaches have been employed [24]–[27], [33], [35]–[47] to study the molecular structure of MscL in different conformational states. Figure 1 shows examples of the molecular structures of MscL obtained for Staphylococcus aureus (SaMscL) and Myobacterium tubercolosis (MtMscL). In particular, Fig. 1(A) displays the tetrameric structure of SaMscL solved most recently [46] using x-ray crystallography. This structure may correspond to an expanded state which is intermediate between the closed and open states of MscL. Figure 1(B) shows pentameric structures of the closed and open states of MscL proposed for MtMscL using crystallographic, biochemical, and computational approaches. The closed state of MscL displayed in the left-hand panel of Fig. 1(B) was obtained on the basis of x-ray crystallography [40], while the right-hand panel displays a molecular model suggested for the open state of MscL [43]–[45]. For MscL in Escherichia coli (EcoMscL), hexameric [38], [39] as well as pentameric [43]–[45] molecular models have been proposed.
The contour lines approximating the cross sections of the transmembrane domains in Fig. 1 represent the bilayer-MscL boundary curves used in our membrane-mechanical model of MscL gating. Similar fits are obtained for the hexameric [38], [39] and pentameric [43]–[45] models proposed for EcoMscL (in particular, see Fig. 3 in Ref. [39] and Fig. 5 in Ref. [44]). The subscript in denotes the oligomeric state (symmetry) of MscL with tetrameric, pentameric, and hexameric structures of MscL corresponding to , , and , respectively. As discussed further in the Models and Methods section, we express the bilayer-MscL boundary curves in terms of the variables and , which are the radial coordinate and the polar angle associated with a polar coordinate system having the MscL protein at its center. The cylinder model of MscL [51]–[54] corresponds to choosing and in the closed and open states of MscL, where and are the cylinder radii in the closed and open channel states. However, as apparent from Fig. 1, the proposed hydrophobic cross sections of MscL [24]–[27], [33], [35]–[47] often deviate from a circle. Indeed, inspired by the structural models of MscL in Fig. 1 and Refs. [38], [39], [43]–[45], we distinguish between two basic shapes of boundary curves. The “polygonal boundary curves” correspond to the tetragonal boundary curve shown in Fig. 1(A) (see Fig. 5 in Ref. [44] for examples of pentagonal boundary curves), while the “clover-leaf boundary curves” correspond to the pentameric propeller shapes in Fig. 1(B) (see Fig. 3 in Ref. [39] for examples of hexameric clover-leaf shapes).
Following the approach summarized in Eq. (3), we employ the cylinder model of MscL [51]–[54] as a means to isolate the role played by the oligomeric state and hydrophobic shape of MscL [24]–[27], [33], [35]–[47] in the regulation of MscL by the surrounding lipid bilayer. In particular, in the simplest model of MscL the hydrophobic thickness of MscL is assumed to be constant when transitioning between closed and open channel states [51], [52], while a more general model [53], [54] allows for changes in the hydrophobic thickness of MscL [43], [44], [47]. We consider here both models of the hydrophobic thickness of MscL but, to systematically study the role played by MscL shape in MscL gating, focus on the case of a constant hydrophobic thickness (see the Models and Methods section for details). In either case we always use the same hydrophobic thickness when making comparisons between different shapes of MscL so as to isolate the role played by MscL shape. Moreover, in order to compare membrane inclusions of equal size, and in light of the central role played by the protein area in Eq. (1), we generally contrast different oligomeric states and hydrophobic shapes of MscL for a fixed area of the hydrophobic cross section. This assumption allows us to make direct comparisons with previous work on bilayer-MscL interactions [51]–[54], and eliminates any spurious effects resulting from MscL occupying different membrane areas in different oligomeric states, but would need to be relaxed for a more detailed description of the membrane deformations induced by MscL. In particular, we use for the closed and open states of MscL the cross-sectional areas and with nm and nm, which were estimated previously [51], [52], [54] for the cylinder model of MscL on the basis of the available structural models of MscL [27], [33], [40]–[45], [47]. Setting the cross-sectional area equal to or fixes the size of the polygonal and clover-leaf shapes, with all other parameters in determined by the respective symmetries and morphologies of the MscL boundary curves. For comparison, we also consider polygonal shapes having the same circumference, rather than the same area, as the cylindrical reference shape in the closed and open channel states.
Figure 2 shows the difference in the membrane deformation fields induced by some of the structural models of MscL in Fig. 1 and Refs. [39], [44] and the cylinder model of MscL [51]–[54]. As described in greater detail in the Models and Methods section, we estimated the membrane deformation field due to a given oligomeric state and molecular structure of MscL by minimizing the elastic membrane energy with respect to the thickness deformation field in the limit of weak deviations from the cylindrical reference shape. In particular, Figs. 2(A), 2(B), and 2(C) show the difference in the thickness deformation fields induced by the tetragonal, pentagonal, and pentameric clover-leaf models of MscL in Fig. 1 and Ref. [44] and the cylinder model of MscL. The cross sections of all membrane inclusions in Fig. 2 are of the area corresponding to the closed state of the cylinder model of MscL. The deformation profiles in Fig. 2 demonstrate that the symmetry and shape of the hydrophobic surface of a membrane protein are reflected in the structure of the membrane deformations required to accommodate the protein within the lipid bilayer.
Figure 2 allows us to gain some intuition regarding the membrane deformations associated with different oligomeric states and hydrophobic shapes of MscL. First consider the deformation fields in Figs. 2(A) and 2(B) due to polygonal boundary curves. Tetragonal and pentagonal boundary curves yield membrane deformations exhibiting four- and five-fold symmetry, respectively. However, while polygonal boundary curves of four-fold and lower-order symmetry produce considerable deviations from the deformation field of the cylindrical reference shape, the shallow angles of pentagonal boundary curves only produce relatively small deviations. Indeed, for hexagonal and higher-order symmetries the deviations from the cylindrical deformation field are even smaller than those shown in Fig. 2(B). For clover-leaf shapes, however, the overall deviation from the deformation field induced by the cylinder model of MscL increases with increasing symmetry of the oligomeric state. As illustrated in Fig. 2(C), clover-leaf shapes of pentameric and higher-order symmetry can, in addition to clover-leaf shapes of lower-order symmetry, yield substantial modifications of the deformation field associated with cylindrical membrane inclusions. Thus, for the polygonal structures of MscL in Fig. 1 and Refs. [39], [44] the overall deviation from the elastic deformation footprint of the cylinder model of MscL decreases with increasing symmetry, but for clover-leaf shapes the overall deviation becomes more pronounced with increasing symmetry.
Figure 3(A) shows the difference in membrane deformation energy between some of the structural models of MscL in Fig. 1 and Refs. [39], [44] and the cylinder model of MscL [51]–[54] as a function of lipid tail length (bilayer hydrophobic thickness). Irrespective of the oligomeric state or hydrophobic shape of MscL, deviations of the cross section of MscL from the circle, and the corresponding non-trivial structure of the membrane deformation field, are seen to increase the elastic energy required to embed MscL within the bilayer membrane. Consistent with the deformation profiles in Fig. 2, the elastic energy difference between polygonal shapes of MscL and the cylinder model of MscL is largest for the tetragonal structure in Fig. 1(A) and decreases with increasing symmetry of the oligomeric state, with hexagonal and higher-order boundary curves inducing elastic membrane deformations of essentially the same energetic cost as the cylinder model of MscL. These conclusions do not change if we consider polygonal models of MscL which have the same circumference, rather than the same cross-sectional area, as the cylindrical reference shape. The pentameric clover-leaf shape of MscL in the closed state [see Fig. 1(B)] induces membrane deformations which carry a greater energetic cost than any of the polygonal shapes considered in Fig. 3(A). In contrast, due to its decreased deviation from the cylindrical reference shape, the hexameric clover-leaf shape of MscL in Ref. [39] carries a relatively small cost in membrane deformation energy. Overall, Fig. 3(A) shows that the various structural models of MscL proposed in previous studies [24]–[27], [33], [35]–[47], and the polygonal or clover-leaf boundary shapes associated with these structural models, yield considerable differences in the membrane deformation energy required to embed MscL within a lipid bilayer membrane.
In Fig. 3(B) we compare the elastic energy difference between the open and closed states of MscL for the structural models of MscL gating in Fig. 1 and Refs. [39], [44] (tetragonal shapes in light blue, pentagonal shapes in orange, pentameric clover-leaf shapes in purple, and hexameric clover-leaf shapes in red) and the cylinder model of MscL [51]–[54] (black). For completeness, we also consider in Fig. 3(B) transitions between a closed pentagonal shape and an open pentameric clover-leaf shape of MscL (dark blue), as well as the reverse case of transitions between a closed pentameric clover-leaf shape and an open pentagonal shape of MscL (green). For all of these plots we used the parameter values characterizing bilayer-MscL interactions estimated in Refs. [51], [52] with zero membrane tension. As discussed in greater detail in the Models and Methods section, this parameterization of bilayer-MscL interactions allows the systematic study of the effect of the structure of membrane deformations on the gating characteristics of MscL, without the further complications introduced by MscL having different hydrophobic thicknesses in the closed and open channel states. We also include in this plot the total free energy differences between the open and closed states of EcoMscL estimated by Perozo et al. [27] for PC16, PC18, and PC20 bilayers at zero membrane tension. In the case of transitions between the polygonal structures in Fig. 1 and Ref. [44], we again find that the deviation from the cylindrical reference shape is more pronounced for tetragonal shapes than for pentagonal shapes, and that in either case the free energy of gating is increased relative to cylindrical membrane inclusions.
In addition, Fig. 3(B) shows that, for transitions between the pentameric clover-leaf shapes in Fig. 1, the difference in membrane deformation energy between the open and closed states of MscL is strongly decreased relative to cylindrical inclusions. We attribute this to the larger deformation of the circular boundary curve for the closed pentameric clover-leaf shape in Fig. 1(B) [see also Fig. 3(A)] as compared to the corresponding open pentameric clover-leaf shape. Allowing for (hypothetical) transitions between different families of boundary curves, the situation becomes more complex. Transitions from a closed pentagonal to an open pentameric clover-leaf shape show a strongly increased gating energy, whereas transitions from a closed pentameric clover-leaf shape to an open pentagonal shape carry a small penalty as far as the elastic membrane deformation energy is concerned. This trend is amplified if pentagonal shapes of the same circumference, rather than of the same cross-sectional area, as the cylindrical reference shape are considered. In summary, Fig. 3(B) indicates that, for the proposed structural models of MscL gating [24]–[27], [33], [35]–[47], the term in Eq. (3) is generally of the same order of magnitude as , with different structural models of MscL displaying a characteristic dependence of the sign and numerical value of on the bilayer hydrophobic thickness.
Figure 4 provides a systematic comparison of the membrane deformation energy associated with different oligomeric states of MscL for the polygonal and clover-leaf boundary shapes inspired by the molecular models in Fig. 1 and Refs. [39], [44] (see Fig. S1). As in Fig. 3B, we used for Fig. 4 the same hydrophobic mismatch for closed and open states of MscL [51], [52]. For the clover-leaf shapes in Fig. 4 we considered shapes which were perturbed by the same amplitude about the cylindrical reference shape in open and closed states. The left-hand panel of Fig. 4(A) shows a clear progression in membrane deformation energy as a function of the oligomeric protein state, with lower-order clover-leaf shapes being energetically favorable compared to higher-order clover-leaf shapes. All clover-leaf shapes induce a membrane deformation energy which is greater than the deformation energy associated with the cylinder model of MscL [see Fig. S2(A) for more comprehensive results]. The elastic energy differences between the open and closed states of clover-leaf shapes are displayed in the right-hand panel of Fig. 4(A). We find that the gating energy of clover-leaf shapes decreases with increasing channel symmetry. Intriguingly, oligomeric states of high enough symmetry yield a gating energy which is reduced relative to cylindrical inclusions of the same cross-sectional area (see Fig. S3 for more comprehensive results).
The left-hand panel of Fig. 4(B) illustrates the membrane deformation energy of the closed state of MscL for trigonal, tetragonal, pentagonal, and hexagonal boundary curves. In contrast to clover-leaf shapes, the membrane deformation energy corresponding to polygonal inclusion shapes decreases with increasing symmetry, and eventually approaches the deformation energy associated with cylindrical inclusions. For membrane inclusions of equal circumference the convergence of the membrane deformation energies induced by polygonal and cylindrical inclusions is rendered more rapid as compared to membrane inclusions of the same cross-sectional area [see Fig.S2 (B)]. The elastic energy differences between the open and closed states of polygonal boundary curves are illustrated in the right-hand panel of Fig. 4(B), and exhibit characteristics which are qualitatively different from the corresponding results for clover-leaf shapes in the right-hand panel of Fig. 4(A). For polygonal shapes the energy difference between the open and closed channel states decreases with increasing symmetry of the membrane inclusion, and is always greater than the elastic gating energy associated with the cylindrical reference shape. These conclusions hold for membrane inclusions of equal circumference as well as inclusions of the same cross-sectional area (see Fig. S3). Polygonal boundary curves with six-fold or higher-order symmetry yield, for the parameter values appropriate for MscL [51], [52], a gating energy which closely approaches the corresponding gating energy associated with the cylinder model of MscL (see Fig. S3 for more comprehensive results). Thus, Fig. 4 predicts systematic trends in the total membrane deformation energy required to accommodate MscL (or other membrane proteins with comparable hydrophobic surfaces) within the bilayer membrane, and in the elastic gating energy, as the oligomeric state and protein shape are being varied.
We now turn to the dependence of the channel opening probability in Eq. (1) on the oligomeric state and hydrophobic shape of MscL. It should be emphasized that we thereby focus solely [11]–[15], [51]–[54] on the lipid bilayer contribution to the total free energy difference between the open and closed channel states, and neglect any contributions to the gating energy due to changes in the internal protein conformation. While it was argued previously [51]–[54] that, in certain situations, the total free energy difference between the open and closed states of MscL can be of the same order of magnitude as the difference in membrane deformation energy between the open and closed states of MscL, other contributions to the free energy difference must generally be considered. Note, however, that our results in Fig. 3(B) indicate that the term in Eq. (3) capturing contributions to the membrane deformation energy due to deviations of the hydrophobic cross section of MscL from the circle is generally of the same order of magnitude as the elastic energy difference calculated previously using the cylinder model of MscL [51]–[54]. Thus, the structure of lipid bilayer deformations associated with different oligomeric states and shapes of MscL is expected to affect the gating characteristics of MscL.
In order to facilitate the systematic investigation of the connection between the oligomeric state and the gating energy of MscL in Fig. 3(B) we employed the parameterization of bilayer-MscL interactions in Refs. [51], [52] and used the same hydrophobic mismatch for closed and open states of MscL. Applying these parameter values to the fits to the structural models in Fig. 1 and Refs. [39], [44] we found the gating curves shown in Fig. 5(A). The tetragonal model of MscL in Fig. 1(A) is seen to gate at a larger tension than the pentagonal model of MscL in Ref. [44], with both models yielding a larger gating tension than the cylindrical reference shape. In contrast, the pentameric clover-leaf model of MscL in Fig. 1(B) produces a smaller gating tension than the hexameric clover-leaf model of MscL, the cylinder model of MscL, as well as the tetragonal and pentagonal models of MscL. Moreover, for a pentagonal shape of MscL in the closed state and a pentameric clover-leaf shape in the open state, Fig. 5(A) predicts a relatively large gating tension. In contrast, the reverse case of a pentameric clover-leaf shape in the closed state and a pentagonal open state yields a markedly smaller gating tension than any of our other models of MscL gating motivated by Fig. 1 and Refs. [39], [44].
Figure 5(B) displays the same gating curves as Fig. 5(A), but using the distinct values of the hydrophobic thickness of the closed and open states of MscL suggested by structural studies of MscL [40], [41], [47]. In this parameterization of bilayer-MscL interactions [53], [54], closed and open states of MscL are distinguished not only by their hydrophobic cross section but also by their hydrophobic thickness. As a result, gating is driven by a more complex interplay between the energetics of thickness deformations and the structure of membrane deformations induced by a non-circular cross section of MscL. In comparison to Fig. 5(A), the gating curves in Fig. 5(B) are shifted to a larger tension into the regime of the measured gating tension [30], [33] for which in Eq. (1). Moreover, for the parameter values used in Fig. 5(B), the gating tension associated with the structural models in Fig. 1 and Refs. [39], [44] is generally larger than the gating tension of the cylinder model of MscL. Contrary to Fig. 5(A), Fig. 5(B) implies that the hexameric clover-leaf model of MscL gates at a smaller tension than the pentameric clover-leaf model of MscL. Similarly as Fig. 5(A), however, Fig. 5(B) predicts that the tetragonal model of MscL gates at a larger membrane tension than the corresponding pentagonal model. Moreover, Figs. 5(A) and 5(B) both imply that for a pentagonal shape of MscL in the closed state, and a pentameric clover-leaf shape in the open state, the gating tension is increased relative to most other scenarios suggested by Fig. 1 and Refs. [39], [44], with the reverse result for the case of a closed pentameric clover-leaf shape and an open pentagonal shape of MscL. Collectively, Fig. 5 shows that, even if only membrane contributions to the gating energy are considered, different oligomeric states and hydrophobic shapes of MscL yield considerable and distinctive modifications of the gating characteristics of MscL.
In analogy to Fig. 4, we have also carried out a systematic comparison between the gating characteristics associated with different oligomeric states of MscL for the polygonal and clover-leaf boundary curves inspired by Fig. 1 and Refs. [39], [44] (see Fig. S4). For this comparison we used, as in Fig. 5A, the same hydrophobic mismatch for closed and open states of MscL [51], [52]. As already suggested by the results in Fig. 4 we found that, for clover-leaf shapes, higher-order oligomeric states gate at a smaller membrane tension. Moreover, depending on the oligomeric state considered, clover-leaf membrane inclusions can gate at a smaller or at a larger tension than the cylinder model of MscL. For polygonal shapes, higher-order oligomeric states are also found to gate at a smaller membrane tension than lower-order oligomeric states but, in contrast to clover-leaf shapes, polygonal channels always gate at a larger tension than the cylindrical reference shape. These features of the gating characteristics of polygonal membrane inclusions do not change if inclusions of equal circumference, rather than equal cross-sectional area, are compared, although the differences in the gating tensions associated with the various oligomeric states of polygonal inclusions become less pronounced.
Inspired by structural studies of MscL [24]–[27], [33], [35]–[47] we have determined the membrane deformation energy associated with a variety of oligomeric states and hydrophobic shapes of MscL. Our analysis focused on the limit of weak perturbations about the cylinder model of membrane proteins, which was employed previously to study bilayer-protein interactions for MscL [51]–[54] as well as for a number of other membrane proteins [11]–[15]. It would desirable to complement the analytic approach developed here with numerical schemes allowing the accurate solution of the elastic membrane equations for complicated protein shapes. Such numerical schemes will be crucial for connecting membrane-mechanical models of bilayer-protein interactions more closely to the shapes of real membrane proteins. Moreover, in our analysis we have focused solely [11]–[15], [51]–[54] on contributions to the total gating energy due to thickness deformations of the bilayer membrane. In particular, we did not consider contributions to the free energy difference between the open and closed states of MscL due to changes in the internal protein free energy. While it has been argued [51]–[54] that, at least for some strains of MscL [27], the thickness deformation energy may play a dominant role in MscL gating, other contributions to the free energy budget must generally be considered.
Our mathematical approach for determining the energetic cost of membrane deformations associated with different oligomeric states and hydrophobic shapes of MscL is general and directly applicable to other membrane proteins. Thus, the methodology developed here establishes a quantitative relationship between the oligomeric state and hydrophobic shape of a membrane protein and the elastic energy required to accommodate the membrane protein within the lipid bilayer membrane. However, the quantitative details of our predictions depend on the parameter values characterizing the hydrophobic shape of the membrane protein under consideration. In particular, crucial inputs for our model are the hydrophobic thickness and cross section of membrane proteins. Recent experimental results [3]–[6], [9], [10] on bilayer-protein interactions suggest that it may be feasible to substantially refine these model inputs to arrive at a more realistic description of protein-induced membrane deformations. For instance, we assumed here that the hydrophobic surface of MscL is perpendicular to the bilayer membrane and of a constant thickness, while a more realistic description of bilayer-MscL interactions would allow [64] for variations in the hydrophobic thickness of MscL along the bilayer-MscL interface.
The physiologically relevant oligomeric states and molecular structures of MscL remain a matter of debate [26], [35]–[37], with tetrameric [46], pentameric [40], and hexameric [39] states of MscL having been reported. The oligomeric state and molecular structure of MscL have so far mainly been studied [24]–[27], [33], [35]–[47] using crystallographic, biochemical, and computational approaches. Our results suggest that, for cases in which there is a significant membrane contribution to the gating energy, functional properties of MscL, such as the predicted discrepancies in the gating energy and gating tension between different oligomeric states and structural models of MscL [24]–[27], [33], [35]–[47], may also be used to shed light on the physiologically relevant oligomeric states and molecular structures of MscL. While we have illustrated our approach for MscL, the methods developed here are general and applicable to other membrane proteins. We predict that the oligomeric state and hydrophobic shape of a membrane protein are reflected in the energetic cost of the lipid bilayer deformations necessary to accommodate the protein within the membrane. Thus, our results suggest that, in addition to the hydrophobic mismatch between membrane proteins and the surrounding lipid bilayer [11]–[17], the symmetry and shape of the hydrophobic cross section of membrane proteins, and resulting structure of elastic membrane deformations, play an important role in the regulation of protein function by bilayer membranes.
In accordance with the standard framework for describing elastic bilayer-protein interactions [11]–[15], [51]–[63], we model MscL as a rigid membrane inclusion inducing bilayer deformations as a result of a hydrophobic mismatch between lipid bilayer and membrane protein. In mathematical terms, the lipid bilayer is represented within the Monge representation of curved surfaces using the functions and , which define the positions of the hydrophilic-hydrophobic interface at the Cartesian coordinates in the top and bottom (outer and inner) membrane leaflets. Focusing on thickness deformations induced by MscL [14], [51]–[54], we consider the elastic energy [18], [20], [56](4)where the thickness deformation field is defined by(5)in which is the equilibrium thickness of the unperturbed bilayer, is the bending rigidity, is the stiffness associated with thickness deformations, and is the membrane tension. Energy functionals of the form in Eq. (4) have been employed in a range of studies [11]–[15], [51]–[63] of membrane deformations induced by MscL as well as other membrane proteins.
The terms and in Eq. (4) provide lowest-order descriptions of the energetic cost of membrane bending and compression or expansion of the lipid bilayer, respectively. For generality we allow for the two tension terms and in Eq. (4), which were employed previously to describe the effects of membrane tension on lipid surface area [18], [53], [54] and on membrane undulations [18]–[20], [51], [52], [56]. While Eq. (4) provides a simple description of protein-induced membrane deformations, more sophisticated models of membrane deformations can be developed [20], [52], [55]–[59] in order to account for detailed properties of lipid bilayers such as lipid structure and spontaneous curvature. Finally, the elastic model of bilayer membranes in Eq. (4) is completed by accounting for the midplane deformations(6)To leading order, midplane deformations decouple from thickness deformations in the total membrane elastic energy [56]. It was found previously [14], [51]–[54] that energetic contributions to MscL gating due to midplane deformations can generally be neglected relative to energetic contributions due to thickness deformations, and we therefore focus here on Eq. (4).
The specific properties of MscL enter Eq. (4) through the boundary conditions at the bilayer-MscL interface [12]–[15], [51]–[54]. For convenience, we specify these boundary conditions along some boundary curve using polar coordinates:(7)(8)where is the unit normal vector along the bilayer-inclusion interface. If MscL is described as a cylindrical membrane inclusion [51]–[54], is a constant and . The quantity corresponds to one-half the hydrophobic mismatch between MscL and the surrounding lipid bilayer, and corresponds to the gradient of the thickness deformation field at the bilayer-inclusion interface. We denote the values of and associated with the closed and open channel states by and , and by and , respectively. The crystallographic structure of the closed state of MscL suggests [40], [41], [54] nm, while it has been proposed [41], [47], [54] that nm for the open state of MscL. To our knowledge, no experimental estimates of the values of and are available for MscL but, within the membrane-mechanical model of MscL gating, these parameters were found previously [51]–[54] to play a minor role compared to and , and are commonly set to zero. We set in all calculations presented here. An approach alternative to that in Eq. (8) would allow [55], [57]–[63] for a free contact slope along the bilayer-inclusion interface.
The membrane-mechanical model of bilayer-MscL interactions outlined above yields a qualitative framework for understanding MscL gating, is in broad agreement [14], [51]–[54] with available experimental data, and provides a machinery for making quantitative predictions. In particular, within the framework of this model, MscL gating is understood on a qualitative level as driven by two competing physical mechanisms. On the one hand, closed channels generally leave a smaller elastic deformation footprint in the membrane, which makes the closed state favorable compared to the open state. On the other hand, in membranes under tension, the increase in membrane area associated with open channels makes this state favorable compared to the closed state. Put differently, MscL gating harnesses the mechanical properties of lipid bilayers for channel function, which penalize the more pronounced membrane deformations which are generally necessary to accommodate larger channels, but favor the relaxation of the tension-inducing loading device [26], [54] brought about by an increased channel area. This physical picture of mechanosensitive gating [14], [51]–[54] relies on the implicit assumption that, in the closed state of MscL, and in Eqs. (7) and (8) are of a similar or smaller magnitude as in the open state of MscL.
While the elastic model in Eq. (4) provides a general description of membrane shape [12]–[15], [18]–[20], quantitative tests of the relevance of this model for mechanosensitive gating rely [14], [51]–[54] on comparing theoretical estimates of to measured values of . In the absence of reliable measurements of in Eq. (2), and presence of large experimental uncertainties, any such comparison can only be of broad character. In the simplest case, the closed and open states of MscL are assumed to take cylindrical shapes with the same hydrophobic thickness, which is then fitted to experimental data. In agreement with the experimental results in Ref. [27], it is thus found [51], [52] that varies from to as the lipid tail length is varied from 16 carboxyl groups to 20 carboxyl groups, and that this variation approximately takes the shape of a quadratic function. This result is obtained at zero tension with the fitted hydrophobic mismatch nm, which corresponds to a hydrophobic thickness of MscL matching a PC12 bilayer and lies in between the aforementioned values of and proposed on the basis of the crystallographic structure of the closed state of MscL [40], [41] and molecular modeling of the open state of MscL embedded in doped bilayers [41], [47]. For a finite tension , which approximately corresponds to the critical gating tension at which in Eq. (1), one finds [54] for the cylinder model of MscL with the values of and proposed on the basis of structural studies of MscL [40], [41], [47] that for a model lipid bilayer. This estimate does not involve any free parameters, and agrees quite well with the corresponding experimental estimate in Refs. [30], [33]. We employ the fitted value nm [51], [52] in Figs. 2–4 and 5(A), as well as Figs. S2, S3, S4, for our systematic study of the effect of protein shape on the membrane deformation energy and gating tension. This parameterization of bilayer-MscL interactions allows us to avoid any spurious effects resulting from different hydrophobic mismatches in the closed and open channel states. In Fig. 5(B) we use the estimates nm and nm suggested in Refs. [40], [41], [47], [54].
We follow Refs. [11]–[15], [51]–[63] and use Eq. (4) with the boundary conditions in Eqs. (7) and (8) as our basic model of the membrane deformations induced by MscL. The Euler-Lagrange equation associated with Eq. (4) is given by(9)To proceed, we introduce the function(10)in terms of which Eq. (9) reduces to(11)where(12)The solution of Eq. (11) is of the form [11], [65](13)where are solutions of the Helmholtz equations(14)For the exterior of a circle of radius , the above Helmholtz equations are readily solved by separation of variables [65], [66]. Thus, for the exterior of a circle, the solution of Eq. (11) can be written as the Fourier-Bessel series(15)in which(16)where and are constants, are modified Bessel functions of the second kind, and we have assumed that membrane deformations decay away from the membrane inclusion [57]. At each order in the Fourier-Bessel series in Eq. (15), two boundary conditions at the membrane-inclusion interface are required to fix all constants and .
Boundary curves are obtained by fitting the Fourier representation of ,(17)in which we take(18)and , to the transmembrane cross sections of MscL in Fig. 1 and Refs. [39], [44]. We focus here on the weak perturbation limit of Eq. (17) and only consider leading-order terms in .
The molecular structures in Fig. 1 and Refs. [39], [44] suggest two basic families of as models of the hydrophobic cross section of MscL: polygonal boundary shapes and clover-leaf boundary shapes. Polygonal shapes are obtained using the Fourier representation of regular -gons in the complex plane [67],(19)in which is the imaginary unit and the tetragonal and pentagonal oligomeric states in Fig. 1(A) and Ref. [44] correspond to and , respectively. Higher orders of in Eq. (19) yield increasingly sharp polygonal corners. For all polygonal shapes in this manuscript we considered terms up to in Eq. (17). As described in the Results section, all parameters in Eq. (17) are then fixed for polygonal shapes by setting the areas of polygonal shapes equal to the cross-sectional areas of closed and open MscL suggested by structural studies [27], [33], [40]–[45], [47] and used in previous membrane-mechanical models of MscL gating [51], [52], [54].
The clover-leaf shapes in Fig. 1 are obtained using boundary curves of the form(20)where the pentameric and hexameric clover-leaf shapes in Fig. 1(B) and Ref. [39] correspond to and , respectively. As for polygonal shapes, the overall coefficient in Eq. (20) is determined by fixing the area of clover-leaf shapes in closed and open channel states [27], [33], [40]–[45], [47], [51], [52], [54]. For the clover-leaf shapes considered in Figs. 2–5, we determined through fits to the models of MscL shape shown in Fig. 1 and Ref. [39], yielding (closed pentameric clover-leaf shape), (open pentameric clover-leaf shape), and (closed and open hexameric clover-leaf shapes). For the model clover-leaf shapes shown in Figs. S1, S2, S3, S4 we used for closed states and for open states so that the amplitude of perturbations about the cylindrical reference shape, , took the same magnitude in closed and open states.
In general, and in the boundary conditions in Eqs. (7) and (8) at may both exhibit an angular dependence, and our approach is able to handle such cases. Here we focus on the effect of deviations from the circular shape on the elastic membrane deformations induced by MscL. For simplicity, we therefore take and to be constants. Assuming small deviations from circularity in Eq. (17), we use a perturbative approach and expand [68] at the boundary curve around to leading order in ,(21)in which(22)from the general solution in Eq. (15) to in , where(23)for . Note, in particular, that any term in Eq. (15) involving an angular dependence must at least be of in . Similarly,(24)to leading order in , in which(25)from the general solution in Eq. (15) to in , where(26)for , and is determined by the terms in Eq. (15).
Thus, using Eqs. (21) and (24), we can recast the boundary conditions in Eqs. (7) and (8) for non-cylindrical inclusions as boundary conditions for cylindrical inclusions of variable hydrophobic thickness,(27)(28)to leading order in . Matching Eqs. (27) and (28) with Eq. (15) at each order in the Fourier-Bessel series, we find(29)(30)(31)where, for , and . Equations (29)–(31) together with Eq. (15) constitute, in the limit of weak perturbations about cylindrical inclusion shapes, the general solution of the membrane deformation profile for arbitrary oligomeric states of MscL.
The membrane deformation energy associated with the equilibrium deformation profile in Eq. (15) with Eqs. (29)–(31) is obtained by evaluating the surface integral in Eq. (4). To this end, we note from Eq. (11) that(32)Hence, we can use Gauss's theorem in the plane to transform the surface integral in Eq. (4) to a line integral:(33)where is a constant. For simplicity, we choose the zero of the energy such that .
To evaluate the integrals in Eq. (33) it is convenient to note that . Substituting the Fourier-Bessel series in Eq. (15) into Eq. (33) then generates integrals of the form(34)Thus, we find the elastic thickness deformation energy(35)where(36)(37)(38)(39)(40)(41)for . Equation (35) with Eqs. (36)–(41) and Eqs. (29)–(31) provides the general solution of the thickness deformation energy in Eq. (4) for arbitrary oligomeric states of MscL in the limit of weak perturbations about cylindrical inclusion shapes.
The deformation profiles in Fig. 2 were obtained from Eq. (15) with Eqs. (29)–(31), the energy curves in Figs. 3, 4, S2, and S3 were obtained from Eq. (35) with Eqs. (36)–(41) and Eqs. (29)–(31), and the gating curves in Figs. 5 and S4 were obtained from Eq. (1) together with Eq. (35), Eqs. (36)–(41), and Eqs. (29)–(31). For all plots we used the elastic moduli [54] and , with for Figs. 2–4, S2, and S3. The results in Figs. 2–4, 5(A), and S2, S3, S4 were obtained with nm [51], [52]. For Fig. 5(B) we used the estimates nm and nm [40], [41], [47], [54]. We used a bilayer hydrophobic thickness corresponding to PC14 lipids for Fig. 1, to PC18 lipids for Figs. 4, 5(A), and S4, and to PC14 lipids for Fig. 5(B). We related membrane hydrophobic thickness to PC lipid tail length using the simple interpolation described in Ref. [51].
The primary accession numbers (in parentheses) from the Protein Data Bank are: Pentameric MscL (2OAR, formerly 1MSL; Resolution of 3.50 Å; Ref. [40]) and tetrameric MscL (3HZQ; Resolution of 3.82 Å; Ref. [46]).
|
10.1371/journal.ppat.1006558 | O-acetylation of the serine-rich repeat glycoprotein GspB is coordinated with accessory Sec transport | The serine-rich repeat (SRR) glycoproteins are a family of adhesins found in many Gram-positive bacteria. Expression of the SRR adhesins has been linked to virulence for a variety of infections, including streptococcal endocarditis. The SRR preproteins undergo intracellular glycosylation, followed by export via the accessory Sec (aSec) system. This specialized transporter is comprised of SecA2, SecY2 and three to five accessory Sec proteins (Asps) that are required for export. Although the post-translational modification and transport of the SRR adhesins have been viewed as distinct processes, we found that Asp2 of Streptococcus gordonii also has an important role in modifying the SRR adhesin GspB. Biochemical analysis and mass spectrometry indicate that Asp2 is an acetyltransferase that modifies N-acetylglucosamine (GlcNAc) moieties on the SRR domains of GspB. Targeted mutations of the predicted Asp2 catalytic domain had no effect on transport, but abolished acetylation. Acetylated forms of GspB were only detected when the protein was exported via the aSec system, but not when transport was abolished by secA2 deletion. In addition, GspB variants rerouted to export via the canonical Sec pathway also lacked O-acetylation, demonstrating that this modification is specific to export via the aSec system. Streptococci expressing GspB lacking O-acetylated GlcNAc were significantly reduced in their ability bind to human platelets in vitro, an interaction that has been strongly linked to virulence in the setting of endocarditis. These results demonstrate that Asp2 is a bifunctional protein involved in both the post-translational modification and transport of SRR glycoproteins. In addition, these findings indicate that these processes are coordinated during the biogenesis of SRR glycoproteins, such that the adhesin is optimally modified for binding. This requirement for the coupling of modification and export may explain the co-evolution of the SRR glycoproteins with their specialized glycan modifying and export systems.
| Bacteria express a variety of structures on their surfaces that promote interactions with the host. The serine-rich repeat (SRR) glycoproteins have emerged as an important group of surface proteins on many streptococci and staphylococci that mediate their attachment to a range of host tissues. The SRR glycoprotein GspB of Streptococcus gordonii is a key virulence factor for developing infective endocarditis (an infection of heart valves), due to its ability to mediate bacterial binding to human platelets. Like all SRR adhesins, GspB is extensively glycosylated in the cytoplasm and is then exported to the cell surface via a specialized transporter, the accessory Sec (aSec) system. The reason why a dedicated transport system is necessary for export has remained elusive. Here we show that in addition to its role in export, one aSec component (Asp2) is also an enzyme that can transfer an acetyl group to N-acetylglucosamine residues on GspB during transport. Loss of GspB acetylation significantly impaired the ability of streptococci to bind human platelets, indicating that acetylation is essential for the binding activity of GspB. These findings add to our knowledge of novel protein glycosylation mechanisms utilized by bacteria, and highlight the importance of SRR glycan acetylation in bacterial-host interactions.
| The serine rich repeat (SRR) glycoproteins are a large family of adhesins on the surface of many Gram-positive bacteria. Expression of the SRR adhesins has been directly correlated with colonization and the ability of these organisms to cause invasive disease [1] [2] [3] [4] [5] [6]. The biogenesis of these adhesins involves the intracellular O-linked glycosylation [7] [8] and transport of the glycoproteins to the bacterial surface by the accessory Sec (aSec) system [9] [10]. Glycosylation is initiated by a two-protein glycosyltransferase (Gtf) complex (GtfAB) that adds N-acetylglucosamine (GlcNAc) to serine and threonine residues within the SRR domains of the adhesins [7] [11]. Other Gtfs sequentially further modify the SRR domains by adding other glycan moieties to the GlcNAc core. The number and type of these Gtfs varies considerably between species, as do the resulting extent of glycan modification [12] [13] [14].
The aSec system is a dedicated transporter that exclusively mediates the transport of SRR glycoproteins [9] [15]. This system is comprised of SecA2 (the motor protein), SecY2 (the translocon channel) and three to five accessory Sec proteins (Asps) that are essential for SRR glycoprotein transport (reviewed in [16]). The exact role of the Asps in transport is uncertain. Our previous studies with the SRR adhesin GspB of Streptococcus gordonii have identified numerous protein-protein interactions between Asps1-3, which are located intracellularly. Disruption of some of these interactions abolishes aSec transport, indicating a coordinated role for the Asps in export [17]. Asp2 and Asp3 can bind GspB directly and Asp1-3 appear to enhance the binding of the GspB preprotein to membrane-associated SecA2 [18] [19]. These findings indicate that the Asps are essential for key stages of SRR protein translocation, either by acting upon the substrate or other members of the aSec system [16].
Although the post-translational modification and export of the SRR adhesins have been largely viewed as separate pathways, our previous studies indicate that at least Asp2 may have a role in both processes. Modeling of the predicted structure of Asp2 suggests that it shares similarities with numerous glycan-modifying enzymes [20]. Moreover, a point mutation within a predicted catalytic triad of Asp2 resulted in an altered glycoform of GspB, which was exported freely by the aSec system, but had reduced binding to its platelet-localized ligand, sialyl-T antigen. These findings suggested that in addition to the known Gtfs, Asp2 may also modify the glycan on GspB [20]. However, the precise role of Asp2 in the post-translation modification of the SRR adhesins was unknown. Here, we show that Asp2 mediates the O-acetylation of GlcNAc residues on GspB. Furthermore, this modification is transport dependent, occurring exclusively when the substrate is transported through the aSec pathway. These results indicate that Asp2 serves as a nexus for the two major processes in SRR adhesin biogenesis, through its role in both transport and glycan modification. The finding that the acetylation of GspB only occurs during aSec transport indicates that glycan modification and transport are tightly linked processes and explains at least in part the need for a dedicated export system.
We have shown previously that the site-specific replacement of residues that comprise a predicted Ser362-Asp452-His482 catalytic triad within Asp2 altered the reactivity of GspB with the GlcNAc-specific lectin sWGA, indicating a change in the glycan decorating the adhesin [20]. However, it remained unclear as to precisely how the glycan had been altered, as a consequence of the Asp2 mutations. The GtfAB complex mediates the O-linked transfer of GlcNAc to the SRR regions of GspB [7] [8]. Two additional Gtfs (Nss and Gly) sequentially add glucose residues to this glycan moiety (Fig 1B, S1 Fig and S2 Fig) (consistent with glycosylation of the SRR adhesin Fap1 [14]). Based on these findings, we first examined which components of the glycan were affected by an Asp2S362A mutation. To assess changes in glycosylation, we used two variants of GspB (GspB736flag and GspB1060flag) (Fig 1A), containing C-terminal truncations and a 3xFLAG tag. These GspB variants lack cell wall anchoring domains, and are thus secreted into the culture media by S. gordonii strain M99, thereby facilitating the analysis of transport activity.
As compared with GspB variants made in a WT background, both GspB736flag and GspB1060flag showed a marked increase in sWGA reactivity, when exported by the Asp2S362A expressing M99 variant, (Fig 1C, lane 2 versus 1 and lane 6 versus 5). Higher sWGA reactivity was seen in the longer GspB forms (Fig 1C and Fig 1D) suggesting the glycan change resulting from the Asp2S362A mutation is present throughout the SRR2 region. The amounts of GspB exported were comparable among WT and Asp2 mutant strains as determined by anti-FLAG reactivity (Fig 1C and 1D), indicating that the changes in sWGA reactivity seen in GspB as a consequence of the Asp2S362A mutation were not due to increased GspB production, but instead, resulted from differences in glycan composition.
To better define the glycan modified by Asp2, the Asp2S362A mutation was introduced to a variant of M99, in which gly and nss had been deleted. Loss of gly and nss results in GspB variants glycosylated only by GtfAB, which is known to deposit GlcNAc only (Fig 1B, [7] [11]). Mutagenesis of asp2 in this variant also resulted in forms of GspB736flag and GspB1060flag with increased sWGA binding (Fig 1C, lanes 4 versus 3 and lanes 8 versus 7). In contrast, no sWGA reactivity was observed with the GspB variants expressed by gtfA-deletion strains (S3 Fig), demonstrating that Asp2 modifies GlcNAc exclusively.
We previously found via predictive modeling that the putative catalytic triad of Asp2 was conserved in numerous esterases [20]. However, O-linked GlcNAc typically does not have any ester-linkages. As some esterases and acyltransferases share a common fold and have similar mechanisms of action [21], we explored whether Asp2 functions as an acyltransferase that modifies GlcNAc residues. Since ester linkages to carbohydrates are base-sensitive [22], we assessed whether mild-base ester hydrolysis (saponification) altered the sWGA reactivity of GspB. Culture media containing secreted GspB736flag or GspB1060flag were treated with 100 mM NaOH and then probed with sWGA.
In control studies, mild-base treatment did not alter the overall amounts of either GspB variant (Fig 2A) demonstrating that the conditions used for saponification did not degrade GspB. However, saponification did significantly increase the sWGA reactivity of secreted wild-type GspB736flag and GspB1060flag (Fig 2B, lanes 1 versus 5 and lanes 3 versus 7). In contrast, the sWGA reactivity of the same two variants exported from an Asp2S362A background was unchanged by this treatment (Fig 2B, lanes 2 versus 6 and lanes 4 versus 8). Moreover, the increase of sWGA reactivity following mild base treatment of both GspB variants produced in a WT Asp2 background was comparable to the sWGA reactivity seen from their untreated catalytic mutant counterparts (Fig 2B, lanes 2 versus 5 and lanes 4 versus 7). These findings demonstrate that the increase in sWGA reactivity seen in Asp2S362A mutant strains is due to a loss of an ester-linked chemical group on GlcNAc residues within the glycan of GspB, and suggest that Asp2 is a GlcNAc modifying enzyme.
To examine the influence of Asp2 activity on glycosylation, we subjected the Δgly-nss glycoform of GspB736flag from both Asp2 and Asp2S362A backgrounds to Q-TOF LC/MS analysis with collision-induced dissociation (CID) fragmentation. Because glycosylation of GspB occurs in the highly repetitive SRR2 region, which lacks Lys and Arg residues, we digested GspB736flag with trypsin, Lys-C, and Glu-C proteases simultaneously, to achieve better peptide coverage. To detect glycopeptides within the digests, we searched MS/MS scans for the production of diagnostic GlcNAc oxonium ions (m/z 204.08) that were generated from broken O-glycosidic linkages. Four major precursor ions were found in the extracted ion chromatogram (peaks 1 to 4) of the GspB736flagΔgly-nss glycopeptides (Fig 3A). MS/MS analyses of these ions displayed characteristic neutral loss of GlcNAc (m/z = 203.08) and provided modest coverage of y- and b- type ions within a tolerance of 0.02 (Fig 3C, S4A, S5A and S6A Figs and S1 Table). These glycopeptides were mapped to the SRR2 region of GspB and the number of GlcNAc modifications could be determined (Table 1 and S1 Table). However, this method did not allow for the precise mapping of the O-GlcNAc sites within the glycopeptides.
Fractions containing the four SRR2-derived glycopeptides were found to elute from the reverse-phase column a second time with respective longer retention times (peaks 1* to 4* of Fig 3A). Each of the additional precursor ions detected in these fractions were 42.01 Da larger in mass compared to their corresponding partner glycopeptides, indicating the presence of an acetyl group. Their MS/MS spectra displayed neutral loss of 245.09 Da, and fragment ions at m/z 246.09 were also detected, which is consistent with an O-acetyl GlcNAc oxonium ion (Fig 3D, S4B, S5B and S6B Figs and S1 Table). Although the exact position of the O-acetylated GlcNAc residue within the glycopeptide could not be determined by MS/MS, some of the O-acetylated glycopeptides formed two peaks in the extracted ion chromatogram, suggesting that alternate sites of addition change peptide retention time. These findings directly demonstrate that SRR2 glycopeptides of GspB736flag from a Δgly-nss background contain variable subpopulations of unmodified and O-acetylated GlcNAc residues.
The same glycopeptides were detected in GspB736flag from the Δgly-nss asp2S362A background and the extracted ion chromatogram indicated that their degree of glycosylation was not affected (Fig 3). However, whereas the LC chromatogram suggested the presence of a peak corresponding to peak 1* of Fig 3A, the extracted ion chromatogram was completely devoid of any ions corresponding to O-acetylated glycopeptides (S7 Fig). Of note, saponification of the above four acetylated glycopeptides of GspB736flag resulted in an extracted ion chromatogram identical to that seen of GspB736flag from an Asp2S362A background (S8 Fig). These data indicate that Asp2 is responsible for the O-acetylation of GlcNAc residues within the SRR2 region of GspB and that GlcNAc O-acetylation results in the decrease in sWGA binding to secreted GspB observed above.
To directly assess the enzymatic activity of Asp2, we examined the ability of the recombinant protein to hydrolyze the acetyl donor p-nitrophenyl acetate (pNP-Ac). Of note, many acetyltransferases function as weak esterases in the absence of acceptor substrates when assayed in vitro as water will serve as the acceptor ligand for the acetyl group resulting in the release of acetate [23]. Using a MalE-Asp2H6 fusion protein and its predicted catalytic mutant, MalE-Asp2S362A-H6 (Fig 4A), hydrolysis of pNP-Ac was only seen with wild-type Asp2. This hydrolysis was increased in the presence of the glycosylated SRR1 domain of GspB (Fig 4B). MalE-Asp2S363AH6 only exhibited pNP-Ac hydrolysis levels on par with background levels and failed to further stimulate pNP-Ac hydrolysis following co-incubation with the SRR1 acceptor substrate. Western blot analysis of SRR1 after co-incubation with pNP-Ac and MalE-Asp2H6 revealed no change in sWGA reactivity, an indicator of GspB glycan O-acetylation (S9 Fig), suggesting that in vitro, the conditions used above do not entirely recapitulate the conditions in vivo for the transfer of O-acetyl groups to the glycan. Nonetheless, collectively these findings demonstrate that the Asp2 has O-acetylesterase activity and can hydrolyze acetyl donor substrates.
Asp2 is required for transport of GspB by the aSec system, with deletion of asp2 abolishing GspB export and resulting in the retention of the SRR adhesin in the bacterial cytosol [9] [10] [17] [20]. To assess whether a loss Asp2 catalytic activity affects GspB transport, we compared the export of GspB736flag and GspB1060flag in a series of Asp2 catalytic mutants. As compared with the WT strain expressing Asp2, GspB736flag and GspB1060flag were comparably transported by isogenic variants expressing Asp2S362A, Asp2E452A or Asp2H482A. The levels of the SRR adhesins detected in the culture supernatant were similar to those observed in the WT strains, while only trace amounts were detected in the protoplasts (Fig 5, S10 Fig, lane 2–5). To ensure that the enzymatic activity of Asp2 was entirely abolished, we also tested a variant containing an alanine replacement in all three putative catalytic residues. As was seen with the single amino acid substitutions, aberrantly glycosylated GspB was freely secreted by this variant, indicating that aSec transport was intact (Fig 5, S10 Fig, lane 6–7). Collectively, these findings show that the loss of Asp2 catalytic activity has no impact on aSec transport, and thus, the aberrant glycoform generated by Asp2 mutagenesis is not an artifact of altered transport. Instead, these results demonstrate that Asp2 is a bifunctional protein that directly acetylates GlcNAc moieties on GspB.
When export of GspB was abolished by an asp2 or secA2 deletion, the retained, intracellular glycoform of the adhesin differed in sWGA reactivity, as compared to the WT, secreted glycoform (Fig 6A). In particular, the intracellular glycoforms of GspB736flag showed a high level of sWGA reactivity, unlike the wild-type secreted form, and instead resembled the GspB glycoform secreted by the Asp2S362A mutant. This indicated that the intracellular glycan of GspB lacked O-acetylated GlcNAc, suggesting that acetylation may be transport-dependent.
To further address this issue, we expressed GspB736flag in an isogenic variant of M99 (PS1226), in which aSec export was abolished because of a short, in-frame deletion within secA2 [24]. As expected, the GspB variants were entirely retained in the cytosol (Fig 6B). Intracellular GspB736flag showed high levels of sWGA reactivity, which were unchanged following mild-base treatment (Fig 6C, lanes 3 & 4 versus lanes 1 and 2). These findings indicate that the GlcNAc moieties on intracellular GspB are not acetylated, suggesting that Asp2 does not modify GspB independently of transport, unlike GtfAB, Gly and Nss [25]. Instead, O-acetylation of GlcNAc by Asp2 appears to only occur concomitant with substrate transport.
We next asked whether Asp2-mediated acetylation of GspB was specific to the aSec pathway. To address this question, we compared the acetylation of GspB736flag transported via the aSec system, with that of GspB736flag*G3, which contains G75L/G79A/G80C mutations within the hydrophobic core of the signal peptide. This altered signal peptide re-routes the preprotein to the general Sec pathway [26]. The GspB variants were expressed in strains M99, M99asp2S362A, or in two strains deficient in aSec transport (M99ΔsecA2 and M99Δasp123). Levels of GspB736flag*G3 transported via the general Sec pathway were comparable to those of GspB736flag transported by the aSec pathway, as measured by Western blotting with anti-FLAG. Of note, the sWGA reactivity of GspB736flag secreted through the general Sec pathway was similar to that of GspB736flag exported from an Asp2S362A mutant (Fig 7B, lanes 3 and 4 versus 2), even when Asp2 was present (M99ΔsecA2). Moreover, upon base treatment, GspB736flag secreted via the general Sec was resistant to saponification, displaying no significant changes in sWGA reactivity (Fig 7B, lanes 3 and 4 versus Fig 7C lanes 3 and 4). These findings show that GspB736flag exported via the general Sec pathway does not undergo acetylation, even when Asp2 is present. Instead, acetylation of GspB by Asp2 only occurs during aSec transport.
We have previously shown that the Asp2-dependent modification of GspB is essential for M99 to bind to immobilized sialyl-T antigen [20]. This glycan is found on platelet glycoprotein GPIbα, and is the major ligand on platelets for GspB [27]. Reduced platelet binding via GspB is associated with decreased virulence in an animal models of infective endocarditis [1], [28]. To determine whether loss of acetylation of GspB affects this interaction, we directly compared the platelet binding by M99, an isogenic Asp2S362A mutant, and other M99 variants expressing altered GspB glycoforms.
As expected, M99 exhibited high levels of binding to human platelets, as compared with the ΔgtfA strain (PS666)(Fig 8A), which does not express GspB on the bacterial surface due to instability of the protein from a loss of glycosylation [12] (Fig 8B). Consistent with our previous reports, deletion of gly and nss, Δgly-nss (PS3319) led to only a small decrease in platelet binding compared with M99. In contrast, M99 expressing the Asp2S362A catalytic mutant (PS3536) had significantly reduced binding to platelets, comparable to those observed with the ΔgtfA deletion strain. These results indicate that the O-acetylation of GlcNAc moieties on GspB is essential for adhesive properties of the glycoprotein.
The biogenesis of the SRR adhesins is a complex process, involving both their intracellular glycosylation and export to the cells surface by the aSec system (reviewed in [16]). Studies in S. gordonii and S. parasanguinis have shown that glycosylation is mediated by a series of glycosyltransferases that sequentially add glycans to the SRR regions of the adhesin [13], [25]. The deposited glycan can range from a single O-linked GlcNAc residue to more complex glycan structures, depending on the number and type of Gtfs encoded within each SRR adhesin-aSec operon [13], [14]. Although the roles of the Asps in export are not fully defined, Asps1, 2 and 3 have been shown to enhance the engagement of the SRR preprotein with SecA2 [19], while Asp4 and Asp5 form a membrane complex with SecY2 [29].
Glycosylation and transport of the SRR adhesins have been viewed as independent processes, in part because SRR glycosylation can be reconstituted in vitro [14] [25] and because glycosylation of the SRR adhesin is not required for aSec transport [17]. However, our results demonstrate that the modification and transport of the SRR adhesins are linked by Asp2. In addition to its essential role in aSec transport, Asp2 functions as an enzyme mediating the O-acetylation of proximal GlcNAc residues within the SRR glycan. The catalytic activity of Asp2 was entirely dispensable for GspB export, thus demonstrating that Asp2 is a bi-functional protein mediating separate events in biogenesis.
Asp2 is found throughout all SRR adhesin-aSec operons but not in other loci, and its Ser-Asp-His catalytic triad is uniformly conserved [20]. These findings, along with the presence of O-acetylated GlcNAc moieties within the glycan of the Srr1 adhesin (one of two SRR adhesins expressed on the surface of S. agalactiae) [30], suggest that acetylation is likely to be a common glycan modification on SRR glycoproteins. It was previously hypothesized that the OatA/B system, which mediates the MurNAc/GlcNAc O-acetylation of bacterial peptidoglycan (PG) [31] [32], might be responsible for the acetylation of Srr1 [30]. Our findings indicate, however, that Asp2 is responsible for this modification. Indeed, mutagenesis of S. gordonii Asp2 resulted in the complete loss of O-acetylated GlcNAc on GspB, suggesting that Asp2 may be the sole O-acetyltransferase modifying the SRR adhesins.
The loss of GlcNAc O-acetylation had a profound effect upon adhesin function, where GspB-mediated streptococcal binding to human platelets was markedly reduced, to levels seen with a GspB-deficient strain. This differed significantly from what was seen with M99 lacking Nss or Gly modifications, where binding was minimally affected. The precise mechanism by which the loss of O-acetyl groups impairs the binding of M99 to platelets is as yet unknown. Since a non-glycosylated recombinant form of the binding-region of GspB binds to platelets with relatively high affinity [33], it is likely that acetylation of the SRR glycan is needed to maintain the proper conformation of GspB for binding to its platelet receptor. It is also possible that acetylation alleviates occlusion of the binding region caused by glycosylation of the neighboring SRR regions. We have previously shown that loss of GspB-mediated platelet binding (either by deletion of the adhesin or mutagenesis of the binding region) is associated with reduced virulence in animal models of streptococcal endocarditis [1] [28], and thus it is highly likely that acetylation of GspB is required for maximal pathogenicity. Indeed, as GspB O-acetylation is critical for platelet binding, Asp2 could provide an effective antimicrobial target, as has been proposed for other bacterial O-acetyltransferases linked to virulence [31], [23].
Our finding that the O-acetylation of GspB was present only when the substrate had undergone aSec transport indicates that these two processes are coordinated, with Asp2 serving as a nexus for biogenesis. As glycan O-acetylation is necessary for optimal GspB function, the coordination of this modification with transport may serve to assure that the SRR adhesion is fully functional. Moreover, analysis of glycopeptides from secreted GspB736flag revealed that not all GlcNAc residues were O-acetylated (Table 1), further suggesting that there may be a requirement to limit or regulate this modification.
These results also provide additional insights as to why a dedicated transporter is needed for the SRR adhesins. The general Sec system is responsible for exporting a large number and variety of substrates, the export of some being essential for cell viability [34]. It is possible that the requirement to coordinate a post-translation modification with export via the Sec pathway would impede export of other Sec substrates, which could potentially be detrimental to the cell. Thus, having a dedicated transporter enables the cell to alleviate any undue pressure upon its secretome. It is also possible that O-acetylated glycan is incompatible with engagement of the Sec machinery. Our findings show that the Sec system can export a non-acetylated SRR glycoform (Fig 7). However, just as O-acetylation affects GspB binding to its host ligand, this modification could prevent interactions with the Sec machinery. The latter would explain in part the evolution of a SecA2/SecY2 paralogue, to accommodate an acetylated substrate.
We have previously shown Asp1-3 can localize with SecA2 at the bacterial membrane to facilitate translocation (reviewed in [16]) and our current findings suggest that Asp2 acetylates GspB at this same location (Fig 9). These findings, along with the absence of O-acetylated glycoforms within the cytosol, suggest that acetylation occurs at the membrane and in association with aSec transport. Collectively, our results indicate the aSec system is a highly specialized export and modification system, where linking of these two processes via Asp2 ensures that the correct SRR glycofrom is expressed on the bacterial surface. The coupling of glycan O-acetylation with transport may serve as means by which this critical modification can be more precisely controlled, and may further explain why the aSec system has evolved as a separate protein secretion system. The roles of other aSec components towards SRR glycoprotein O-acetylation and export are currently under active investigation.
Human platelets were collected from volunteers, under a protocol approved by the UCSF Committee on Human Research (IRB number 11–06207).
The bacterial strains and plasmids used in this study are listed in Table 2. S. gordonii strains were grown in Todd-Hewitt broth (THB, Becton, Dickinson and Company) or on 5% sheep blood agar (Hardy Diagnostics) at 37°C in a 5% CO2 environment. Antibiotics were added to the media at the following concentrations: 60 μg mL-1 erythromycin and 100 μg mL-1 spectinomycin. E. coli strains XL1-Blue and BL21(DE3) were grown in Luria-Bertani (LB) broth or on LB agar containing 30 μg mL-1 kanamycin, 100 μg mL-1 ampicillin, 50 μg mL-1 spectinomycin or 300 μg mL-1 erythromycin when appropriate.
Routine molecular biology techniques for cloning, sequencing and PCR amplification were performed by as described previously [35]. Chromosomal DNA was isolated from S. gordonii according to Madoff et al. (1996) [36]. Plasmid DNA was isolated from E. coli using miniprep columns (Qiagen). DNA restriction and modification enzymes were used according to manufacturer’s recommendations (NEB). E. coli cells were transformed following CaCl2 treatment, while S. gordonii was transformed as described previously [9].
Mutagenesis of asp2 was conducted using the QuikChange Lightning site-directed mutagenesis kit (Agilent Technologies) as previously described [20] and either pET.H6Asp2 or pCOLA.H6asp123 as the template. Following PCR, Dpn I was added to the reaction mixture to remove the original methylated plasmid. The remaining plasmids in the reaction mixture were then used to transform E. coli XL1-Blue, and the resulting clones were screened for the correct point mutations by DNA sequencing (Sequetech DNA, Mountain View).
Point mutations within the asp2 gene of S. gordonii strain M99 were achieved through allelic replacement. Using primers previously described [20], the codons Ser-362, Glu-452 and His-482 of asp2 were replaced with those for Ala by site-directed mutagenesis using the plasmid pCOLA.H6asp123 (encoding asp123) as a template. Plasmids were then used to transform M99 Δasp2::spec strains expressing either GspB736flag or GspB1060flag, resulting in a replacement of the spectinomycin cassette with the mutated asp2. Transformants were plated on sheep blood agar plates and scored for the loss of spectinomycin resistance. Chromosomal DNA was isolated from spectinomycin sensitive clones and the asp2 gene was PCR amplified and sequenced to confirm the correct mutant replacement.
The construction of the GspB736flag*G3 (i.e. G75L/G79A/G80C) signal peptide mutant strain PS1129 and the GspB736flag*G3 ΔsecA2 strain PS1146 have previously been described [26]. Replacement of the gly-nss genes within PS1146 by a spec cassette was performed using a modification of the method used to delete gly or nss individually [12]. In brief, a chromosomal segment upstream of gly was amplified by PCR using primers glyKO4 and glyKO5, and a chromosomal segment downstream of nss was amplified using primers nssKO3 and nssKO5. The fragments were cloned upstream or downstream, respectively, of the spec gene in pS326. The resulting plasmid, pGLYNSSK, was introduced to PS1146 by natural transformation, and allelic replacement was monitored by selection on spectinomycin. Similarly, replacement of the gly-asp3 genes in PS1129 with a spec cassette was accomplished by replacing the upstream chromosomal segment of asp3 in pORF3K [12] with that of the chromosomal segment upstream of gly as described above. The resulting plasmid, pGLYASP3K, was introduced to PS1146 by natural transformation, and allelic replacement was monitored by selection on spectinomycin.
Overnight cultures of M99 were diluted 1:6 in fresh THB, grown for 5 hr at 37°C, and the cells harvested by centrifugation. For analysis of secreted proteins, samples of clarified culture media were either mixed with protein sample buffer (Novagen) prior to SDS-PAGE separation and Western blot analysis or used directly in saponification analysis. For analysis of non-exported proteins, pelleted cells were resuspended in protoplast lysis buffer (PLB: 60 mM Tris pH 6.8, 150 mM NaCl, 10% raffinose and 0.5U μL-1 mutanolysin). The PLB suspensions were incubated for 1 hr at 37°C, boiled in protein sample buffer, followed by SDS-PAGE and Western blot analysis. For saponification analysis of non-exported proteins, pelleted cells were suspended in lysis buffer (LB: 60 mM Tris pH 7, 150 mM NaCl) and lysed using a MiniBeadbeater (Biospec) by using 2× 60-s bursts at RT at full speed. Insoluble material was pelleted by centrifugation, and lysates were normalized for protein concentration prior to saponification (see below).
Secreted GspB736flag was purified from M99 culture supernatants (14 Liters) as previously described [10]. In brief, M99 strains PS3309 and PS3540 were grown overnight in THB and cells removed by centrifugation. Proteins secreted into the culture media were precipitated overnight in ammonium sulfate (NH4)2SO4 (25% final concentration), recovered by centrifugation and reconstituted in Tris buffered saline (TBS: 50 mM Tris pH 7.5, 150 mM NaCl). Glycosylated GspB736flag was subsequently purified from TBS under native conditions by affinity chromatography using sWGA agarose (Vector) and eluted in 300 mM GlcNAc. GspB736flag fractions were pooled, concentrated by ultrafiltration using an Amicon Ultra centrifugal filter device (100 kDa cutoff) and reconstituted in dH2O until further analysis.
To quantify differences in GspB transport and glycosylation, blots of GspB736flag or GspB1060flag were incubated at room temperature (RT) simultaneously with mouse anti-FLAG antibody (Sigma) and sWGA (Vector) used at concentrations of 1:5000 and 0.4 μg mL-1 respectively. Blots were incubated for 2 hr, followed by another 90 min incubation with a 1:20,000 dilution of both HiLyte Fluor800-anti-Mouse IgG and HiLyte Fluor680-streptavidin (Anaspec). Immunoreactive bands were visualized using an infrared imager (LI COR Biosciences) at both 680 nm and 800 nm. Band intensity was analyzed using Odyssey v3.0 software.
Base-promoted ester-hydrolysis was achieved through mild NaOH treatment. Culture supernatants containing secreted GspB variants were buffered in 10 mM Tris pH 7 and incubated with 100 mM NaOH for 1 hr at 37°C. Protoplast generated by beadbeating were clarified by centrifugation at 14,000 rpm for 10 min and the supernatants were incubated with 100 mM NaOH for 1 hr at 37°C.
GspB736flag glycoforms from both Δgly-nss and Δgly-nss Asp2S362A backgrounds were overproduced and then purified by affinity chromatography on sWGA-agarose as described above. To produce glycopeptides suitable for LC/MS analysis, both GspB glycoforms (10 μg) were prepared by in-gel digestion using 10% (v/v) acrylamide gels. Following SDS-PAGE and staining with Coomassie Brilliant Blue, bands corresponding to glycosylated GspB were extracted using a clean scalpel and destained with 50% (v/v) acetonitrile (ACN)/25mM ammonium bicarbonate, pH 7.5 and placed into microfuge tubes. The destained gel slices were then dehydrated with ACN and dried in vacuo using a centrifugal evaporator. The slices were resuspended in 50 mM ammonium bicarbonate, pH 7.5 and treated with both sequencing grade trypsin/Lys-C protease mix (20 μg; Promega, Madison, WI) and endoproteinase Glu-C (20 μg; Thermo-Pierce, Waltham, MA). Following digestion for 18 h at 37°C, peptides were recovered from the slices by sonication for 10 min in a water bath sonicator followed by vortexing for 5 min. The acrylamide gel was pelleted by centrifugation (1000 x g, 1 min), and the recovered supernatants were dried in a centrifugal evaporator.
The recovered glycopeptide/peptide mixtures were dissolved in 15 μL water and analyzed by LC/MS using an Agilent 1200 LC coupled to an Agilent UHD 6530 Q-TOF mass spectrometer (Agilent, Santa Clara, CA). Samples were loaded onto the C18 column (Agilent AdvanceBio Peptide Map; 100 mm × 2.1 mm; 2.7 μm) previously equilibrated with solvent A (water, 2% ACN, 0.1% formic acid) and they were resolved using a linear step-gradient of 0–45% B (ACN, 0.1% formic acid) over 40 min, then 45–55% B over 10 min, followed by a wash step at 95% B at a flow rate of 0.2 mLmin-1. The Q-TOF was operated in extended dynamic range positive-ion auto MS/MS modes with an m/z range of 300–2000 m/z and a capillary voltage of 4 kV. Three precursor ions were chosen for collision induced dissociation (CID) fragmentation. To identify GspB, the MS data were analyzed using PEAKS 7 software (Bioinformatics Solutions Inc., Waterloo, ON). Glycopeptides and their O-acetylated forms were identified with MassHunter Qualitative Analysis software (Agilent, Santa Clara, CA) by manually searching the MS/MS data for the presence of diagnostic HexNAc and O-acetylHexNAc oxonium fragmentation product ions (m/z = 204.08 and 246.09, respectively).
We have previously shown that a GST-GspB fusion encompassing the entire SRR1 region of GspB (GST-SRR1) can be glycosylated by all the glycosyltransferases within the GspB operon when reconstituted in E. coli [20]. Moreover, this GST-GspB fusion proved to be highly soluble and easily purified from E. coli, and was therefore considered representative of the GspB glycan and suitable for subsequent glycan profile analysis. Glycosylated GST-SRR1 proteins were GST purified as described below and subjected to glycan profiling through MALDI-TOF Mass spectrometry profiling, performed as a service by the Glycotechnology Core Resource at the University of California, San Diego. In brief, glycans were released from GST-SRR1 by reductive beta elimination using base-borohydride treatment. The O-glycans were then purified and per-methylated and dissolved in MeOH. Dissolved permethylated glycans were mixed with super-DHB matrix in a 1:1 (v/v) ratio and spotted on a MALDI plate and MALDI-TOF MS analysis was performed in positive ion mode. The proposed structures for mass peaks were extracted from the CFG database using GlycoWorkbench software.
For inducible expression in E. coli, asp2 and its catalytic mutant (asp2S362A) were cloned into the expression vector pMAL-c2x (NEB) resulting in an in-frame fusion of MalE at the N-terminal of Asp2 and a His6 -tag (H6) at the C-terminus. The pMal-asp2-H6 constructs were introduced to E. coli BL21 (Lucigene) and grown in Low Salt LB medium containing 2% glucose at 37°C supplemented with ampicillin. Upon reaching an OD600 of 0.6, the cells were induced with 1 mM IPTG and allowed to grow for an additional 18 h at 17°C. The cells were resuspended in Lysis buffer (20 mM Tris pH 8.5, 200 mM NaCl, 1 mM EDTA and 0.5% Triton-X100) supplemented with lysozyme (50 μg mL-1) and lysed by passage through a French press cell (15, 000 psi). MalE-Asp2-H6 fusion proteins were purified from clarified lysates under native conditions by affinity chromatography using amylose resin (NEB). Further purification of MalE-Asp2-H6 was achieved by affinity purification using Ni2+-nitrilotriacetic acid agarose (Qiagen). Semi-purified MalE-Asp2-H6 protein was reconstituted in His6-binding buffer (50mM Sodium phosphate, pH 8, 150 mM NaCl and 10 mM imidazole) and mixed with pre-equilibrated resin under constant rotation at 4°C. Mal-Asp2H6 proteins were eluted in His6-binding buffer containing 300 mM imidazole, concentrated by ultrafiltration using an Amicon Ultra centrifugal filter (100 kDa cutoff) and dialyzed against 50 mM Sodium phosphate, pH 7 containing 150 mM NaCl overnight at 4°C before use.
Construction of a gstSRR1-gtfAB co-expression plasmid, resulting in the expression a GlcNAc-glycosylated GST-SRR1 protein has been described previously [26]. E. coli strains transformed with this construct (PS875) were grown to an OD600 of 0.6 in LB medium at 37°C supplemented with ampicillin and induced as described above. Cells were resuspended in GST-Lysis buffer (50 mM Tris pH 8, 150 mM NaCl and 0.5% Triton X-100) supplemented with lysozyme (50 μg mL-1) and lysed by sonication. GST-SRR1 fusion protein was purified from clarified lysates under native conditions by affinity chromatography using glutathione agarose (Pierce) according to the manufacturer’s instructions. Purified GST-SRR1 protein was eluted in 50 mM Sodium phosphate, pH 7 containing 150 mM NaCl and 30 mM glutathione and concentrated and dialyzed as described above.
The in vitro acetylesterase activity of Asp2 was assessed as described previously [24]. In brief, reaction mixtures contained 2 mM p-nitrophenyl acetate (pNP-Ac) (dissolved in ethanol, 2% final) (Sigma), 10 μg of MalE-Asp2-H6 or its mutant form and 10 μg of GtfAB glycosylated GST-SRR1 in a total volume of 200 μL in 50 mM Sodium phosphate, pH 7. Reactions, performed in triplicate, were initiated by the addition of pNP-Ac, and were monitored continuously at 405 nm for the release of pNP over 30 min at 25°C using a Spectra Max 250 microplate reader (Molecular Devices). The hydrolysis of pNP-Ac by Asp2 results in the release of pNP which was monitored as an increase of absorbance at 405 nm. Spontaneous pNP-Ac hydrolysis was subtracted from all enzyme catalyzed reactions at time zero. A calibration curve for pNP was obtained under the reaction conditions and used to calculate rate of pNP release.
The binding of S. gordonii to immobilized platelets was performed as described previously [9]. In brief, strains were grown for 18 hr, washed twice in Dulbeccos PBS (DPBS), sonicated briefly to disrupt aggregated cells and diluted to approximately 2x107 cfu mL-1. Bacterial suspensions were applied to wells of a microtiter plate coated with human platelets. After a 2 hr incubation at room temperature, the unbound bacteria were removed by aspiration. Wells were washed three times with DPBS and the bound bacteria were released by trypsinization. The number of input and bound bacteria were determined by plating serial dilutions of bacterial suspensions on sheep blood agar plates, and the binding was expressed as the percent of the input bound to human platelets. The differences in binding between groups were examined by one-way ANOVA with post-hoc Tukey HSD (honestly significant difference) (http://vassarstats.net/anova1u.html). P < 0.01 was considered statistically significant.
|
10.1371/journal.ppat.1000524 | Target Cell Cyclophilins Facilitate Human Papillomavirus Type 16 Infection | Following attachment to primary receptor heparan sulfate proteoglycans (HSPG), human papillomavirus type 16 (HPV16) particles undergo conformational changes affecting the major and minor capsid proteins, L1 and L2, respectively. This results in exposure of the L2 N-terminus, transfer to uptake receptors, and infectious internalization. Here, we report that target cell cyclophilins, peptidyl-prolyl cis/trans isomerases, are required for efficient HPV16 infection. Cell surface cyclophilin B (CyPB) facilitates conformational changes in capsid proteins, resulting in exposure of the L2 N-terminus. Inhibition of CyPB blocked HPV16 infection by inducing noninfectious internalization. Mutation of a putative CyP binding site present in HPV16 L2 yielded exposed L2 N-terminus in the absence of active CyP and bypassed the need for cell surface CyPB. However, this mutant was still sensitive to CyP inhibition and required CyP for completion of infection, probably after internalization. Taken together, these data suggest that CyP is required during two distinct steps of HPV16 infection. Identification of cell surface CyPB will facilitate the study of the complex events preceding internalization and adds a putative drug target for prevention of HPV–induced diseases.
| Human papillomaviruses (HPV), especially HPV types 16 and 18, are a major cause of cancer in women worldwide. HPV16, like most genital HPV types, relies on heparan sulfate proteoglycans (HSPGs) to attach to host cells and to the extracellular matrix. Attachment is mediated by surface-exposed basic residues of the major capsid protein, L1. This triggers conformational changes affecting L1 and the minor capsid protein, L2. However, it is not known what interaction triggers these structural changes and if any host cell protein is involved. Now we have identified a host cell chaperone, Cyclophilin B (CyPB), as essential for efficient HPV16 and HPV18 infection. CyPB, which is present on the cell surface in association with specific forms of O-sulfated HSPG as well as in the lumen of intracellular membrane structures, is an energy-independent enzyme, which catalyzes cis/trans isomerization of peptidyl-prolyl bonds. We demonstrate that CyPB facilitates conformational changes resulting in exposure of the L2 N-terminus, which is required for infectious entry. In addition, we present some evidence suggesting that members of the cyclophilin family are required for a second, probably intracellular, step of HPV16 infection. This is the first report implicating cell surface chaperones as essential host factors for viral infection.
| Cyclophilins (CyP) comprise a family of peptidyl-prolyl cis/trans isomerases, which are evolutionarily conserved and ubiquitously expressed [1],[2]. CyP facilitate folding of nascent proteins and through this have been implicated in RNA splicing, stress responses, gene expression, cell signaling, mitochondrial function, and regulation of kinase activity [3]. The 16 human family members differ mainly by terminal extensions, which are probably responsible for subcellular localization and protein-protein interactions, and by tissue specific expression. CyP were initially identified as high affinity binding proteins for cyclosporin A (CsA), an immunosuppressive agent [4]. CsA blocks the enzymatic acitivity of CyP. Cyclophilin A and B (CyPA and CyPB) are the most abundant among the family, where CyPA mainly localizes to the cytoplasm and CyPB, which encodes a signal peptide, is associated with the endoplasmic reticulum (ER). CyPB can be secreted and is detected on the cell surface, where it colocalizes with heparan sulfate proteoglycans (HSPGs) like syndecan-1 [5]. Recent reports suggest that CyPB preferentially bind HSPG molecules that carry a 3-O-sulfated N-unsubstituted glucosaminoglycan residue in the heparan chain [6]. 3-O-sulfation is the least abundant modification of heparan sulfate and thus only few HSPG molecules on the cell surface are associated with CyPB. The core protein required for triggering biological function of cell surface CyPB is most likely syndecan-1 [5].
Several viruses exploit CyP for life cycle completion. The capsid protein of human immunodeficiency virus type 1 (HIV-1) harbors a CyPA binding site resulting in the incorporation of this chaperone into the virion [7]. In addition, target cell CyPA is required for efficient infection of human cells [8],[9]. Inhibition of CyPA prevents the transport of reverse transcribed viral genome to the nucleus without interfering with reverse transcription [10]. A number of observations were interpreted as CyPA preventing the interaction of the viral capsid protein with restriction factors rather than it promoting viral uncoating. In some nonpermissive cells, CyPA activity is required for binding of the restriction factor TRIM5 to the capsid protein (for review see [7]). Hepatitis C virus (HCV) is another example requiring CyPB activity for efficient replication. It interacts with the viral polymerase NS5B thus promoting RNA binding [11]. Furthermore, mouse cytomegalovirus (MCMV) infection of neural stem/progenitor cells is facilitated by CyPA by an unknown mechanism [12].
Here we demonstrate that CyPB activity facilitates infection of human papillomavirus type 16 (HPV16) and HPV18. HPV are non-enveloped epitheliotropic DNA viruses with a circular, chromatinized, double stranded DNA genome of approximately 8000 bp. They induce benign lesions of the skin and mucosa that in some instances progress to malignancies. HPV induced malignancies, including cervical carcinoma, contribute to more than 7% of all cancers in women worldwide. The viral capsid is composed of 360 copies of the major capsid protein, L1, and up to 72 copies of the minor capsid protein, L2 [13]–[15]. L1 protein, which is organized in 72 pentamers, called capsomeres, mediates the primary attachment of viral particles to the cell surface [16]–[18] and/or extracellular matrix (ECM) of susceptible cells [19], most probably via HSPG [20]. The need for HS can be bypassed by treatment of immature HPV16 pseudovirions with furin convertase [21]. The primary attachment is mediated by surface-exposed lysine residues located at the rim of capsomeres [22]. HPV33 binding to the cell surface requires O-sulfation of HS, whereas both N- and O-sulfation are needed for HSPG to function as an initiator of the infectious entry pathway [23]. These data suggested that secondary HSPG interactions may play a role in infection, which was recently supported by the use of the HS binding drug DSTP-27 [18]. Virus attachment triggers conformational changes in both capsid proteins [23]–[25], which seem to be required for transfer to putative secondary receptors and infectious internalization [18]. Conformational changes result in the exposure of the N-terminus of L2 protein, which contains a highly cross-reactive neutralizing epitope, and subsequent cleavage of 12 N-terminal amino acids catalyzed by furin convertase [24],[26]. Data presented below suggest that cell surface CyPB facilitates exposure of the L2 N-terminus, which is required for infectious internalization.
We used a well established pseudovirus system for our studies, which relies on the expression of codon-modified forms of L1 and L2 in human embryonic kidney 293TT cells harboring a high copy number packaging plasmid [27]. We packaged a green fluorescent protein (GFP)–based marker plasmid that has been successfully used before to study early events of HPV infection [28]–[31]. We observed that CsA efficiently blocked HPV16 infection of 293TT cells with an inhibitory concentration 50 (IC50) of approximately 2 µM (Figure 1A). Similar results were obtained for HaCaT, which is currently the most commonly used keratinocytes-derived cell line for analysis of HPV infection, and HPV-harboring HeLa (Figure 1B). CsA has been shown to block activity of calcineurin and CyP as well as P-glycoproteins, also known as ABC transporters. We used more specific inhibitors to narrow down the cellular target responsible for the observed inhibition. Neither INCA-6 nor the cell permeable R-VIVIT peptide and FK506, inhibitors of the interaction between calcineurin and nuclear factor of activated T cells (NFAT), blocked infection (Figure 1C). Similarly, Virapamil and nifedipine, specific inhibitors of P-glycoproteins, had no effect. In contrast, NIM811, which blocks both P-glycoproteins and CyP, inhibited HPV16 infection as efficiently as CsA. Identical results were obtained for key inhibitor NIM811 in HaCaT cells (data not shown). All inhibitors reduced cell growth of 293TT (Figure 1D), HaCaT and HeLa (not shown) cells to a similar extent. Cell growth inhibition of these inhibitors is well established. Taken together, these results strongly suggest that CyP facilitate HPV16 and HPV18 infection.
In order to determine, which CyP family member may be involved and to confirm our findings, we used an siRNA approach to knock down individual CyP. First, we used an siRNA, si-CyP[broad], which has been shown to target several members of the CyP family including CyPA, CyPB, CyPE, and CyPH [11]. 293TT cells were transfected with si-CyP[broad] 48 prior to infection with HPV16. Western blot confirmed the significant reduction of steady state CyPA and CyPB protein levels (Figure 2B) and infection was reduced to 11% (p<0.01) compared to cells transfected with a control siRNA (Figure 2A). Individual knock down of CyPA and CyPB with specific validated siRNAs [11] also reduced infection to 59% (p<0.05) and 35% (p<0.01), respectively. Specificity of the siRNA knock down for their target was confirmed by Western blot (Figure 2B). The data indicated that both CyP may play a role in HPV16 infection. Compared to CyPA, knockdown of CyPB consistently resulted in stronger inhibition (p<0.05). Similar results were obtained for HaCaT cells. However, due to reduced transfection efficiency of HaCaT cells (70% vs. 95% for 293TT) the inhibitory effect was not as pronounced (Figure 2C and 2D).
To identify the stage at which infection is blocked by CyP inhibitors, we first measured internalization using immunofluorescence (IF). It was shown by several groups that most surface-exposed conformational epitopes that are recognized by neutralizing monoclonal antibodies (NmAb) are destroyed following entry and that L1 protein segregates from the L2/DNA complex in acidic endocytic compartments [18],[29],[30]. During this process reactivity of antibodies specific for hidden linear L1 epitopes is gained [32]. We used NmAb H16.56E to determine if conformational epitopes are lost in the presence of NIM811. H16.56E binding site includes but is not restricted to the N-terminal portion of the FG loop (HPV16 L1 residues 260–270) [33]. We also used mAb 33L1-7, which binds a linear epitope (residues 303–313) that is neither accessible in capsomeres nor in intact particles [34],[35] and recognizes L1 protein late in HPV entry [32]. In untreated cells at 18 h post infection (hpi) with HPV16 pseudovirus, H16.56E reactivity was hardly detectable but perinuclear 33L1-7 staining was obvious indicative of particle internalization and accessibility of the 33L1-7 epitope (Figure 3A). In contrast, we observed a strongly increased perinuclear signal with H16.56E when infection was performed in the presence of 10 µM NIM811. The signal for 33L1-7 was greatly diminished under these conditions (Figure 3A). Similar results were obtained when NIM811 was replaced by CsA (data not shown). We will use the term ‘stabilized capsid phenotype’ to describe the increased reactivity of internalized pseudovirions with H16.56E. These data demonstrate that, first, viral particles are indeed internalized in the presence of CyP inhibitor and, second, the conformational L1 epitope recognized by H16.56E is stabilized.
We again used siRNA knock down to identify the CyP family member responsible for the stabilized capsid phenotype. For this, HaCaT cells were transfected with unspecific control siRNA, si-CyP[broad], si-CyPA or si-CyPB 48 h prior to infection with HPV16 pseudovirus. Successful down regulation of CyPB was confirmed by IF (Figure 3B). Down regulation of CyPA could not be determined by IF because of lack of CyPA-specific antibody reactivity in this assay. However, successful transfection was monitored using FITC-labeled siRNA (not shown) and knock down of CyPA was confirmed by Western blot. Cells with reduced levels of CyPB following transfection with si-CyPB or si-CyP[broad] displayed a stabilized capsid phenotype at 18 hpi, whereas adjacent cells, which were not transfected as indicated by strong staining for CyPB, showed much less reactivity with H16.56E (Figure 3B). A stabilized capsid phenotype was not detected in cells transfected with si-CyPA, even though basal level of reactivity with H16.56E is evident. Taken together these data suggest that blockage of CyPB activity may be responsible for the stabilized capsid phenotype.
Previously we observed a stabilized capsid phenotype when transfer to secondary receptors on the cell surface was blocked by antibodies or drugs [18]. Furthermore, CyPB is found on the cell surface where it is associated with HSPG [6],[36]. Therefore, we hypothesized that CyPB may facilitate the conformational shifts reported for both capsid proteins upon interaction with cell surface HSPG [23],[25]. Currently, the only reliable test for these changes measures the exposure of the L2 N-terminus using the L2-specific NmAb RG-1. RG-1 binds to a peptide encompassing HPV16 L2 residues 17 to 36 [37]. RG-1 reactivity with L2 protein incorporated into virions requires cell attachment-induced exposure of the L2 N-terminus and furin cleavage [24]. To test the role of CyP in conformational shifts, HPV16 pseudovirus was bound to HaCaT cells for 2 h at 4°C and was chased for 4 h at 37°C prior to cell surface staining with RG-1 (a kind gift of R.B. Roden, John Hopkins University) and K75 polyclonal VLP antisera. In control infection we found strong RG-1 signal, which perfectly overlapped with cell-associated L1-specific K75 binding (Figure 4A). RG-1 reactivity was greatly diminished albeit not completely abolished when HaCaT cells were infected in the presence of NIM811 (Figure 4A), whereas reactivity with K75 was not decreased. Similarly, CsA treatment decreased RG-1 signal albeit not as pronounced (data not shown). We quantified the RG-1- and K75-specific signals using software provided by Zeiss and found statistically significant reductions of over 70% and 59% of relative RG-1 signal strength in presence of NIM811 and CsA, respectively (p<0.01) (Figure 4B). Taken together, these data strongly suggest that CyPB activity is required for exposure of the RG-1 epitope on the viral capsid and lend support for a function of cell surface CyPB in HPV16 infection.
Recently, it has been shown that the presence of RG-1 antibody during infection of HaCaT cells with HPV16 pseudovirions prevents infection and virus internalization and relocates viral particles from the cell surface to ECM [24]. We took advantage of this observation to strengthen our findings. We reasoned that, irrespective of presence of RG-1, viral particles should still be internalized and display a stabilized capsid phenotype in presence of NIM811, if the RG-1 epitope is indeed not accessible to antibody binding after drug treatment. To test this, HPV16 pseudovirus was bound to HaCaT cells for 2 h at 4°C in the presence or absence of this drug. After washout of unbound virus, cells were incubated overnight in presence of NIM811 and RG-1. Confirming previous findings [24], RG-1 treatment alone induced deposition of the majority of viral particles to ECM in the absence of NIM811, as evidenced by colocalization of capsid-specific H16.56E signal with the ECM marker Laminin 5 (Figure 5A). We also confirmed the neutralizing capacity of RG-1 using 293TT cells (Figure 5C) to ascertain that this antibody is functional in our hands. Inhibition of HPV16 pseudovirus infection by this antibody using HaCaT cells was previously demonstrated by others [24],[37]. However, the presence of RG-1 antibody in addition to drugs did not prevent internalization of viral capsids, as evidenced by a stabilized capsid phenotype (Figure 5B) and did not result in increased deposition of viral particles on ECM (not shown). It should be noted that RG-1 treatment in the absence of NIM811 displayed a weak but reproducible stabilized capsid phenotype (Figure 5B) suggesting that not all particles are displaced from the cell surface and are instead internalized in a noninfectious manner. These data further support our notion that CyPB action on the cell surface is required for the conformational change resulting in exposure of the RG-1 epitope, which is a prerequisite for infectious internalization.
Not much information is available regarding CyPB substrate binding sites. However, CyPA binding to the HIV capsid protein has been mapped to 85-PXXXGPXXP-93, which is located between Helix 4 and 5 [7]. We found similar sequence elements at the N-terminus of L2 conserved among many but not all members of the Papillomaviridae family (Figure 6A). We exchanged glycine and proline residues of L2 at positions 99 and 100 within the putative CyP binding site for alanine to test their importance for HPV16 infection. We hypothesized that this mutant is either defective for infection due to loss of CyP binding or does not require active CyP for exposure of the L2 N-terminus due to higher flexibility in this L2 region induced by amino acid exchanges. We found that 16L2-G99A-P100A (16L2-GP-N) is incorporated into particles similar to wt L2 (not shown). Mutant pseudovirus retains full infectivity in 293TT (Figure 6B) and HaCaT cells (data not shown), which is consistently and statistically significantly increased compared to wt (p<0.01). When we bound 16L2-GP-N mutant pseudovirus to HaCaT cells and surface-stained with RG-1 and K75 after a 4 h chase at 37°C, we observed similar reactivity of RG-1 with cell-bound pseudovirions in absence or presence of NIM811 (Figure 6C). Quantitative analysis of signal strength confirmed that reactivity of RG-1 with mutant pseudovirus is not significantly reduced by this drug (Figure 6D) in contrast to wt pseudovirus (Figure 4). These data suggested that 16L2-GP-N mutant pseudovirus does not require CyP activity for exposure of the RG-1 epitope. Nevertheless, infection was still sensitive to CsA (Figure 7A) and siRNA knock down of CyP (Figure 7B). However, unlike wt pseudovirus mutant pseudovirus did not produce the stabilized capsid phenotype after treatment with drugs (Figure 7C) or siRNA knock down of CyP (not shown), although H16.56E was still able to detect mutant viral particles on the cell surface and on ECM (data not shown). Taken together, these data indicate not only that 16L2-GP-N mutant pseudovirus bypasses the requirement for cell surface CyPB but also that HPV16 infection requires CyP at a second, possibly intracellular, stage of entry and transport. Furthermore, they strongly support our previous notion that, in presence of CyP inhibitors, wt virus is shunted into a noninfectious entry pathway.
To determine whether the requirement for CyP is a conserved feature among papillomaviruses we tested a number of low and highrisk HPV types as well as BPV-1 for sensitivity to CsA. We found that HPV6, HPV45 and HPV58 were inhibited by CsA similar to HPV16 and HPV18, whereas BPV1, HPV5, HPV31, and HPV52 were relatively resistant to CsA (Table 1). These data suggest that different papillomavirus types have different requirements for CyP, which may be reflective of the entry strategies these viruses evolved.
Here, we report that CyP facilitate infection of the oncogenic HPV16 and 18 among other HPV types. Focusing on HPV16 and using specific drugs, siRNA knock down and mutant pseudovirus we provide evidence that CyP are required at two different stages following primary attachment to host cells. In addition, siRNA knock down data point to the involvement of two members of the CyP family in the infection process: CyPA and CyPB. Combined knock down using siCyP[broad] affected infection more severely than individual knock downs suggesting they may facilitate different steps of HPV16 infection. Our data indicate that CyPB is functioning on the cell surface. However, we were not yet able to identify the step requiring CyPA. Also, at the moment we cannot completely rule out the involvement of additional CyP family members, like CyPE and CyPH, whose expression should also be affected by siCyP[broad].
We provide evidence that cell surface CyPB is essential for triggering events that lead to infectious internalization of viral particles probably by catalyzing conformational changes of viral capsid proteins. It is well established that both HPV16 capsid proteins undergo conformational changes on the cell surface prior to internalization. Conformational changes induced in L1 are not well defined but seem to involve the BC loop [23]. Conformational changes induced in L2 protein result in exposure of some forty N-terminal amino acids, which allows furin convertase-mediated cleavage of L2 and binding of the L2-specific NmAb RG-1 [24],[25],[38]. CyP inhibition greatly reduced exposure of the RG-1 epitope following cell attachment as measured directly by IF and indirectly by determining the fate of cell bound pseudovirus in the presence of CyP inhibitors and RG-1. This strongly indicates that CyP activity is required to make the RG-1 epitope accessible to antibody binding. However, the block was not complete and residual reactivity with RG-1 was observed in presence of inhibitors, which could possibly be attributed to the presence of activated particles in the pseudovirus preparation [18],[28] and/or to the baseline spontaneous conformational change in absence of CyP activity due to receptor engagement. Nevertheless, the reduction in RG-1 reactivity by CyP-specific drugs was found to be correlated with reduction of infectivity by drugs.
We also provide evidence that L2 protein may be the substrate for CyP. First, we were able to bypass the requirement for cell surface CyP by introducing amino acid changes in a putative CyP binding site within L2, which is accessible in mature virions [39]. Mutant pseudovirus did not require CyP activity for exposure of L2 as demonstrated by IF. However, at this point we cannot completely rule out that CyP functions rather indirectly by (i) modifying cell surface receptors, since CyP have been shown to isomerize prolyl peptide bonds of cell surface markers, thus modifying their biological function [40],[41], and by (ii) regulating cell trafficking and cell surface expression of proteins [42]. However, this is rather unlikely since BPV1, which uses the same route of internalization as HPV16 [43],[44], is not blocked by CyP inhibitors. Second, specific blockage of CyPB induced noninfectious internalization with the hallmark of a stabilized capsid phenotype. In this respect CyPB inhibition is similar to post-attachment treatment with the BC loop-specific antibody H33.J3, heparinase, or the HS binding drug DSTP-27, which also induce noninfectious internalization and stabilization of viral capsids [18]. It was suggested that these treatments all block secondary receptor interactions, which seems to require an exposed L2 N-terminus [18]. It is unlikely that L1 rather than L2 protein is the substrate of CyP. This is based on our unpublished observations that CsA, NIM811 or CyP-specific siRNAs do not block L1 conformational changes occurring on the cell surface.
Interestingly, mutant 16L2-GP-N pseudovirus remained sensitive to CyP inhibitory drugs and siRNA knock down. However, bypassing the need for cell surface CyPB using mutant pseudovirus yielded an inhibition phenotype distinct from wt particles. We no longer observed capsid stabilization. This suggests that CyP activity is required at a subsequent step during internalization and/or intracellular transport. So far, we were not able to identify the exact step(s) that require CyP activity and therefore cannot predict which specific CyP family member may be involved. However, a second putative CyP binding site is located near the C-terminus of L2 (409-PLVSGPDIP-417). This sequence is close to a region that has been shown to mediate interaction of L2 with L1 capsomeres in HPV11 [45]. It is therefore tempting to speculate that endocytic CyP mediates segregation of L2 from L1. As the C-terminal section of L2 is required for membrane destabilization and passage of membranes [30] as well as for interaction with dynein [46], this could free the C-terminus allowing association and penetration of surrounding membranes and consequently the L2/DNA complex to egress from endosomes and retrograde transport towards the nucleus. CyPB encodes a signal peptide and is therefore found in the luminal compartment of intracellular membranes making it a likely candidate. However, CyPA is also secreted into the extracellular space, even though it lacks a signal peptide, suggesting it finds its way into the luminal compartment of at least the secretory pathway.
Host cell CyP do not facilitate infection of all papillomaviruses. Support for this notion came from our finding that HPV5, HPV31, HPV52, and BPV1 are rather resistant to CyP-specific drugs. This may reflect the evolution of different internalization strategies. For example, HPV31 is internalized via caveolae-dependent endocytosis [31],[47] whereas HPV16 uses a caveolae- and clathrin-independent pathway [32]. BPV1 L2 does not harbor putative CyP binding sites and replacing key proline residues by more flexible amino acids may make a catalytic activity dispensable for L2 exposure, as we have shown with HPV16L2-GP-N. The entry pathways of HPV5 and HPV52 have not been investigated yet. However, the L2 protein of both HPV types has a putative N-terminal CyP binding site.
Attachment-induced conformational changes are a common theme in virus infection. They are usually triggered by interaction with specific receptors, which allows interaction with secondary receptors or, more often, trigger cell fusion events. Although chaperones present in endocytic vesicles or the endoplasmic reticulum have been shown to facilitate virus uncoating and translocation across membranes [48],[49], this is the first report to implicate chaperones in mediating conformational changes of capsid proteins on the surface of target cells.
With this report we are adding another virus family to the list of viruses dependent on CyP activity for completion of their life cycle. Despite 15 years of study, the role of CyPA in HIV-1 infection is not yet fully defined (for review see [7],[50]). Similarly, its involvement in MCMV infection of neural progenitor cells has not been characterized in molecular detail [12], whereas it was convincingly demonstrated for HCV that ER-resident CyPB enhances the RNA binding activity of the NS5B RNA polymerase and consequently genome amplification [11]. With the identification of CyPB as modifier of oncogenic HPV capsid protein conformation, which activates the virus for entry via an infectious pathway, for the first time we have characterized its role at the molecular level during cell surface events of viral infections. This should allow characterizing the complex events preceding internalization in more detail and adds a putative drug target for prevention of HPV-induced diseases, especially since CsA has been approved for and is already being used in clinical settings.
293TT cells and expression plasmids for codon-optimized structural genes coding for HPV5, HPV6, HPV18, HPV31, HPV45, HPV52, HPV58 as well as BPV1 were kindly provided by John Schiller and Chris Buck, Bethesda [27],[51]. Codon-optimized HPV16 L1 and L2 expression plasmids were a kind gift from Martin Müller, Heidelberg [52]. HPV16L1-specific rabbit polyclonal antisera K75, mouse monoclonal antibody H16.56E and 33L1-7 have been described previously [34],[35]. Anti-CyPA polyconal rabbit antibody was obtained from Dharmacon (cat #: 07-313). CyPB polyclonal rabbit antibody was purchased from Affinity BioReagents Inc (Golden, Colorado; cat #: PA1-027). However, we noticed that only lot number 328-120 and prior lots were reactive in IF. All subsequent lots tested were not reactive in IF analyses. Laminin 5 rabbit polyclonal antibody was from Abcam (cat #: ab14509). AF488-labeled GFP-specific rabbit polyclonal antibody was obtained from Invitrogen. Mouse monoclonal L2-specific RG-1 antibody was kindly provided by Richard Roden, John Hopkins University, Baltimore. AlexaFluor (AF)–labeled secondary antibodies and phalloidin were purchased from Invitrogen. Pseudovirions were generated and purified using Optiprep gradient centrifugation following published procedures [27]. Pseudovirus yield was determined by green fluorescent protein (GFP)–specific quantitative real time polymerase chain reaction (qRT–PCR).
Cyclosporin A was obtained from Toronto Research Chemicals (cat #: C988900). NIM811 was a kind gift from Novartis. Verapamil (cat #: 676777), Nifedipine (cat #: 481981), 11R-VIVIT (cat #: 480401) and INCA-6 (cat #: 480403) were obtained from Calbiochem. FK 506 (cat #: F1030) was purchased from A.G. Scientific (San Diego). The cell viability and proliferation assay ‘CellTiter96 Aqueous One Solution’ was purchased from Promega (Madison, WI). This assay measures the quantity of formazan product, which is directly proportional to the number of living cells.
293TT cells were seeded a day before and allowed to attach. Next day, drugs were serially diluted in complete DMEM in 24 well-plates and adequate amounts of pseudoviruses were added to achieve infection levels of 10 to 30%. Infectivity was scored by counting GFP expressing cells at 72 hpi using flow cytometry. Similar protocol was followed for infection assay using HaCaT and HeLa cells except that cells were fixed with 2% paraformaldehyde, permeabilized with 0.2% Triton X-100 in phosphate buffered saline (PBS), stained with AF488-labeled GFP-specific antibody and counted using a Leica DMBI 6000 fluorescence microscope. Unless otherwise stated standard deviation was based on at least five replicates from at least two independent experiments.
RNA interference was carried out using synthetic siRNA duplexes with symmetric 3′-deoxythymidine overhangs. siRNA duplexes si-CyPA, 5′-AAGCATA CGGGTCCTGGCATC-3′; si-CyPB, 5′-AAGGTGGAGAGCACCAAGACA-3′; and si-CyP(broad), 5′-AAGCATGTGGTGTTTGGCAAA-3′), which have been described and validated before [11], were purchased from Integrated DNA Technologies Inc. Non-specific siRNA, si-NS, 5′-AAGTCCGTGCCGTCAGTTCTCAGAA-3′ was obtained from Invitrogen. Cells were transfected with 3 µg of siRNA duplexes in serum-free medium using MATra reagent (IBA biotagnology, Goettingen; cat. #: 7-2001-100) according to manufacturer's protocol. Typical siRNA transfection efficiency was found to be 70% for HaCaT and 95% for 293TT cells as monitored by fluorescein-labeled control siRNA duplex. CyP knockdown was confirmed 48 h post siRNA transfection (hpTx) by Western blot.
HaCaT and 293TT cells were transfected with siRNA as mentioned above. 48 hpTx, HaCaT were harvested with trypsin and reseeded onto cover slip for immunofluorescence study. Few hours later, when cells had attached, they were infected. At 18 hpi samples were fixed with 4% paraformaldehyde and stained. Alternatively, cells were incubated for 72 h and subsequently stained for GFP as described above to score infection [18]. For infection assay using 293TT, cells were harvested, reseeded into 96 well plates and allowed to attach. Few hours later, they were infected and scored at 72 hpi by counting GFP positive cells.
HaCaT cells were grown on cover slips till ∼50% confluency and infected with HPV16 pseudovirus in presence of NIM811, antibody, or DMSO. 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 0.2% Triton X-100 in PBS for 2 min, washed, and blocked with 5% goat serum in PBS for 30 min, followed by a 1 h incubation with primary antibodies at 37°C. After extensive washing, cells were incubated with AlexaFluor-tagged secondary antibodies and fluorescently labeled phalloidin for 1 h. After extensive washing with PBS, cells were mounted in ‘Gold Antifade’ containing Dapi (Invitrogen). Images were captured by confocal microscopy (Zeiss 510 Laser Scanning Confocal Microscope operated by LaserSharp2000 software) or by standard fluorescence microscopy (Leica DMBI 6000 microscope). Within individual experiments the same microscope settings and exposure times were used. For quantification of fluorescent signal intensity, the LSM server software provided with the confocal microscope was used. Signal strength was acquired from randomly selected single cells (n>15 for each group). The average region of interest was not significantly different among all groups. Background was determined using mock infected cells and subtracted prior to calculations.
RG-1 staining was performed as described [24]. In brief, infected HaCaT cells were shifted to 4°C and incubated with RG-1 and K75 for 1 h in presence of 2% normal goat serum. After extensive washing and incubation with fluorescently labeled secondary antisera, cells were fixed for 20 min in 2% paraformaldehyde. After washing, cells were incubated for 5 min with phalloidin-AF647 conjugate and mounted.
CyPA: NM_021130; CyPB: NM_000942; codon optimized HPV16 L1: AJ313179; codon optimized HPV16 L2: AJ313180
|
10.1371/journal.pcbi.0030243 | Predicting Gene Expression from Sequence: A Reexamination | Although much of the information regarding genes' expressions is encoded in the genome, deciphering such information has been very challenging. We reexamined Beer and Tavazoie's (BT) approach to predict mRNA expression patterns of 2,587 genes in Saccharomyces cerevisiae from the information in their respective promoter sequences. Instead of fitting complex Bayesian network models, we trained naïve Bayes classifiers using only the sequence-motif matching scores provided by BT. Our simple models correctly predict expression patterns for 79% of the genes, based on the same criterion and the same cross-validation (CV) procedure as BT, which compares favorably to the 73% accuracy of BT. The fact that our approach did not use position and orientation information of the predicted binding sites but achieved a higher prediction accuracy, motivated us to investigate a few biological predictions made by BT. We found that some of their predictions, especially those related to motif orientations and positions, are at best circumstantial. For example, the combinatorial rules suggested by BT for the PAC and RRPE motifs are not unique to the cluster of genes from which the predictive model was inferred, and there are simpler rules that are statistically more significant than BT's ones. We also show that CV procedure used by BT to estimate their method's prediction accuracy is inappropriate and may have overestimated the prediction accuracy by about 10%.
| Through binding to certain sequence-specific sites upstream of the target genes, a special class of proteins called transcription factors (TFs) control transcription activities, i.e., expression amounts, of the downstream genes. The DNA sequence patterns bound by TFs are called motifs. It has been shown in an article by Beer and Tavazoie (BT) published in Cell in 2004 that a gene's expression pattern can be well-predicted based only on its upstream sequence information in the form of matching scores of a set of sequence motifs and the location and orientation of corresponding predicted binding sites. Here we report a new naïve Bayes method for such a prediction task. Compared to BT's work, our model is simpler, more robust, and achieves a higher prediction accuracy using only the motif matching score. In our method, the location and orientation information do not further help the prediction in a global way. Our result also casts doubt on several biological hypotheses generated by BT based on their model. Finally, we show that the cross-validation procedure used by BT to estimate their method's prediction accuracy is inappropriate and may have overestimated the accuracy by about 10%.
| Developing computational strategies for predicting transcription factor binding sites (TFBSs) and transcription regulatory networks has been a central problem in computational biology for more than a decade. Reviews on this problem and various proposed methods can be found in [1–3]. A popular strategy is to search from upstream sequences of a set of co-regulated genes for over-represented (i.e., enriched) sequence features (motifs) [4–7]. With the help of gene expression microarray technology, the expression level of thousands of genes can be measured at the same time [8–10], which makes the discovery of sets of co-regulated genes and their respective regulatory signals at the genome-wide level a reality for many species.
Bussemaker et al. [11] pioneered the use of regression models to relate a gene's expression with numbers of occurrences of certain k-mer “words” in the upstream sequence of this gene. Motivated by their work, researchers have developed various methods to extract features that are predictive of gene expression levels. Keles et al. [12,13] tackled the problem using logic regression, which treats motif occurrences as binary covariates and selects important predictors adaptively. Conlon et al. [14] proposed a stepwise regression procedure called Motif Regressor, which uses motif matching scores at promoter regions instead of k-mer occurrences as covariates. Zhong et al. [15] extended these methods by introducing a more flexible regression model with an unspecified nonlinear link function. Das et al. [16] implemented a smoothing-spline regression in the place of the linear regression used by Motif Regressor. Further along this general direction, Segal et al. [17] showed that DNA sequence and gene expression information can be combined to construct transcriptional modules. Lee et al. [18] used the ChIP-chip technology and genome-wide location analysis to infer transcriptional regulatory networks in S. cerevisiae.
Beer and Tavazoie (BT) [19] proposed a novel formulation of the sequence–expression problem. They asked the very intriguing, but seemingly impossible, question: how much can we predict gene expressions from gene upstream sequences? To address the question, they first clustered a large portion of genes in S. cerevisiae into 49 tight co-expression groups, found enriched sequence patterns (motifs) among the promoter sequences of genes in each group using de novo motif prediction tools [6,20], and then trained a set of Bayesian network models to predict the group membership of each gene using the matching scores of its promoter sequence to the set of sequence motifs as well as the orientation and position of the predicted binding sites. They conducted a 5-fold cross-validation (CV) procedure to estimate their model's prediction power and found its prediction accuracy to be as high as 73%. A great benefit of the Bayesian network, as shown by BT, is its ability to learn “combinatorial codes” for gene regulation. Hvidsten et al. [21] have applied a similar approach to infer “IF–THEN” rules for transcription regulation. While Bussemaker et al. [11] and Conlon et al. [14] aimed at using gene expression information to help discover transcription factor binding motifs (TFBMs) and binding sites, BT focused directly on the prediction problem.
However, a few key questions remain. First, BT's assessment of their method's prediction power is over-optimistic, as their CV procedure did not include the motif-finding step (more details later). But, how much can we really predict? Second, is the Bayesian network an appropriate model for the task or just too complex a black box, prone to overfitting for the stated tasks? Third, do those inferred combinatorial rules have real predictive power, or are they only observational oddities after the model fitting? How should we think about and quantify uncertainties inherent in such inferred models? Given the limited amount of data and the vast number of potential predictors (e.g., 666 sequence motifs, orientations, and positions of candidate motif sites, etc.), it is not clear if a complex-structured model can be fitted with any confidence.
Our plan to address the above concerns is as follows. We first use the same data and the same (but wrong) CV procedure as in [19] to develop our predictive models, naïve Bayes classifiers with feature preselections, so as to study the problem of model fitting. Then, we study contributions of various sequence features, such as orientations and positions of the predicted binding sites, to the prediction accuracy. Lastly, we implement a correct CV procedure and show the difference of prediction accuracies resulting from correct versus incorrect CV procedures.
Based on the same gene clustering information, putative TF binding motifs, and gene upstream sequences as in [19], our naïve Bayes classifiers outperformed BT's Bayesian network without using any information regarding the position and orientation of the predicted TFBSs. Our classifiers typically select more motif features, but have far fewer model parameters than the Bayesian network models in [19]. We also found that adding the information regarding TFBS orientation and position cannot further improve the naïve Bayes classifier's predictive power in a global way, which casts doubts on several biological predictions made in [19] regarding combinatorial rules of gene regulation. We further studied a few cases in detail and found that the supports for the inferred combinatorial rules are at best circumstantial. Finally, we speculate that the incorrect CV procedure used in [19] has likely overestimated the accuracy rate of their method by 10%.
The data used in this study were obtained from the supplemental Web site of [19], which contains matching scores (i.e., the likelihood of a promoter sequence to contain good sequence matches to a candidate TFBM), and orientations and positions of the predicted matches of 666 putative TFBMs for 2,587 genes in S. cerevisiae. In [19], these 2,587 genes were clustered into 49 different co-expression groups according to their expression profiles in 255 conditions, such as environmental stress [22] and cell cycle [8]. We trained a set of naïve Bayes classifiers to predict the cluster label (membership) for each gene using only its motif matching scores. Since genes in the same cluster have very similar expression profiles, a gene's cluster membership can serve as a surrogate of its expression behavior under different conditions.
We built one naïve Bayes model for each cluster, resulting in a total of 49 classifiers. For each cluster, we first ranked all the 666 sequence motifs according to a Chi-square test procedure, which reflects these motifs' capability of differentiating genes in this cluster from all other genes. Then, we selected the top m most significant motifs as explanatory variables to train a naïve Bayes classifier (for this cluster), where m can range from 1 to 666. We used the same 5-fold CV procedure as that in BT to test the predictive power of our models. As shown in Figure 1, using the same criteria for classification accuracy as in [19] (i.e., for any pair of clusters, if the correlation between their mean expression is greater than 0.65, then misclassifying genes in one cluster into the other is not counted as errors), naïve Bayes classifiers correctly predicted expression patterns for 75% of the genes when the number of preselected motifs m is 5. When m is increased to 20, naïve Bayes classifiers achieved a 79% prediction accuracy (see Table S1). In addition, the naïve Bayes models contain almost all the motif features selected by BT in [19] and include many more (see Figures S1 and S2). It can also be seen that, although the training accuracy always increases as m increases, the prediction accuracy starts to plateau and then decrease as m exceeds 20, which is indicative of overfitting as more variables are included. Following BT, we also calculated the mean correlation of each gene to its predicted expression pattern. For a gene, its predicted expression pattern is the mean expression pattern of the cluster that it is predicted to belong to. With our 20-motif naïve Bayes model, we obtained a mean correlation of 0.56 without using any position and orientation information, which is also higher than BT's result of 0.51.
Having fitted the classification models, we now study how the 666 motifs are present in the model of each cluster. Our first observation is that most clusters have their distinct sets of motif features. But a few motifs are selected by multiple clusters, which may indicate that either the transcription factors corresponding to these motifs are somewhat multi-taskers, or the clusters that share these common motifs are closely related. For example, Motifs PAC and RRPE are selected in the models for clusters 4, 10, 17, 26, and 29. This suggests that many genes in these five clusters may be targeted by the TFs that bind to PAC and RRPE. Clusters 47 and 48 share 17 out of 20 motifs in their models (p < 1 × 10−21). Coupled with the fact that the correlation of the mean expression patterns of these two clusters is more than 0.8, it strongly suggests that genes in these two clusters are co-regulated.
Motif PAC is associated with polymerase A and C subunits [20,23]. Motif RRPE specifically exists in genes involved in rRNA processing [20]. BT extracted from their model a combinatorial prediction rule for cluster 4 [19]: PAC should have a score higher than 0.6 and be within 140 bp of ATG; RRPE should have a score higher than 0.65 and be within 240 bp of ATG. Table 1 shows numbers of genes in a few different clusters that satisfy these constraints. The statistics suggest that PAC and RRPE are both significantly enriched in cluster 4, but not uniquely. Clusters 10, 17, 26, and 29 also have significant portions of genes that satisfy the constraints of both motifs. Our naïve Bayes method successfully picked PAC and RRPE for all these five clusters, whereas BT did not select RRPE for cluster 10, or PAC for cluster 29. It suggests that, due to its complex nature, the Bayesian network model in [19] can easily miss important features. Furthermore, our method using no information about TFBS orientation and position correctly predicted 94% of the genes in cluster 4 and 87% of the genes in clusters 10, 17, 26, and 29, which is comparable to the 92% and 87% accuracy of [19] for the same clusters.
RAP1 is a main regulator of ribosomal proteins in S. cerevisiae, and many ribosomal protein coding genes are reported to have RAP1 binding site(s) in their upstream sequences [24]. BT [19] found that cluster 1 is enriched with RAP1 binding sites, and their Bayesian network inferred a rule for genes in this cluster: their RAP1 score on upstream sequences has to be greater than 0.6, and their RAP1 sites have to be oriented toward a certain direction. We examined this rule carefully and observed the following. First, we found that 82 genes in cluster 1 (a total of 124 genes) and 165 genes in other clusters (a total of 2,463 genes) have putative RAP1 binding sites (i.e., with RAP1 matching score >0.6), which gives rise to a p-value of 1 × 10−59 (based on Fisher's exact test) for the enrichment of RAP1 sites in cluster 1. Seventy-three genes in cluster 1 and only 85 genes in other clusters satisfy both the orientation and the site score requirements, which yields an even more significant contrast p-value, 1 × 10−64. It seems that the RAP1 orientation can indeed help enhance the prediction specificity, although only slightly.
However, our naïve Bayes model selected motif M198 as its main predictor for genes in cluster 1. This motif has a very similar weight matrix to that of RAP1 but includes an extra position (Figure 2). By setting 0.6 as the score threshold of M198, we found that 100 genes in cluster 1 and 126 genes in other clusters contain the M198 site, which gives us a p-value of 4 × 10−94 for the M198 enrichment in cluster 1. Thus, if judged by statistical significance of the prediction specificity, the naïve Bayes model with one simple predictor easily outperformed the more complex combinatorial rule inferred by BT's Bayesian network.
In order to evaluate the effectiveness of RAP1 (with orientation constraint, denoted as RAP1d for short) and M198 as covariates in our classifier, we compared two procedures. In both procedures, one single best motif was selected for each cluster. The only difference was that, for cluster 1, M198 was used in Procedure One and RAP1d was used in Procedure Two. As a result, Procedure One predicted 20 more genes correctly than Procedure Two, and the improvement is mainly in cluster 1. For cluster 1 alone, Procedure One has a 30% false positive and 18% false negative rates, while Procedure Two has a 38% false positive and a 34% false negative rate. These results further suggest that M198 is a better motif for cluster 1 than the oriented RAP1. In the next subsection, we provide a more thorough investigation on the biological relevancy of motif site orientation and its effect on the classification accuracy.
The result in the previous subsection does not mean that the motif site orientation is not biologically important. In fact, we found that 91 of the 100 predicted M198 sites for genes in cluster 1 are oriented toward one direction. In comparison, only 56 of the 126 predicted M198 sites for genes in other clusters are oriented the same way. Clearly, including both the M198-score and its site orientation constraints can improve the prediction specificity for cluster 1, as observed by BT for RAP1. However, in a similar procedure comparison as in the previous subsection, adding the orientation constraint of M198 does not improve the global prediction. This orientation constraint may help reduce the false positive rate for cluster 1, but it at the same time increases false positive rates in other clusters. Thus, a fundamental question is: is it appropriate to justify the “authenticity” of a prediction model based on its prediction performance? Our analysis suggests that a combinatorial regulation rule, and perhaps many other causal relationships, may not be reliably inferred using an automatic “learning machine” under a global classification accuracy criterion.
To assess globally whether the TFBS orientation and position information can further help predict gene expression, we added the covariates representing TFBS orientations and positions to the feature list of our model. We performed the same feature preselection and naïve Bayes procedures as described above on the augmented dataset. The classification accuracies for the training sets were very close to the result from using motif score alone. However, the classification accuracies for the test sets were slightly worse than before. This result implies that, although it may be biologically true that orientations and positions of authentic TFBSs have an effect on the binding of the corresponding TFs in some cases, such information for predicted TFBSs do not help in predicting co-expression of genes globally when motif matching scores are given. Even in BT's Bayesian network models, position and orientation constraints were selected only 5.1% and 0.6% of the time, respectively. In both of the strong cases detailed in [19], we were able to find a simpler rule (matching scores only) that is as sensitive and specific as or better than the combinatorial rules reported by BT.
We would like to caution the reader again, however, that our results cast doubts on some of these delicate model interpretations of BT but do not imply that the position and orientation of TFBSs are biologically unimportant.
So far we have followed BT's approach as closely as possible: using the same set of motif features generated by [19] and employing exactly the same CV procedure as theirs. The only difference between our and their approach is that we used the naïve Bayes model, whereas they used the more complex Bayesian network.
However, we cannot help notice that the 615 de novo motifs (excluding the 51 known motifs) generated by [19] were found by using the Gibbs motif sampler AlignACE [20] to search the upstream sequences of all genes in both the training and the test datasets for each cluster. These motifs were further optimized so as to be more specific to the respective clusters they were discovered from by a simulated annealing procedure [19], still using all genes in both the training and test datasets. These steps inevitably generate motifs (features) that are already biased in favor of the existing clustering in the test set. In a valid CV procedure, only the information for the training set genes, including both their upstream sequences and their cluster labels, are allowed to be used in both feature extraction and model training.
To correctly measure how much of gene expression information can be predicted by DNA sequence features, we implemented a valid 5-fold CV procedure, still using the gene clustering result of BT. First, genes in each cluster were divided into five sets of approximately equal sizes at random. Each time, we left out 20% of genes (one subset of genes for each cluster), and used the remaining 80% of genes (i.e., the training set) and their upstream sequences for de novo motif finding via AlignACE [20]. These motifs were then optimized by a simulated annealing algorithm. The total number of motifs we found ranged from 600 to 700 for each training set, which is consistent with the number of 666 motifs in [19]. We then preselected the top 20 motifs (see Figure S3) for each cluster and trained naïve Bayes classifiers based on the training set and the preselected motifs. Finally, the classifiers so trained were used to predict the cluster memberships of the left-out 20% genes. The classification accuracy of this correct CV procedure is 61% according to the criterion in [19], which is still significantly higher than random guessing. When we further added the 51 known motifs to the motif sets, the classification accuracy increased to 64%.
Note that we cannot directly use the motif finding and model-fitting procedure of [19] because their complete algorithm is not publicly available. Furthermore, their-model fitting procedure needs bootstrapping replications and can be overly time consuming, unstable, and nonreproducible. Thus, there is a possibility that the low accuracy of our correct CV procedure is caused by the lower capability of our motif finding strategy compared to that of [19]. To calibrate with BT's approach, we also applied the exact same incorrect CV procedure as in [19] using our own motif finding, optimization, and model-fitting strategies described above. When using all the genes in all clusters, our de novo motif discovery strategy found altogether 650 motifs, and the whole procedure yielded a classification accuracy of 75%, which is slightly higher than the result of [19] (73%). Based on these results, we conclude that the incorrect CV procedure of [19] has likely overestimated the true prediction accuracy of their expression prediction method by 10%–15%.
The naïve Bayes model we adopted is essentially the simplest version of the Bayesian network. The assumption of conditional independence of the covariates is far from realistic in most applications, as well as in this study. However, it outperformed the more complicated Bayesian network, as well as SVM, CART, logistic regression, and Bayesian logistic regression [25] (unpublished data) for this study. As described by Domingos and Pazzani [26], optimality in terms of zero-one loss (classification error) is not necessarily directly connected to the quality of the fit of a probability distribution. Rather, as long as both actual and estimated distributions agree on a most-probable class, the classifier will have a reasonable performance.
Although it is not rare to see successful examples of the naïve Bayes method, the feature selection step is always challenging. In our method, features are considered independently. Each feature is dichotomized to 0 or 1 according to a threshold that maximizes a Chi-square test statistic. In this way, features that are highly associated with a target cluster will be selected as covariates in the naïve Bayes model of this cluster. Our method selects not only the features that are enriched in the target cluster, but also those that are “depleted” in the target cluster but enriched in other clusters. The latter type of features can be explained as a logic operator “NOT”.
Dichotomization of motif scores in our procedure is a gross simplification. Although the binding of a TF to DNA may not be a simple 0–1 trigger, it is easier to model it in this way, and it is also interesting to see whether this simple model can help predict gene expression. We expect to lose some information through discretization, but it is not clear how much the lost information can help the classification problem. It is a worthwhile future project to explore possibilities of using the continuous data, both motif scores, and gene expression values, directly and more efficiently.
Our study has shown that it is perhaps not very sensible to justify a model's “authenticity” by its global prediction performance, and one may easily inject subjective interpretations into the inference results, especially when the prediction uncertainty is not explicitly quantified. This in fact is a challenge for many machine learning approaches, and researchers have begun to pay attention to the problem of estimating prediction uncertainties. In this regard, it is perhaps beneficial to act more like a real Bayesian when using Bayesian tools. That is, these tools not only provide point estimates, but also posterior distributions, which summarize all the information in the data and quantify uncertainties of the estimates.
The keen difference between the correct and incorrect CV procedures reminds us how easy it is to be overconfident. Similar mistakes have also been uncovered in some computational biology studies in which knowledge from literature is used to help construct gene clusters or biological networks and these results are then evaluated and validated by GO analysis, which is by itself a product partially based on the literature.
Although it has been accepted as common knowledge in biology that TFBSs' orientation and position have a functional role in affecting gene regulation activities, and anecdotal examples abound [27,28], it is still nonconclusive how the orientation and position information of putative TFBSs can help one discern true TFBSs from sporadic sequence matches that exert no regulatory functions. In particular, the TFBS orientation and position information did not help us improve the classification accuracy globally, and was not even obviously useful in the two strongest cases detailed in [19]. Since the Bayesian network in [19] is more prone to overfitting, the danger of overinterpreting the fitted models can be a serious threat. In a recent study of nucleosome positioning in yeast, Yuan et al. [29] observed that true regulatory elements are highly enriched in nucleosome depleted regions. Thus, certain sequence information at a scale of nucleosome binding regions (larger than TF binding sites) may be more useful than orientation and position information in differentiating true TFBSs from false ones.
For motif j, its score for gene i is denoted as sij, which is computed in [19] as either zero, when motif j has no predicted occurrence in the promoter of gene i, or the highest matching score among all predicted occurrences of the motif in the promoter of gene i. In this way, a score matrix S = (sij)2587×666 can be built directly from the supplement data of [19].
The continuous scores sij are discretized into 0 or 1 by a thresholding procedure described below. In a word, a threshold for the scores corresponding to a motif is chosen so as to maximize the specificity of TFBSs for the cluster of interest. Let N be the number of all the genes in consideration (i.e., 2,587) and let yi be the class label of gene i (i ∈ {1,
···,}N). Among these N genes, Nk,1 of them are in class k (defined as positive set) and Nk,0 are not in class k (defined as negative set). Thus Nk,1 = #{i:yi = k}, Nk,0 = #{i:yi ≠ k} and Nk,1 + Nk,0 = N. For motif j (j ∈ {1,
···,666}) and a threshold c, define
The best threshold for motif j in model k is defined as:
where
More intuitively, the above procedure finds the most significant Chi-square test result for the 2 × 2 contingency table of the N's. This procedure makes the distribution of TFBSs in positive set and negative set most different. The thresholds calculated above discretize the score matrix S into a 0–1 matrix and it is denoted as X. Note that the discretized covariate matrix X will be different for fitting models in different classes.
The feature preselection step is simply an extension of the threshold finding procedure. For model k, the best threshold
is calculated for motif j along with its highest χ2 statistic. Features (motifs) are sorted by their χ2 statistics, and the top m ones are included the models. This selection is done for each model separately.
The naïve Bayes method has been widely used in statistical learning. It is based on the very simple assumption that all feature variables (covariates) are independent given the class label of the sample. We use cluster 1 and its preselected m motifs as an example to describe our naïve Bayes model fitting procedure. Denote the class label variable as Y and the preselected top m covariates as X1,
···, X m. Using the Bayes theorem, we have
Thus, the odds ratio can be computed as
We further assume Bernoulli models for each Xj given Y and class label variable Y itself, i.e.,
The prior distributions for py, p0j, and p1j are set to be uniform. The training set consists of a class label vector y = (y1,
···,yN) and the discretized TFBS score matrix X = (xij),i = 1,
···,N; j = 1,
···,m. Given the training set, the posterior distribution of py, p0j, and p1j can be easily calculated as
For a new observation with the covariates vector Xnew = (X1,new,...,Xm,new), we have
Thus, we have the predictive odds ratio for this new observation as
For the 49 classes, 49 models are fitted and the genes in the test set are assigned to the class with the respective model that fits the data best. Specifically, for k = 1,
···,49, the odds ratio
can be calculated and a gene will be assigned to a class k* with the highest odds ratio.
To reduce the complexity, for each motif on each gene we only consider the orientation and position of the site with the highest matching score. The site orientation is coded into two separate binary variables, xl and xr, where xl = 1 indicates that the predicted site is left-oriented (away from ATG), xr = 1 for right-oriented, and xl = 0 or xr = 0 otherwise. Note that when a gene does not contain TFBS for a specific motif, the corresponding xl and xr are both 0. The TFBS position in [19] is a continuous variable representing the distance of the TFBS to ATG. We set it to a very large number if a motif has no occurrence in the promoter region of a gene. In our naïve Bayes procedure, the new variable d is a dichotomized version of the original position variable based on an optimized distance threshold, so that d = 1 means that the distance from the predicted site to ATG is smaller than the chosen threshold. |
10.1371/journal.ppat.1002522 | Biochemical Properties of Highly Neuroinvasive Prion Strains | Infectious prions propagate from peripheral entry sites into the central nervous system (CNS), where they cause progressive neurodegeneration that ultimately leads to death. Yet the pathogenesis of prion disease can vary dramatically depending on the strain, or conformational variant of the aberrantly folded and aggregated protein, PrPSc. Although most prion strains invade the CNS, some prion strains cannot gain entry and do not cause clinical signs of disease. The conformational basis for this remarkable variation in the pathogenesis among strains is unclear. Using mouse-adapted prion strains, here we show that highly neuroinvasive prion strains primarily form diffuse aggregates in brain and are noncongophilic, conformationally unstable in denaturing conditions, and lead to rapidly lethal disease. These neuroinvasive strains efficiently generate PrPSc over short incubation periods. In contrast, the weakly neuroinvasive prion strains form large fibrillary plaques and are stable, congophilic, and inefficiently generate PrPSc over long incubation periods. Overall, these results indicate that the most neuroinvasive prion strains are also the least stable, and support the concept that the efficient replication and unstable nature of the most rapidly converting prions may be a feature linked to their efficient spread into the CNS.
| Prion diseases are fatal neurodegenerative disorders that are also infectious. Prions are composed of a misfolded, aggregated form of a normal cellular protein that is highly expressed in neurons. Prion- infected individuals show variability in the clinical signs and brain regions that selectively accumulate prions, even within the same species expressing the same prion protein sequence. The basis of these divergent disease phenotypes is unclear, but is thought to be due to different conformations of the misfolded prion protein, known as strains. Here we characterized the neuropathology and biochemical properties of prion strains that efficiently or poorly invade the CNS from their peripheral entry site. We show that prion strains that efficiently invade the CNS also cause a rapidly terminal disease after an intracerebral exposure. These rapidly lethal strains were unstable when exposed to denaturants or high temperatures, and efficiently accumulated misfolded prion protein over a short incubation period in vivo. Our findings indicate that the most invasive, rapidly spreading strains are also the least conformationally stable.
| Prion diseases are fatal neurodegenerative disorders that include Creutzfeldt-Jakob disease (CJD) in humans, bovine spongiform encephalopathy (BSE) in cattle, and chronic wasting disease (CWD) in cervids (reviewed in [1]). These disorders are caused by misfolding of the cellular prion protein, PrPC, into a β-sheet-rich, aggregated isoform known as PrPSc [2]–[5]. PrPSc can exist as distinct conformational variants or strains, which show strikingly different disease phenotypes even when PrPC sequences are identical [6]–[10]. Prions strains may vary in their aggregate size, stability in chaotropes, PrP epitopes exposed, glycosylation profile, and core of shielded hydrogen atoms as assessed by H/D exchange and mass spectrometry [11]–[15]. Nonetheless, the critical conformational features of PrPSc that drive rapidly lethal disease remain unclear.
Conformational determinants of PrPSc that impact the key events in prion pathogenesis are emerging. Incubation periods from time of exposure to terminal disease vary widely among prion strains, sometimes by more than two-fold [16]. Prion stability, or resistance to denaturation, has been assessed in chaotrope-based assays and has revealed that short incubation period strains correlated with unstable PrPSc in mouse [17], yet with stable PrPSc in hamster [18]. Together these findings suggest that other factors such as differences in prion structure or cellular processing influence survival times.
PrPSc particle size varies among prion aggregates, from oligomers to long fibrils, with the most highly infectious PrPSc size identified as small oligomers [19], [20]. Prion strains also show differences in the amount of proteinase K (PK)-sensitive versus PK-resistant aggregates [21], [22], with some of the most virulent strains having an estimated 80% of aggregates being PK-sensitive [13], [21], [23]. Additionally, the electrophoretic mobility of the PK-resistant core can differ between strains [11]. Thus, the conformational variability among distinct prion strains is pronounced and correlates with some features of disease.
Following entry of prions, many prion strains accumulate in lymphoid tissue during the early stages of disease. Prions spread to the CNS through peripheral nerves in a process known as neuroinvasion [24], [25]. Although it has not yet been directly demonstrated how PrPSc transits via peripheral nerves, a wealth of indirect evidence exists that prion transport by nerves is a major entry route into the brain [26]–[30]. Manipulating the density of nerves or the distance between nerves and the PrPSc peripheral reservoirs in lymphoid tissue has a robust impact on neuroinvasion [26], [27], [31]. Additionally, gastric, ocular, or lingual exposure to prions induces prion replication initially at CNS sites with direct neural connections to the entry site, also suggestive of neural transport of prions into the CNS [30], [32]–[34]. Yet little is known about how strains differ in their capacity for neuroinvasion.
Neuroinvasion is dependent upon (1) host and (2) the specific strain. Certain critical host factors that enhance neuroinvasion, such as CD21/35 complement receptor or C1q, have been identified through prion infection studies in which receptors or complement were depleted [35]–[37]. Yet within the same host species, distinct strains show varying abilities to invade the CNS. For example, the mouse-adapted sheep scrapie strain 87 V is poorly neuroinvasive whereas strain 22A is highly neuroinvasive [38]. Yet, the structural determinants that underlie strain differences in CNS entry are unknown. Here we investigated the biochemical properties of prion strains that efficiently or poorly invade the CNS. We demonstrate that the most rapidly lethal strains are highly neuroinvasive, physically unstable, and form primarily diffuse aggregates in the brain.
To compare the pathogenesis of prion strains, we inoculated WT mice (VM/Dk background) with mouse-adapted strains 22 L or 87 V by intracerebral (IC) or intraperitoneal (IP) routes. Prion infection in brain was determined by three independent assays: western blot, histoblots, and immunohistochemistry. Following IC inoculation, both 22 L and 87 V led to terminal prion disease in all mice, although with a significantly longer incubation period in 87 V-exposed mice (22 L: 200±2 days; 87 V: 302±2 days) (Figure 1A). Yet after IP inoculation, only 22 L was neuroinvasive and led to the consistent development of clinical prion disease and PrPSc in the brain (5/5, 100% attack rate, Figure 2A, left panel). No prions were detected in the brains of mice inoculated IP with strain 87 V (0/6), consistent with previous reports indicating that CNS invasion was inefficient for this strain (Figure 2A, right panel).
To determine how additional strains vary in their ability to neuroinvade, we inoculated mice expressing WT mouse PrPC (Tga20 mice) with mouse-adapted CWD (mCWD) and mouse-adapted scrapie strains RML, 22 L, and ME7 by the IC route. Tga20 mice overexpress a WT PrP sequence that varies from VM/DK mice by two dimorphisms (108L/F and 189T/V) [39]. Here we again found that following IC inoculation, the strains led to disease with 100% attack rate but showed significant differences in the time to terminal prion disease. The mean incubation period for strains RML, 22 L, and ME7 was 71–85 days, whereas the mean incubation period for strain mCWD was significantly longer at 164 days (Figure 1A). Upon IP inoculation of RML and mCWD, the rapid strain RML developed prion disease with an incubation period that was extended 40 days beyond that of the IC route and had an attack rate of 100% (n = 10). In contrast, the slow mCWD strain developed prion disease with an incubation period of approximately 460 days and a very low attack rate of 20% (3/15) (Figure 2B). ME7 and 22 L are well-established to be strongly neuroinvasive strains following IP inoculation [35], [40]–[43]. Taken together with the WT mouse infections, we found that the rapid strains (22 L, RML) tended to be strongly neuroinvasive with all mice developing prion disease following a peripheral exposure, whereas the slower strains (87 V, mCWD) tended to be weakly or non-neuroinvasive.
We found dramatic differences in the PrPSc aggregate morphology in brains of mice that were IC inoculated with strongly and weakly NI strains. In the WT brain, the strongly NI strain, 22 L, led to diffuse PrPSc deposits that did not stain with the amyloid binding dye Congo red, hence deposits are referred to as noncongophilic. In contrast, the weakly NI strain, 87 V, led to dense, large aggregates that bound CR, and are referred to as congophilic (Figure 1B). Similarly, in the Tga20 brain, RML, ME7, 22 L, all strongly NI strains [31], [35], [40]–[43], led to diffuse, fine 2–5 µm aggregates and occasionally small 10–15 µm plaques that were noncongophilic, whereas the weakly NI strain, mCWD, led to focal, dense, large 20–50 µm plaques that were congophilic (Figure 1B). Ultrastructurally, the strongly NI strains had no visible fibrils consistent with previous reports for ME7 and RML [44], [45], whereas the weakly NI prion mCWD consistently had long fibrils visible within the plaques (Figure 1C and Figure S1). Plaques of fibrils have been previously observed for 87 V [45].
We then compared the prion aggregate morphology of the same strain entering the brain following either an IC or IP exposure. The strongly NI strains showed identical aggregate morphologies in the brains of mice exposed either by the IC or IP routes (Figure 2A,B). However in the weakly NI strains, there were surprising differences in the PrPSc aggregate morphology. After IP inoculation, the weakly NI strains showed none of the large congophilic aggregates in the brain that were seen after the IC inoculation, although scattered diffuse aggregates were seen in the brain of three Tga20 mice IP exposed to mCWD (data not shown).
We next assessed whether PrPSc accrued in the lymphoid tissues after IP inoculation with the different strains. Here we found that PrPSc was detectable in spleens of all mice inoculated with the strongly NI strains RML and 22 L, but in none of the weakly NI 87 V-inoculated and only approximately 20% of mCWD-inoculated mice (Figure 2). Thus, the strongly NI strains also developed a more robust accumulation of prions in the lymphoid tissues.
Having found that the less neuroinvasive prions form large, dense congophilic core plaques in histologic sections suggested that there were structural differences among the strains that may alter their ability to spread to the brain. To test this hypothesis, we first investigated the structural and biochemical differences underlying the biological properties of the strongly and weakly NI prions. We compared the mobility of the PK resistant core fragment and the ratios of di-, mono- and unglycosylated PrPSc glycoforms. In assessing the PK-resistant core fragment, we found no differences in the electrophoretic mobility among the strains. The glycoform ratios varied among the strains, however neither the PK-resistant PrP core size nor the glycoform profile correlated with neuroinvasion ability (data not shown).
Prion strains have been shown to vary dramatically in their conformational stability. To next determine how the strongly and weakly NI prions varied in their stability in chaotropes, we exposed brain homogenates to 14 concentrations of guanidine hydrochloride (GdnHCl) ranging from 0 M to 6 M. We then diluted the samples to 0.15 M GdnHCl, digested with PK, denatured and measured the PrPSc remaining by ELISA, and calculated the [GdnHCl]1/2. ELISA was used to quantify PrPSc since western blots of murine prion strains showed no visible mobility shift in PK-resistant PrPSc at the higher GdnHCl concentrations (data not shown). This suggests global PrPSc denaturation and complete PrP digestion with PK occurs, consistent with the results of Peretz and colleagues [46]. Interestingly, the strongly NI, nonfibrillar strains showed the lowest [GdnHCl]1/2, which was 0.9±0.2, 1.1±0.1 and 1.6±0.1 for the 22 L, RML and ME7 strains, respectively. The [GdnHCl]1/2 was significantly higher for the weakly neuroinvasive strains mCWD at 1.9±0.1 and 87 V at 3.2±0.1. (Figure 3A–C). These results indicate that the strongly NI, nonfibrillar, noncongophilic strains 22 L, RML, and ME7 were less stable in chaotropes as compared to the weakly NI, fibrillar, congophilic strains mCWD and 87 V.
To further confirm the stability differences among strains, we also assessed the thermostability of the prion aggregates. Here we digested brain homogenate with PK and exposed PrPSc to a thermal gradient from 25–99°C in the presence of 1.6% SDS for 6 minutes, then performed one dimensional denaturing gel electrophoresis and quantified levels of monomeric PrP, as has been previously shown with yeast prions [47]. These measurements indicate how readily the aggregates disassemble into monomers at increasing temperatures. We determined the temperature at which half of the total PrPSc, measured at 99°C, appeared as monomers (T1/2). The thermostability and chaotrope stability measurements showed remarkable agreement. Whereas the strongly NI strains showed T1/2 from 61–70°C, the weakly NI, fibrillar strains disassembled at higher temperatures, with a T1/2 from 86–90°C (Figure 4). These results again indicated that the strongly NI, noncongophilic strains formed less stable aggregates.
To assess the relative efficiencies of PrP conversion, we measured the amount of soluble and insoluble PrP in brain at terminal disease for each WT strain. The insoluble PrP fraction was significantly higher in the strongly NI 22 L versus the weakly NI 87 V, at 90%±1 versus 53%±7 insoluble PrP, respectively (Figure 5). The 22 L strain notably accumulated 11±4-fold more total PrPSc at terminal disease than 87 V, which occurred over a shorter incubation period. These data suggest that the noncongophilic 22 L is more efficient at rapidly converting PrPC in vivo as compared to the congophilic 87 V strain. The Tga20 strains showed highly variable insoluble∶soluble PrP ratios, likely due to the high PrPC expression.
Since the relative level of PK-sensitive PrPSc may influence the neuroinvasive ability of a strain, we quantified the total insoluble and PK-resistant PrPSc of strains 22 L and 87 V. PK-digested and non-digested aliquots of brain homogenate were ultracentrifuged, and PrPSc was measured in the pellet fractions. Interestingly the levels of PK-resistant PrPSc were similar and consisted of 40±3% and 46±6% of the total insoluble PrP for 22 L and 87 V, respectively, indicating that >50% of the total insoluble PrPSc was PK-sensitive (Figure S2).
To determine whether the total insoluble PrPSc more readily dissembled into monomers at lower temperatures as compared to the PK-resistant fraction, the thermostability measurements were repeated on total insoluble PrPSc in the absence of PK. The T1/2 was similar to the values determined for the PK-resistant PrPSc for both the 22 L and 87 V (T1/2: 22 L = 65°C±5; 87 V = 92°C±3) (Figure S3).
Infectious prions invade the body through peripheral sites such as the gastrointestinal tract, and following amplification in lymphoid tissues for some strains, spread mainly via peripheral nerves into the CNS [48]. Experimentally, certain fibril-forming prion strains replicate peripherally and do not enter the CNS, leading to a persistence of prions in extraneural tissues [24], [38], [49]. GPI-anchorless prion strains are also fibrillar and also show infrequent neuroinvasion after inoculation by tongue, ocular, intravenous, or intraperitoneal routes [50]. Natural infection with similar weakly or non-neuroinvasive strains may yield asymptomatic, long-term carriers of infectious prions, thus could pose a risk for transmission to other humans or animals. Here we investigated the pathologic phenotype and the biochemical properties of strongly and weakly neuroinvasive prion strains using a range of assays. Our results establish that the strongly neuroinvasive murine strains are less stable and efficiently accumulate PrPSc over a short incubation period.
The strongly and weakly neuroinvasive mouse strains showed profound differences in the PrP aggregate morphology and incubation period to terminal disease. After intracerebral inoculation, strongly neuroinvasive strains form diffuse, nonfibrillar PrP aggregates and mice rapidly progressed to terminal disease. In contrast, weakly neuroinvasive strains form dense, congophilic, fibrillar plaques and mice slowly progressed to terminal disease. These findings suggest that the congophilic, fibrillar PrPSc accumulates slowly or is less toxic, the latter consistent with recent evidence that fibrils in general are less toxic than oligomers [51]. Interestingly, there are new reports of strains that form large, dense plaques and show prolonged incubation periods, such as vCJD prions in transgenic mice expressing human 129VV PrPC [52] and Gerstmann–Straussler-Scheinker (GSS) prions in TgPrP101LL mice [53]. Similar to our mCWD strain, the vCJD and GSS mouse strains show a predilection for the corpus callosum region [52], [53].
We found that the pathogenesis diverged among the strains inoculated IP, in that only the diffuse, noncongophilic strains were strongly neuroinvasive. Since several murine prion strains inoculated IP require amplification in the lymphoid tissues for efficient neuroinvasion to occur [54], [55], it is possible that the strongly neuroinvasive prion strains replicate readily to higher levels in lymphoid tissue, which may facilitate spread through peripheral nerves. Here the strongly neuroinvasive strains had more abundant PrPSc aggregates in the spleen, and had a short incubation period after IC inoculation. Together these findings suggest that the strongly neuroinvasive strains could propagate efficiently in lymphoid tissues, which may enhance neuroinvasion.
What is the biochemical basis for a strain's efficient spread to the CNS? Marked differences were detected in the stability assays where the strongly neuroinvasive strains were the most unstable. The stability could account for the differences in the disease pathogenesis. First, it is possible that the unstable strains have small PrPSc aggregates that efficiently spread via peripheral nerves to the CNS. Consistent with this hypothesis, previous studies have shown that small or low density PrP aggregates were unstable and rapidly lethal after IC inoculation [20]. Whether these strains also spread efficiently after IP inoculation is unknown. Second, the higher stability of the congophilic strains may lead to such slow fragmentation and PrPC recruitment that incubation times exceeding the mouse lifespan would be required for prion replication in spleen and spread to the CNS. Although we did not detect splenic PrPSc in 87 V-infected mice, others have shown with sensitive infectivity assays that 87 V accumulates in the spleen at early timepoints post-inoculation [38]. This suggests that the failure of 87 V to spread to the CNS was not due to a delayed replication in the spleen [38]. A third possibility is that instability of a strain generates high levels of infectious particles necessary to initiate peripheral nerve spread. Indeed, a report of 87 V indicates that very high doses of infectious prions can lead to brain entry [56]. The lymphoid tissues may serve to amplify PrPSc to high levels near nerve terminals, and thus efficient lymphoid PrPSc amplification may be crucial for neuroinvasion after IP inoculation. Alternatively, the strain may change after replication in the spleen. Lastly, it is possible that the instability does not underlie the efficient spread of prions, but that there is another yet undiscovered factor, such as a PrPSc interacting protein, that is required for spread. These possibilities are not mutually exclusive.
Previous studies indicate that for many strains, a major fraction of PrPSc consists of PK-sensitive conformers [57], consistent with our findings of 87 V and 22 L. In human CJD cases, higher levels of stable PK-sensitive conformers were associated with extended disease duration [58]. Similarly, in hamster strains, high levels of PK-sensitive PrPSc correlate with longer incubation periods [13]. In mice, the level of PK-sensitive forms was found to be higher in RML as compared to ME7 [59]. Although a full characterization of the role of the PK-sensitive and resistant conformers in neuroinvasion is beyond the scope of this study, it will be interesting in future studies to determine which fraction is more involved in neuroinvasion.
Is congophilia predictive of biological behavior? CR staining has been a gold-standard for the histologic diagnosis of amyloid fibrils, in which tightly interdigitated β-sheets are aligned perpendicular to the fibril axis. Recently, CR was shown to bind parallel to the fibril axis in a surface groove of yeast Het-s fibrils, and to require electrostatic interactions for binding [60]. One key residue exchange in Het-s abrogated the binding of CR, even though fibrils were present, thus CR-negative status of aggregates does not necessarily indicate a lack of fibrillar structure. Nevertheless, the noncongophilic prion strains RML and ME7 used here previously have not revealed fibrils after direct examination of the brain ultrastructure [44], [45]. In studies of partially purified hamster prions, fibrils were found only when PrPSc was truncated by digestion with PK [61], [62]. To our knowledge, fibrils of full length PrP have not been observed for noncongophilic prion strains [62]. In contrast, fibrils are readily detected for strains mCWD [63] and 87 V [45]. Taken together, these studies suggest that for the mCWD and 87 V strains, congophilia correlates with plaques of PrP fibrils.
Notably, the unstable and stable prion strains reported here share properties with the yeast Sup35 prion strains, Sc4 and SCS. Sc4 prions are thermally unstable, readily fragment to generate short fibrils, and have low levels of soluble monomers, whereas the SCS prions are highly thermally stable, rarely fragment, develop long fibrils, and have high levels of soluble monomers [64]. By comparison, strain 22 L, like Sc4, is thermally unstable and has low levels of soluble monomers, whereas 87 V, like SCS, is highly stable with long fibrils and high levels of soluble monomers. Although we did not directly measure fibril breakage or the propensity to shear, in yeast strains, low thermal stability correlates with high fragmentation rate [64]. Whether this is also true for PrP remains to be determined.
Recent studies in yeast have shown that transmission of a prion aggregate to the daughter cell selects for a certain aggregate size [65], which is governed in part by fragmentation properties. In mice and yeast, the prion stability appears to greatly influence prion amplification, and lends support to the proposed mathematical model that fibril breakage is a dominating factor in the kinetics of prion propagation [66].
This study shows that the more unstable, noncongophilic and nonfibrillar murine prion strains efficiently spread from extraneural sites to the central nervous system. Future studies on the relationship between the biochemical properties of misfolded proteins and the disease phenotype are essential for deciphering how aggregates spread in prion and other protein misfolding diseases.
All procedures involving animals were performed to minimize suffering and were approved by the Institutional Animal Care and Use Committee at UC San Diego. Protocols were performed in strict accordance with good animal practices, as described in the Guide for the Use and Care of Laboratory Animals published by the National Institutes of Health.
WT (VM/Dk inbred mice, kindly provided by Dr. Byron Caughey) or Tga20 transgenic mice [67] (groups of n = 4–15 mice) were intracerebrally inoculated into the left parietal cortex with 30 µl of a 0.1% or 1% (w/v) prion-infected brain homogenate prepared from terminally diseased mice. Strain 87 V was a gift from Dr. Thomas Wisniewski. Uninfected brain homogenate served as a negative control. Intraperitoneal inoculations were performed using 100 µl of a 0.1% or 1% prion-infected or uninfected brain homogenate. Mice were monitored three times weekly, and TSE was diagnosed according to clinical criteria including ataxia, kyphosis, stiff tail, hind leg clasp, and hind leg paresis. Mice were sacrificed at the onset of terminal disease and incubation period was calculated from the day of inoculation to the day of terminal clinical disease. Mice were maintained under specific pathogen-free conditions.
Histoblots were performed on tissue cryosections as reported in Taraboulos et al. [68], using up to 100 µg/ml of PK for the digestion of PrPC. Histoblots were developed using the anti-PrP POM1 antibody (epitope in the globular domain, amino acids 121–231, a kind gift from Dr. Adriano Aguzzi) [69].
Two-µm thick sections were cut onto positively charged silanized glass slides and stained with hematoxylin and eosin, or immunostained using antibodies for PrP (SAF84) or GFAP for astrocytes. For PrP staining, sections were deparaffinized and incubated for 5 min in 88% formic acid, then washed in water for 5 min, treated with 5 µg/ml of proteinase-K, and washed in water for 5 min. Sections were then autoclaved in citrate buffer (pH 6), cooled for 3 min, and washed in distilled water for 2 min. Immunohistochemical stains were performed using the TSA Plus DNP kit (PerkinElmer). Sections were blocked and incubated with anti-PrP SAF-84 (SPI bio; 1∶400) for 45 min followed by anti-mouse HRP (Jackson Immunolabs; 1∶500) for 30 min. Slides were then incubated with anti-DNP-HRP (PerkinElmer, 1∶100) for 30 min, followed by 6 min incubation with DAB. Sections were counterstained with hematoxylin.
Samples were electrophoresed in 10% Bis-Tris SDS-PAGE gels (Invitrogen), transferred onto a nitrocellulose membrane, and PrP detected using the anti-PrP primary antibody POM1 and an HRP-conjugated anti-mouse IgG secondary antibody. The blots were developed using a chemiluminescent substrate (ECL detection Kit, Pierce) and visualized on a Fuji LAS 4000 imager. Quantification of PrPSc glycoforms was performed using Multigauge V3 software (Fujifilm). Prior to western blotting, PrPSc was concentrated from the brain homogenates of IP inoculated mice by sodium phosphotungstic acid precipitation as previously described [70].
Tissues were post-fixed in osmium tetroxide, embedded in epon araldite, sectioned with the ultramicrotome, then collected on grids and post-stained using saturated uranyl acetate solution and bismuth subnitrate. Grids were analyzed with a Zeiss EM10 electron microscope.
Brain homogenate in Tris lysis buffer (10 mM Tris-HCl pH 7.4, 150 mM NaCl, 2% sarcosyl) was digested with 50 µg/ml PK for 30 min at 37°C and then treated with phenylmethylsulfonyl fluoride (PMSF) (2 mM final concentration) and Complete TM protease inhibitor (Roche). Individual aliquots were incubated in 1.6% SDS (final) and subjected to temperatures ranging from 25°C to 99°C in 10° intervals for 6 min with shaking at 1000 rpm. Western blotting was performed on the samples by electrophoresis in a 10% Bis-Tris SDS-PAGE gel as described above. PrP signals from monomers were captured and quantified using a Fujifilm LAS-4000 imager and Multigage software. Each strain was analyzed in at least 3 separate experiments using 3–5 mice.
Prion strain stability in guanidine hydrochloride (GdnHCl) was performed as previously described [14] with minor modifications. Briefly, brain homogenates in Tris lysis buffer were incubated in GdnHCl for 1 hr and then diluted with lysis buffer to a final concentration of 0.15 M GdnHCl. Samples were digested with PK at a ratio of 1∶500 (1 µg PK : 500 µg total protein) for 1 hr at 37°C, treated with protease inhibitors, and centrifuged at 18,000 g for 1 hr at 4°C. The pellets were washed with 500 µl of 0.1 M NaHCO3, pH 9.6 and centrifuged for 20 min at 18,000 g. Pellets were denatured in 6 M GdnSCN for 20 min, diluted 2× with 0.1 M NaHCO3 and coated passively onto an ELISA plate. PrP was detected with anti-PrP biotinylated-POM1 antibody and a streptavidin HRP-conjugated anti-mouse IgG secondary antibody. The signals were detected with a chemiluminescent substrate (1-Step Ultra TMB-ELISA, Thermo-Scientific). Each strain was analyzed in at least 3 separate experiments using 3–5 mice. Statistical analysis was performed using a Student's t test.
Brain homogenate in Tris lysis buffer was centrifuged at 150,000 g for 1 hr at 4°C. The supernatant and pellet were separately collected. Proteins in the supernatant were precipitated using cold methanol. Supernatant and pellet proteins were then analyzed and quantified by western blotting for PrP. Each strain was analyzed in at least 3 separate experiments using 3–5 mice.
|
10.1371/journal.pntd.0007261 | In vitro model of postoncosphere development, and in vivo infection abilities of Taenia solium and Taenia saginata | Taenia solium is known to cause human cysticercosis while T. saginata does not. Comparative in vitro and in vivo studies on the oncosphere and the postoncospheral (PO) forms of T. solium and T. saginata may help to elucidate why cysticercosis can occur from one and not the other. The aim of this study was to use in vitro culture assays and in vivo models to study the differences in the development of the T. solium and T. saginata oncosphere. Furthermore, this study aimed to evaluate the expression of cytokines and metalloproteinases (MMPs) in human peripheral blood mononuclear cells (PBMCs), which were stimulated by these oncospheres and PO antigens. T. solium and T. saginata activated oncospheres (AO) were cultured in INT-407 and HCT-8 intestinal cells for 180 days. The T. solium began to die while the T. saginata grew for 180 days and developed to cysticerci in INT-407 cells. Rats were inoculated intracranially with AO and PO forms of either T. saginata or T. solium. Rats infected with T. solium AO and PO forms developed neurocysticercosis (NCC), while those infected with the T. saginata did not. Human PMBCs were stimulated with antigens of AO and PO forms of both species, and the production of cytokines and metalloproteinases (MMPs) was measured. The T. solium AO antigen stimulated a higher production of IL-4, IL-5, IL-13, IFN-γ, and IL-2 cytokines compared to T. saginata AO. In the PO form, the T. saginata PO antigen increased the production of IL-4, IL-5, IL-13, IFN-γ, IL-1β, IL-6, IL-10, TNF-α and IL-12 cytokines compared to T. solium, suggesting that this global immune response stimulated by different forms could permit survival or destruction of the parasite depending of their life-cycle stage. Regarding MMPs, T. solium AO antigen stimulated a higher production of MMP-9 compared to T. saginata AO antigen, which may be responsible for altering the permeability of intestinal cells and facilitating breakdown of the blood-brain barrier during the process of invasion of host tissue.
| Taenia solium and Taenia saginata are two parasites that cause the tissue infection cysticercosis in their intermediate hosts, pigs and cows, respectively. One major difference between them is that T. solium can also cause neurocysticercosis in the human brain, while T. saginata cannot. Neurocysticercosis is thought to be the major cause of adult-onset seizures in developing countries. It is not well understood why only T. solium can survive in human tissue; however, the host inflammatory response likely plays an important role. The authors found that human immune cells stimulated with T. solium in the early stages of the parasite life cycle produced a more robust cytokine response than T. saginata. However, in the mature stage, which occurs once T. solium reaches the brain, T. solium antigens stimulated a lower inflammatory response compared to T. saginata, suggesting the parasite is able to manipulate the host immune response in some way to evade destruction. These findings may support the differences in growth observed by the authors when rat brains were inoculated with either parasite species. This study provides new insights into the different ways T. solium and T. saginata activate the immune response to survive and develop within the host.
| Taenia solium and T. saginata are two taeniid cestodes that cause the diseases taeniasis and cysticercosis [1]. These are zoonotic diseases, and swine and bovine act as intermediate hosts, causing porcine and bovine cysticercosis, respectively. Humans act as the definitive hosts in both T. solium and T. saginata infection leading to taeniasis. In the case of T. solium, humans can also act as accidental intermediate hosts causing human cysticercosis [2]. However, only T. solium causes human cysticercosis, while T. saginata does not [3].
When cysticercosis involves the central nervous system in humans, it is called neurocysticercosis (NCC). NCC is common throughout Latin America, sub-Saharan Africa, most of Asia, and parts of Oceania. Human NCC is believed to be the leading cause of acquired epilepsy worldwide [4,5].
The eggs of T. solium and T. saginata contain a six-hooked larva called the oncosphere [6]. When the eggs hatch, this oncosphere is released into the intestine. Intestinal fluid dissolves the oncospheral membrane, releasing and activating the oncosphere. The T. solium activated oncosphere can then penetrate the intestinal wall. Once in the tissue, usually in the muscle or the central nervous system, the oncosphere can transform into a postoncospheral form, and completely develop into cysticerci—a larval stage that consists of a fluid-filled sac containing an invaginated scolex. When this happens, the parasite produces a variety of molecules, which modulate the host immune response in order to evade parasite destruction [7].
The postoncospheral (PO) form is an intermediate stage between an oncosphere and a fully developed cysticercus in tissue [8]. The PO form of T. solium and T. saginata can be obtained in vitro by co-culture of oncospheres with a monolayer of mammalian feeder cells [9,10]. T. solium oncosphere and in-vitro generated PO forms can develop into cysticercus in rats causing NCC [9,11]. However, little is known regarding the in vitro development of the oncosphere to PO form of T. solium and T. saginata, specifically the immunological events that occur at the host/parasite interface. Also, it is not known if the development of the T. saginata oncosphere to PO form could cause NCC in the rat model as T. solium does. In vitro and in vivo models could lead to a better understanding of host-parasite relationships.
Host immune cells such as macrophages, lymphocytes, and polymorphonuclear leukocytes can produce cytokines and metalloproteinases (MMPs) in order to prevent the development of the parasite [12,13]. Because of this, the parasite has developed mechanisms to evade or modulate the host immune response.
Comparative studies on the oncosphere and the PO form of T. solium and T. saginata are limited. This study focused on the in vitro development of the oncosphere to the PO form for T. solium and T. saginata. These in vitro-developed larvae were then tested for infectivity in rats. Later, we obtained antigens from the T. solium and T. saginata oncosphere and PO forms to stimulate the production of cytokines and MMPs in healthy human peripheral blood mononuclear cells (PBMCs).
HCT-8 and INT-407 cells, obtained from the American Tissue Culture Collection (ATCC, Manassas, VA), were used to obtain T. solium and T. saginata PO forms. Cells were incubated at 37°C in 5% CO2 and grown in a specific medium as recommended by ATCC (EMEM media for INT-407, and RPMI for HCT-8; all medium was supplemented with 10% fetal bovine serum). The medium was changed every two days. Once cell confluency was obtained, cells were harvested using trypsin-EDTA (Sigma Chemical Co). Cells were placed into 24-well plates (1x105 cells per well) for maturation assay. The assays described below were performed when cells formed a monolayer.
Tapeworms were collected after medical treatment of newly diagnosed patients, as described by Jeri et al [14]. Hatching of eggs and oncosphere activation were performed, as described by Verastegui et al [15]. The eggs were obtained from gravid proglottids of adult tapeworms by gentle homogenization in a 2.5% potassium dichromate solution (Sigma, St. Louis, Missouri). Eggs were then washed three times in distilled water with centrifugation steps to collect the eggs between washes in 2500g for 5 minutes. The eggs were hatched, and the oncospheres were released using a solution of 0.75% sodium hypochlorite in water for 10 minutes (Mallinckrodt Baker, Inc, Phillipsburg, NJ). Oncospheres were then washed three times in RPMI medium (Sigma, St. Louis, Missouri), and activated by incubation at 37°C for 45 minutes (in the case of T. solium) or 90 minutes (in the case of T. saginata) with artificial intestinal fluid (1 g pancreatin (Sigma Chemical Co., St. Louis, MO), 200 mg Na2CO3, and 1 ml of fresh porcine bile (for T. solium), or 1 ml of fresh bovine bile (for T. saginata), with enough RPMI 1640 medium (pH 8.04) to make 100mL). After activation, the oncospheres were washed three times with RPMI medium and counted using a Neubauer chamber.
Parallel in vitro maturation assays with INT-407 and HCT-8 monolayer cells, using T. solium and T. saginata activated oncospheres, were conducted in order to compare the morphological characteristics during development of each species using the methodology reported by Chile et al [9]. Ten thousand activated oncospheres were cultured in confluent INT-407 and HCT-8 monolayer cells for two weeks. During that time, the medium was changed every three days. At day 15 of culture, the postoncospheral forms were collected and rinsed twice with fresh medium, then transferred to another well containing a confluent of monolayer cells. This process was repeated every three days for up to six months to allow the postoncospheral forms to continue to develop. Cultures were inspected daily using an inverted microscope (Leitz labovert FS). Parasites were collected at 15, 30, 60, 120, and 180 days of incubation.
To determine if T. saginata AO and PO forms can develop into viable cysts in vivo, 15-day old Holtzman rats, purchased from Universidad Peruana Cayetano Heredia, Lima, Peru, were infected intracranially (in the bregma) with oncospheres and 15-day old PO forms from either species following the methodology reported by Verastegui et al., 2015 [11]. Rats were anaesthetized with ketamine (100 mg/kg body weight) and xylazine (5 mg/kg body weight) before infection.
Six rats were inoculated with 180 T. solium AO in 100 μL of saline solution; seven rats were inoculated with 180 T. saginata AO in 100 μL of saline solution. The negative control was 2 rats inoculated with saline solution.
Eight rats were inoculated with ten T. solium 15-day PO forms in 100 μL of saline solution, eight rats were inoculated with ten T. saginata 15-day PO forms in 100 μL of saline solution, and five rats were inoculated with 100 μL of saline solution as a control. A 24-gauge syringe needle was used.
After four months, necropsy was performed. Rats were anaesthetized with ketamine (100 mg/kg body weight) and xylazine (5 mg/kg body weight). Anaesthetized rats were perfused with 200 ml of PBS and then with 100 ml of 4% paraformaldehyde in PBS. Brains were carefully removed, post-fixed for 24 hours at 4°C with 4% paraformaldehyde in PBS, and stored in 70% ethanol. Brains were observed macroscopically to identify extraparenchymal cysticerci. Five millimeter coronal brain sections were cut until the intraparenchymal cysticerci were observed.
Antigens were obtained from AO and 30 day-PO forms of both parasites. AO and PO were obtained as described above. Parasites were rinsed three times with PBS buffer, sonicated, and centrifuged at 10,000g for 15 min at 4°C. The supernatant (total soluble antigens) was separated, and proteins were quantified using Bradford Protein Assay (Bio-Rad) and stored at -70°C until ready for use.
Healthy volunteers (n = 13) with negative serology for NCC were invited to participate in this study. After the volunteers signed an informed consent form, 10 mL of venous blood was collected from each non-infected donor. PBMCs were collected from 10 ml EDTA blood. The blood samples were centrifuged at 400g on a Ficoll-Hypaque gradient (Ficoll–Paque TM PLUS, GE Healthcare) for 10 minutes at room temperature for the separation of mononuclear cells. Cells were again suspended in RPMI medium plus 5% of inactivated human serum. Cell viability was evaluated with trypan blue and counted in a Neubauer chamber. Each well contained 2x105 group of cells was cultured in a 96-well plate at 37°C with 5% CO2 for 48 hours in an RPMI medium containing 5% of inactivated human serum. Cells were stimulated with 5 μg/ml of phytohemaglutinin (PHA) as the positive control, 20 μg/ml of AO antigens (both species, separately), and 20 μg/ml of antigens from PO forms at 30 days of maturation (both species, separately). At the end of the incubation period, PBMCs were harvested and centrifuged for 10 min at 400g, and supernatants were collected and stored at −70°C until tested for cytokine and metalloproteinase content by multiplex analysis.
MILLIPLEX MAP kit High Sensitivity Human Cytokine Magnetic Bead Panel (Millipore) was used to measure cytokines (IFN-γ, IL-1β, IL-2, IL-4, IL-5, IL-6, IL-10, IL-12p70, IL-13 and TNF-α), and Fluorokine MAP (R&D system, Minneapolis USA) was used to measure MMP (MMP-2 and MMP-9) in the supernatant of stimulated PBMCs following the manufacturer’s instructions. The cytokines and MMP were detected by the Bio-plex 200 system (Bio-Rad Laboratories, Hercules, CA) using Luminex xMAP technology.
Experiments were done according to the Guide for care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Committee on the Ethics Animals of the Universidad Peruana Cayetano Heredia, Lima, Peru (Permit Numbers: 61242).
For cytokines and MMP assays, the median fluorescent intensities of the beads with the cytokines or MMP now bound were converted to concentrations (pg/ml), using a five-parameter logistic model using the Bio-Plex Manager™ 6.0 software (Bio-Rad Laboratories, Hercules, CA). Each MMP´s or cytokine’s concentration was normalized by subtracting the data of medium alone (control). The Mann Whitney test was used to compare two unpaired groups, and ANOVA followed by Tukey’s post-test was used for three or more group using the Prism V 6.0 statistical program (GraphPad). A P-value of <0.05 was considered statistically significant.
T. saginata and T. solium developed from the oncosphere to the PO form in both cell lines, INT-407 and HCT-8 cells (Table 1). In HCT-8 cells, both species of PO forms developed until 60 days of culture, and then began to die. However, after 60 days in INT-407 cells, the T. saginata PO forms continued to develop to cysticerci, while the T. solium PO forms began to die. At 180 days of culture 0.25% of T. saginata oncospheres developed to cysticerci.
The morphological characteristics of the T. saginata and T. solium PO forms obtained from culture in INT-407 monolayer cells were compared. At 15 days of incubation, both species were morphologically similar (Fig 1A), ranging between 60 to 250 μm in size. Both had an oval form without hooks and shared characteristic movement. For both species at 30 days of incubation, PO forms increased in size (between 200 to 1000 μm), maintained their oval form, and formed a protuberance at one end of the body (Fig 1B).
At day 60 of culture, the T. solium PO form had a spherical shape, similar to a cyst without a scolex, and cells accumulated in the protuberance of one end of the body (Fig 1C). The T. saginata PO form had similar characteristics but also showed a neck-like prolongation at one end of the body (Fig 1C), reaching up to 3000 μm. Additionally, at day 60, T. solium PO forms began to die while T. saginata PO forms continued growing until 180 days of culture.
At day 120 of culture, T. saginata developed into a cysticercus with a spherical form and an evaginated pre-scolex containing four characteristic suckers and measuring 4 mm in diameter (Fig 2A). At day 180 of culture, T. saginata cysticerci increased in size compared to 120 day, reaching up to 6 mm and also displayed a well-defined scolex and suckers (Fig 2B).
We infected rats with T. saginata oncosphere and PO forms in order to evaluate if T. saginata can develop into cysticercus in vivo, as T. solium does. Table 2 shows rats that were infected with T. solium oncosphere or 15-days PO form developed NCC, while rats that were infected with T. saginata oncosphere or 15-days PO form did not develop NCC.
Given that T. saginata does not develop to the cysticerci form in humans, we investigated the differences in host immune response to infection by T. solium and T. saginata. To determine this, we evaluated the production of cytokines during stimulation with AO and 30-day PO antigens of both species.
When compared by stages (AO and PO form) species, both T. solium and T. saginata PO forms stimulated a higher production of the majority of cytokines evaluated than the AO forms (S1 Fig). Nevertheless, when compared by species stages (AO and PO form), we observed a higher production of IL-4, IL-5, IL-13, IFN-γ and IL-2 cytokines stimulated by the T. solium AO form (Fig 3A and 3B) compared to the T. saginata AO; but in PO form, the T. saginata 30-day PO form stimulated a higher production of a variety of cytokines, including, IL-4, IL-5, IL-13, IFN-γ, IL-1β, IL-6, IL-10, TNF-α and IL-12 compared to the T. solium 30-day PO form (Fig 4A, 4B and 4C).
To evaluate whether MMPs are involved in the host immune response against the parasites, we stimulated PBMCs with T. solium and T. saginata forms. The T. solium AO stimulated higher MMP-9 production compared to the T. saginata AO, while in PO form there was no significant difference between both species (Fig 5). There was no significant difference in the production of MMP-2 by PBMCs stimulated with different antigens of both parasites (S2 Fig).
The present study demonstrates that T. saginata activated oncospheres can develop into cysticerci in vitro in a human cell line, while T. solium AO do not. In contrast, in the rat model, the T. saginata oncosphere and PO forms did not develop to cysticerci in vivo as T. solium does. We suspect the differences in development between the two species in the in vitro model are due in large part to the environment of the growth media and the specific cell line, whereas in vivo it is the host immune response that plays a predominant role in regulating parasite development and survival. For instance, Heath and Smyth noted that serum used for in vitro culture contains factors unique to each host-parasite system, which can stimulate development of the parasite [16]. In our study, both species were cultured in media containing fetal bovine serum (FBS). FBS is known to contain large amounts of α- and β-globulin, proteins which have been shown to stimulate the development of the T. saginata oncosphere but not the development of other parasites like T. taeniaformis [17]. Additionally, the cell line used for culture likely plays an important role as T. saginata is able to develop to a cysticercus in vitro using cell line INT-407, while T. solium does not. On the other hand, T. saginata and T. solium do not develop into cysticercus using the HCT-8 cell line. INT-407 cells originate from the duodenum while the HCT-8 cell line originates from the colon. Furthermore, it is known that INT-407 cells are contaminated with HeLa cells that express different surface molecules that could promote the development of T. saginata. To our knowledge, this is the first study to achieve in vitro development of T. saginata PO forms to cysticerci.
In the in vivo, T. solium oncosphere and PO forms developed to cysticerci in brain while T. saginata forms did not. This finding is likely due to T. solium’s ability to evade the host immune response, either through binding of host plasma proteins or by synthesizing surface proteins that are antigenically similar to those of the rat [18]. Previous work has demonstrated that T. taeniaeformis oncospheres are able to initiate PO development after acquiring a component of rat serum on their surface, a component that presumably protects the oncosphere in the newly invaded host from recognition as a foreign antigen [17]. In the same study, T. saginata failed to bind this component and was subsequently attacked by the host immune system. A similar observation has been noted in pigs after infection with T. saginata eggs [19]. Another explanation may be genetic factors that play a role in the rat’s resistance to infection from T. saginata [20].
Although the rat is not a natural host of T. solium, we observed in previous studies that the cysticerci that develop in the rat brain are morphologically equal to those that develop in humans, as is the observed inflammatory response and subsequent pathology [11, 21]. Therefore, we believe the rat, like the human, has molecules that prevent the development of T. saginata.
The cytokines response plays an important role in survival of the oncosphere at the time of initial infection. We suspect T. solium stimulates a different cytokines profile than T. saginata allowing the T. solium oncosphere to survive, while T. saginata is destroyed by the host. To probe this hypothesis, we stimulated cytokine and MMP production in healthy human PBMCs using antigens of oncosphere and PO forms from both species.
We observed that T. solium oncospheres stimulate a higher production of IL-4, IL-5, IL-13, IFN- γ and IL-2 compared to T. saginata oncospheres. The IL-4, IL-5, IL-13, are cytokines typically associated with a T helper 2 (TH2) response, the predominant protective immune response against helminthic infections [22]. On the other hand, IFN- γ and IL-2 are associated with a T helper 1 (TH1) response, which is thought to be responsible for destruction of the parasite [23]. The TH2 response is important in protecting against extracellular helminthic parasites through suppression of the TH1 response, neutralization of toxins, and defense of the host against damage [22, 24].
Inflammatory reactions are dependent on a delicate balance between TH1 and TH2 type responses. In the case of the T. solium oncosphere, it appears a mix immune reaction of TH1/ TH2 type response that aids survival of the parasite. Perhaps in doing so, T. solium oncospheres are able to migrate from the vasculature to the brain, where they develop into the PO form. Similar mixed TH1/TH2 phenotypes have been observed in patients with NCC [7, 13] whereas a predominantly TH2-type response is associated with asymptomatic disease [25, 26].
In both species, the 30-day PO form generated an overall greater inflammatory response than the AO form. Levels cytokines increased when exposed to equal concentrations of PO vs AO antigen, suggesting that antigen composition changes as the parasite is maturing [9].
However, the T. saginata 30-day PO form stimulated a profile of pro-inflammatory cytokines (IL-1β, IL-6, IL-12, TNF-α), plus a mix of TH1 and TH2 related cytokines (IL-4, IL-5, IL-13 and IFN-γ) that was stronger than the response produced by the T. solium 30-day PO form. IL-6, a pro-inflammatory cytokine, plays a role in the death of microorganisms by stimulating the behavior of neutrophils [27]; and TNF-α is strongly expressed at the sites of parasite and cell destruction [28]. Together with the overproduction of IL-4, IL-5, and IL-13, these cytokines may mount a response that could destroy the T. saginata PO form and prevents development of the cyst in vivo. Although these cytokines do have the suggested properties mentioned, they also could be involved in complex networks with mixed effects with respect to inflammation, for example IL-6 has been shown to exhibit anti-inflammatory properties [29].
T. solium oncospheres stimulated increased MMP-9 production in PBMCs compared to production by T. saginata oncospheres. MMP-9 is an endopeptidase produced by neutrophils, macrophages, monocytes, and intestinal epithelial cells [30,31]. It can degrade components of the blood brain barrier (BBB) as well as the extracellular matrix, increasing intestinal epithelial permeability [30,32,33]. MMP-9 has been associated with the breakdown of the BBB in a murine model of NCC [34,35] and is present in high concentrations in sera of symptomatic NCC patients [36]. We hypothesize that T. solium stimulates the production of MMP-9 as a means of enhancing epithelial permeability in order to pass restrictive biological barriers like the intestine and the BBB during early stages of infection.
In conclusion, we have found novel evidence to suggest that T. saginata PO forms are capable of developing into cysticerci in the human cell line INT-407 while not in HCT-8 cells. In vivo, T. saginata fails to develop into cysticerci in the rat brain, suggesting there are factors in the host immune system (that are not present in the in vitro culture) that destroy the parasite. In the oncosphere stage, T. solium stimulated a strong mix of TH1 and TH2-related cytokines and MMP-9 production in healthy PBMCs, which may mediate the inflammatory response and promote oncosphere survival in the vasculature, aiding the entrance of oncospheres into the brain. In the PO form, T. saginata stimulated a strong pro-inflammatory and mix of TH1/TH2-related cytokines, responses that could be causing destruction of the parasite in the tissue. These differences between both species of Taenia found in vitro and in vivo could explain why the larval stage of T. saginata does not develop in the human host, while T. solium does, despite having similar life cycles.
|
10.1371/journal.ppat.1002488 | Human-like PB2 627K Influenza Virus Polymerase Activity Is Regulated by Importin-α1 and -α7 | Influenza A viruses may cross species barriers and transmit to humans with the potential to cause pandemics. Interplay of human- (PB2 627K) and avian-like (PB2 627E) influenza polymerase complexes with unknown host factors have been postulated to play a key role in interspecies transmission. Here, we have identified human importin-α isoforms (α1 and α7) as positive regulators of human- but not avian-like polymerase activity. Human-like polymerase activity correlated with efficient recruitment of α1 and α7 to viral ribonucleoprotein complexes (vRNPs) without affecting subcellular localization. We also observed that human-like influenza virus growth was impaired in α1 and α7 downregulated human lung cells. Mice lacking α7 were less susceptible to human- but not avian-like influenza virus infection. Thus, α1 and α7 are positive regulators of human-like polymerase activity and pathogenicity beyond their role in nuclear transport.
| Adaptive mutations in the polymerase complex, such as PB2 E627K and PB2 D701N play a crucial role in influenza virus interspecies transmission. We have shown earlier that PB2 D701N promotes nuclear entry of PB2 by enhanced importin-α binding in mammalian cells (Gabriel et al., PLoS Pathogens 2008; Gabriel et al., Nat. Commun. 2011). In this study, we show that the adaptive mutation PB2 E627K involves an interaction with importin-α by a novel mechanism that does not mediate nuclear transport of viral factors. We found that human importin-α1 and -α7 are required for efficient human- (PB2 627K) but not avian-like (PB2 627E) polymerase activity. Further, human-like polymerase activity correlated with enhanced importin-α1 and -α7 binding to viral ribonucleoprotein complexes (vRNP). Mice lacking importin-α7 were less susceptible to human-like but not avian-like influenza virus infection. Our findings strongly suggest that importin-α isoforms possess functions beyond their primary role in nuclear transport facilitating PB2 E627K mediated host adaptation and pathogenicity.
| Influenza A viruses are able to cross species barriers and transmit to humans leading to disease of various severity. Interspecies transmission of influenza viruses is multigenic involving several viral and cellular factors. Host restriction occurs mainly by two cellular barriers which need to be overcome upon transmission. First, upon entry viruses need to cross the cell membrane by interaction of the viral receptor binding protein, the hemagglutinin (HA), to the adequate host cell receptors consisting of sialic acid-containing glycoproteins [1], [2]. Second, vRNP components (PB2, PB1, PA and NP), especially PB2 and NP need to adapt to the nuclear import machinery in order to establish efficient replication in the nucleus of the new host cell [3]–[5].
The adaptive position in the viral polymerase subunit PB2 at position 627 is a well known determinant of host range and pathogenicity [6]. Influenza viruses of avian origin are characterized by a glutamic acid signature at this position while human viruses predominantly possess a lysine signature [7]. The PB2 E627K mutation has been shown to increase viral polymerase activity [3] and pathogenicity in mammalian hosts [8], [9].
Despite intensive investigations the molecular basis underlying the host adaptive position 627 in PB2 is poorly understood. It has been shown, that polymerase complexes containing PB2 627E display a defect in vRNP complex assembly leading to restricted polymerase activity and impaired virus growth in mammalian cells [10], [11]. It has been postulated that an unknown cellular inhibitor specifically restricts avian-like polymerase activity in human cells [11]. On the other hand, it has also been proposed that there is no evidence for the existence of a mammalian inhibitory factor of avian-like polymerases but instead the absence or low expression of a positive factor is responsible for low avian polymerase activity in human cells [12]. Very recently, the DEAD box RNA helicase DDX17/p72 has been identified among other polymerase interacting proteins [13] to facilitate efficient H5N1 627K virus transcription and replication in human cells [14]. Consistent with the hypothesis that PB2 627K position affects polymerase-host interaction, it was shown that an E627K substitution alters the electrostatic surface potential of the 627-domain resulting in a basic patch, possibly modulating interactions between viral and host factors [15], [16]. However, cellular factors involved in PB2 627K mediated host adaptation and pathogenicity still remain poorly understood.
It was shown that cellular importin-α isoforms play an essential role in influenza virus host adaptation [4], [5]. Importin-α proteins are components of the classical import pathway and act as adaptors recognizing cargo proteins with a nuclear localization signal (NLS). Importin-α/cargo protein complexes facilitate binding to the importin-β1 receptor protein. Thus, cargo proteins are transported into the nucleus as ternary complexes [17], [18]. Adaptive mutations in PB2 D701N and NP N319K have been shown to be adaptations to cellular importins thereby allowing efficient nuclear transport of PB2 and NP and thus resulting in enhanced virus polymerase activity in mammalian cells [3], [4]. It has been recently shown that avian and mammalian influenza viruses possess differential preferences for importin-α isoforms in human lung cells [5]. While growth of highly pathogenic avian influenza (HPAIV) viruses with avian signatures (PB2 627E or PB2 701D) depended on importin-α3, viruses with mammalian signatures (PB2 627K or PB2 701N) depended on importin-α7. Thus, a switch from importin-α3 to importin-α7 dependency occurs upon mammalian adaptation. Analyzing the role of the PB2 701 polymorphism revealed that adaptive mutations in PB2 D701N and NP N319K mediate the switch from importin-α3 to importin-α7 dependency upon avian-mammalian transmission [5]. Whether other host adaptive positions, such as PB2 627K can mediate a switch to importin-α7 dependency similar to PB2 710N was not further investigated. However, functional substitutions between PB2 627K and 701N have been described before. Either position could compensate for the lack of the other position resulting in increased virus transmission in a guinea pig model [19].
We have initiated this study to analyze whether the host adaptive substitution in PB2 E627K is also involved in importin-α dependent host adaptation. Here, we investigated the role of cellular importins in the regulation of PB2 627K mediated influenza virus polymerase activity and pathogenicity.
We have identified human importin-α1 and -α7 as positive regulators of human- (PB2 627K) but not avian- (PB2 627E) like polymerase activity. In contrast, human importin-α3 acts as a general negative regulator of human- and avian-like polymerase activity in vRNP reconstitution assays while virus growth was not affected. Increased human-like polymerase activity correlated with efficient recruitment of importin-α1 and -α7 to vRNPs in human cells. In contrast, avian-like PB2 627E failed to recruit these importin-α isoforms efficiently to vRNPs. Interestingly, subcellular localization of PB2 627 mediated human- and avian-like vRNP components was not affected by importin-α proteins as reported before with PB2 D701N [4] suggesting a novel mechanism triggered by PB2 627K. Consistent with these findings, human-like PB2 627K but not avian-like PB2 627E virus growth depended on importin-α1 and -α7 in human lung cells and displayed reduced pathogenicity in importin-α7 knockout (α7−/−) mice. Our findings described here show that importin-α1 and -α7 act as positive regulators of human- (PB2 627K) but not avian- (PB2 627E) like polymerase activity without affecting nuclear transport of viral vRNPs. Thus, cellular importin-α proteins play an important role in PB2 627K mediated interspecies transmission beyond their primary role in nuclear transport.
We have previously shown that growth of avian and human influenza viruses depends on different importin-α isoforms [5]. Further, it has been proposed that avian-like polymerase complexes containing PB2 627E are specifically restricted by an unknown host factor which is only present in a mammalian host environment [11]. Moreover, the crystal structure of the PB2 domain containing 627K led to a basic patch on the protein surface upon PB2 E627K substitution [15]. Basic amino acids are classical binding motives for importin-α proteins such as found in NLS sequences of cargo proteins.
Therefore, we addressed the question whether cellular importins are involved in the regulation of influenza virus polymerase activity mediated by the PB2 627K polymorphism. We performed a cell-based polymerase activity assay in combination with siRNA-mediated silencing of individual importin-α proteins. vRNPs of WSN containing either PB2 627K or PB2 627E were reconstituted in importin-α silenced 293T cells (Figure 1D) and their activity was compared to vRNPs expressed in control siRNA treated cells which was set 100% (Figure 1A and B). Silencing of importin-α1 and importin-α7 significantly decreased human-like PB2 627K polymerase activity to 31% and 42%, respectively, compared to negative siRNA controls (Figure 1A). In contrast, PB2 627K activity was significantly increased up to 189% in importin-α3 silenced cells. No significant effects were observed in importin-α4 or -α5 silenced cells. In contrast, silencing of either importin-α isoform (importin-α1, -α4, -α5 or -α7) except for importin-α3 did not affect avian-like PB2 627E polymerase activity (Figure 1B). Silencing of importin-α3 led to increased PB2 627E polymerase activity (192%) compared to control siRNA treated cells. However, avian-like polymerase activity was severely impaired in human cells with a 10-fold reduction (10%) compared to human-like polymerase activity (Figure 1C, left). On the other hand, avian-like polymerase activity was slightly increased (132%) in comparison to human-like polymerase activity in avian cells (Figure 1C, right). This further confirms that avian-like PB2 627E polymerase complexes are restricted in human but not in avian cells.
Here we show that importin-α1 and -α7 are required for efficient human-like but not avian-like polymerase activity in human cells. In contrast, human importin-α3 restricts both human and avian polymerase activities. These findings suggest that importin-α1 and -α7 are positive regulators of human-like polymerase activity while importin-α3 is a negative regulator of human- and avian-like polymerase activities.
In order to understand the molecular basis for differential regulation of polymerase activities by importins, we assessed binding properties of PB2 627K-FLAG and PB2 627E-FLAG in human cells by immunoprecipitation analysis. As cargo proteins are transported as ternary complexes into the nucleus, in this case PB2/importin-α/importin-β1 [17], [18], we assessed the binding of PB2 to importin-α isoforms and detected importin-β1 levels precipitated from PB2/importin-α complexes as described before [4]. Both, avian- and human-like PB2 proteins were found to interact with endogenous importin-α isoforms as well as the corresponding importin-β1 receptor to similar affinities (Figure 2A and 2B). Co-immunoprecipitations studies with co-transfected importin-α-FLAG and PB2 627K or PB2 627E were carried out to investigate the importin-α binding properties to both PB2 proteins. Here, no significant differences have been observed in importin-α or importin-β1 binding affinities to PB2 627K or PB2 627E. Both PB2 variants showed highest binding affinity to importin-α4. Importin-α5 and -α7 displayed lowest affinity to the importin-β1 receptor protein (Figure 2C and 2D).
In summary, we show that human- as well as avian-like PB2 proteins possess similar binding affinities to their adaptor importin-α and receptor importin-β1 complexes. These findings suggest that PB2 E627K substitution does not alter importin-α binding or result in differential importin-α/importin-β1 binding properties suggestive of altered nuclear transport.
Previous reports highlighted the involvement of PB2 E627K substitution in vRNP complex formation via improved binding of the PB2-627K containing trimeric polymerase to NP [10], [11]. It has been shown that single NLS-containing cargo proteins might be bound and transported into the nucleus by several importin-α isoforms. However, these cargo proteins might differ in their importin-α isoform binding preferences when expressed in complexes or in competition with other cargo proteins for nuclear import. It has been demonstrated that both, the NLS and the protein context are responsible for distinct importin-α binding specificities upon competition of multiple substrates for the limited amount of importin-α proteins [20]. Since importin-α binding to monomeric PB2 did not provide an explanation for the altered polymerase activities in importin-α silenced cells, we next investigated whether binding of importin-α proteins to vRNP complexes was affected.
Therefore, viral vRNP complexes were reconstituted in 293T cells by co-transfection of expression plasmids for PB2-627K-FLAG or PB2-627E-FLAG, PB1, PA and NP along with pPol-I-NP-Luc to provide an RNA template. vRNP complexes were precipitated from cell lysates using the FLAG-tagged PB2 proteins and the amount of importin-α1, -α3, -α4, -α5, -α7 and -β1 co-immunoprecipitated along with both vRNP complexes was analyzed by Western blotting and quantified (Figure 3A and 3B). vRNP complex formation was confirmed by detection of co-immunoprecipitated PA as well as NP protein. PB2-627K-FLAG or PB2-627E-FLAG containing vRNP complexes precipitated comparable amounts of PA suggesting that PB2 E627K substitution does not affect trimeric polymerase complex formation. However, vRNP complexes containing PB2-627E-FLAG precipitated less NP protein (55%) compared to human-like vRNP complexes. Additionally, we observed a strong reduction in binding to importin-α1 (40%), -α5 (49%) and -α7 (51%). No obvious differences have been observed for importin-α3 and -α4 with both vRNP complexes. Interestingly, similar amounts of the importin-β1 receptor were precipitated (Figure 3B).
Next, co-immunoprecipitation assays were performed with the polymerase complex (PB1, PB2 and PA) omitting NP (Figure 3C and 3D). Remarkably, differential importin-α binding was lost with PB2 627K and PB2 627E polymerase complexes, indicating that NP is involved in importin-α1, -α5 and -α7 binding to human-like polymerase complexes.
To further understand the role of NP in importin-α binding, interaction between NP and endogenous importins was confirmed in 293T cells (Figure 4A). Co-immunoprecipitation assays were performed with importin-α-FLAG and NP proteins expressed in 293T cells (Figure 4B). NP bound to all importins with strongest affinities for importin-α1, -α5 and -α7. Again, importin-β1 showed lowest affinity to importin-α5 and -α7 as observed with monomeric PB2. The overall abundance of importin-α5 and -α7 compared to importin-β1 levels which are not observed with importin-α1, -α3 or –α4 might suggest additive functions for importin-α5 and -α7 beyond nuclear transport.
Our findings demonstrate that PB2 E627K substitution induces binding of importin-α1, -α5 and -α7 proteins to human-like vRNP complexes mediated by NP. Interestingly, importin-β1 binding properties were not changed despite increased importin-α1, -α5 and -α7 binding to PB2 627K human-like vRNPs. Thus, differential importin-α binding observed here with the PB2 627 signature does not correlate with differences in importin-β1 binding which is indicative of altered nuclear transport as previously shown for the PB2 701 adaptive position [4]. Furthermore, highest binding affinity of NP for importin-α1, -α5 and -α7 correlates with the recruitment of the same importin isoforms to human-like vRNPs. This underlines the bridging role of NP in vRNP/importin-α complex formation.
PB2 627K containing human-like vRNP/importin-α complexes did not result in increased importin-β1 binding which is indicative of improved nuclear import of vRNP components [4]. To further investigate, whether PB2 E627K substitution affects subcellular distribution of vRNP subunits, we analyzed PB2 and NP localization by immunofluorescence assays in unsilenced as well as importin-α silenced cells (Figure 5). vRNP complexes containing either PB2-627K-FLAG or PB2-627E-FLAG were reconstituted in 293T cells and specific antibodies were used to stain for PB2-FLAG and NP. Here, we focused on the importin-α1, -α3 and -α7 isoforms which significantly altered viral polymerase activity (Figure 1A and B). In unsilenced controls, both PB2 variants as well as NP were detected only in nuclear areas of the cell. No differences were observed in PB2 subcellular localization with either PB2 627K or PB2 627E (Figure 5A). In importin-α silenced cells, localization of PB2 and NP was not affected compared to unsilenced controls (Figure 5B–D).
Taken together, we show that adaptive mutation in PB2 E627K does not affect subcellular distribution of the vRNP subunits PB2 or NP. Further, importin-α1, -α3 and -α7 isoforms are not essentially required for nuclear accumulation of these vRNP subunits.
In order to assess whether the regulatory role of importins on polymerase activity affects virus growth, we have performed growth kinetics of recombinant viruses containing either PB2 627K (WSN-PB2627K) or PB2 627E (WSN-PB2627E) in the WSN (H1N1) virus background in importin-α silenced A549 human lung cells (Figure 6E). Silencing of importin-α1 and -α7 significantly decreased growth of WSN-PB2627K virus by 10- to 50-fold, both 72h (2% and 6%) and 96h p.i. (2% and 5%) compared to controls (Figure 6A and B). Silencing of importin-α3 had no significant effect on WSN-PB2627K replication. WSN-PB2627E growth was not affected in any of the importin silenced cells (Figure 6C and D).
These findings show that human-like PB2 627K but not avian-like PB2 627E virus growth depends on importin-α1 and -α7 in human lung cells. However, an inhibitory effect of importin-α3 on viral replication was not observed.
Next, we wanted to study whether the positive regulatory role of importin-α7 on human- but not avian-like polymerase activity is also reflected in vivo. It has been shown that a single amino acid substitution in PB2 E627K converts a non-lethal virus to a lethal virus in mice [9]. Therefore, we have determined the MLD50 of WSN-PB2627K and WSN-PB2627E viruses using serial virus dilutions (Table S1). We chose the minimum inoculation dose for each virus which guarantees 100% lethality in WT mice. This would allow the detection of potential differences in importin-α7−/− mice in case the in vitro data should also correlate in vivo. Therefore, we have infected wildtype and importin-α7−/− mice with 105 p.f.u. (∼20-fold LD50) of WSN-PB2627K or 5×106 p.f.u. (∼10-fold LD50) of WSN-PB2627E. Wildtype animals infected with WSN-PB2627K displayed significant weight loss and succumbed to infection within 7 days (Figure 7A and C). In contrast, importin-α7−/− mice lost less weight and 20% of the infected animals survived an otherwise 100% lethal infection. No significant differences in weight loss or survival were observed among importin-α7−/− and wildtype mice infected with WSN-PB2627E (Figure 7B and D). All mice died within 7 days. Reduced mortality of WSN-PB2627K infected importin-α7−/− mice correlated with 10-fold reduction in lung titers, reduced expression of virus antigens in the lung and inflammatory cells on day 6 p.i. but not on day 3 p.i. compared to WT animals (Figure 7E, Figure S7). This might suggest that initial replication deficiencies are detected at later time points upon viral clearance by the host similar to previous observations in importin-α7−/− mice [5]. In contrast, no significant differences in lung titres, virus antigen positive cells or lung pathology were observed in WSN-PB2627E infected WT and importin-α7−/− mice on day 3 and 6 p.i. (Figure 7F and Figure S7).
Furthermore, serial viral dilution experiments (Table S1 and Figure S8) confirmed the significant effect of importin-α7 on WSN-PB2627K but not WSN-PB2627E infection. Survival further increased in importin-α7−/− mice infected with 104 p.f.u. of WSN-PB2627K up to 60% compared to 70% lethality in WT mice. In contrast, no significant differences were observed upon infection with WSN-PB2627E between WT and importin-α7−/− mice.
In summary, these data demonstrate that human-like PB2 627K but not avian-like PB2 627E confers importin-α7 dependency in mice.
The influenza virus polymerase complex is active in the nucleus of the infected cell where cytoplasmically expressed PB1-PA dimers are imported by RanBP5 [21]–[23], a component of the non-classical import machinery which is independent of importin-α. PB2 appears to be imported separately into the nucleus by the classical importin-α/β1 mediated import pathway [4], [15]. Several mammalian influenza virus strains either having PB2 701N, 627K or other unknown adaptive positions in 2009 pandemic influenza virus strains displayed importin-α7 dependency in human lung cells [5]. PB2 D701N and NP N319K were sufficient to confer importin-α7 dependency in human cells. However, the role of other host adaptive positions, such as PB2 627K was not further investigated. PB2 627K can be functionally substituted by PB2 701N in terms of polymerase activity [24] and virus transmission [19], [24]. These findings suggested a mechanistically similar mode of action between PB2 627 and PB2 701. Therefore, we investigated whether the cellular importin-α proteins are also involved in PB2 E627K mediated host adaptation and pathogenicity as has been shown before for PB2 701N [4], [5].
Two opposing concepts have been proposed on the involvement of cellular factors concerning the PB2 627 signature position: (1) Mehle and Doudna hypothesized that an inhibitory factor exists which restricts avian-like polymerase activity in mammalian but not in avian cells [11], (2) Moncorgé and colleagues suggested that there is no inhibitory factor of avian polymerases but the absence or low expression of a positive factor is responsible for their low polymerase activity in human cells [12]. Indeed, very recently Bortz et al. have identified several cellular proteins which differentially regulate H5N1 polymerase activity in a PB2 627 dependent fashion [14]. Among them, the cellular DDX17 protein was described as a novel factor to facilitate efficient human-adapted PB2 627K H5N1 virus transcription and replication in human cells. Future investigation is needed to analyze the role of these host factors [14] within the network of polymerase interacting proteins [13] in host adaptation and pathogenicity in relevant animal models.
In this study, we could identify the importin-α isoforms as novel regulators of PB2 627K mediated human-like polymerase activity and pathogenicity. Human importin-α1 and -α7 act as positive regulators of PB2 627K mediated human-like polymerase activity while importin-α3 is a negative regulator independently of the viral polymerase origin. vRNP expression levels were not affected in importin-α silenced cells (Figure S1). Thus, it is unlikely that these effects are due to altered expression of vRNP subunits. We have performed all experiments in this study in the WSN (H1N1) background for consistency with previous studies on PB2 627 mediated polymerase activity and vRNP formation in human cells [11], [25]. Thus, our findings with importin-α1 and -α7 as positive regulators of human- but not avian-like polymerase activity are consistent with the model suggested by Moncorgé and colleagues since reduced avian-like polymerase activity correlated with reduced binding of these importin-α isoforms in human cells. However, this does not exclude the existence of another inhibitory factor in human cells since host adaptation is a multigenic process and importins are one of the factors involved in host adaptation and pathogenesis.
In contrast to PB2 D701N, PB2 E627K substitution did not affect importin-α binding affinity when expressed alone. This is consistent with previous reports using the C-terminal region of PB2 only [16]. The authors showed that PB2 D701N increases importin-α binding in general, consistent with previous studies [4], [5]. In contrast, PB2 E627K substitution did not alter importin-α binding properties [16] consistent with our data described here. However, single cargo proteins can interact with most of the importin-α isoforms while expression of multiple cargos can result in distinct importin-α specificities which is determined not only by the NLS but the whole protein context [20]. In case of a viral infection, multiple viral and cellular proteins compete for nuclear entry by the importin-α/β1 pathway. Limited availability of the importin-α isoforms is believed to be a restricting factor.
Therefore, we have further analyzed whether the expression of vRNPs affects binding to individual importin-α isoforms in a PB2 627 dependent manner. While PB2 627K alone had no effect on importin-α binding, it enhanced binding of importin-α1, -α5 and -α7 to vRNPs when expressed with all vRNP subunits. The enhanced binding of importin-α1 and -α7 to human- but not avian-like vRNPs correlated with their positive regulatory effect on human-like but not avian-like polymerase activity. However, reduced binding of importin-α1, -α5 and -α7 to avian-like vRNPs did not correlate with reduced precipitation of importin-β1 from vRNP/importin-α which is indicative for altered nuclear transport [4]. Consistently, subcellular localization of PB2 627K or NP was not affected unlike PB2 701N and NP 319K [4]. This is further confirmed by studies performed here and by others using cell fractionation assays where PB2 E627K did not affect subcellular distribution upon polymerase transfection or virus infection (Figure S4) [11].
According to the literature, cargo proteins are transported into the nucleus as cargo/importin-α/β1 ternary complexes [17], [18]. Therefore, one would expect similar importin-β1 proteins bound to PB2/importin-α complexes. However, it is still unclear how many importin-α isoforms are attached to ternary complexes to mediate efficient nuclear transport. The abundance of importin-α5 and -α7 compared to importin-β1 levels might either suggest that several importin-α proteins are associated with ternary complexes or, more likely that these isoforms have additional functions to those of nuclear transport. Indeed, it has been postulated before that importins can fulfill multiple functions. Importin-β-like proteins were shown to act as chaperons for exposed basic domains [26]. Very recently, importin-α/β1 proteins were suggested to play a role in cellular mRNA quality control [27]. The idea that importin-α isoforms might have additional functions than nuclear transport in the influenza virus life cycle was first put forward by Resa Infante et al. They proposed that host-dependent interaction of importin-α with PB2 is required for virus RNA replication itself [28]. Future investigation is needed to understand whether this function is affected by the PB2 627K signature position.
Another possibility how importin-α might regulate human-like polymerase activity would be indirectly by recruiting other cellular factors to their vRNPs by specific binding to importin-α1 and -α7. However, silencing both, importin-α1 and -α7 did not further suppress PB2 627K polymerase activity in an additive manner (Figure S2). This might suggest that importin-α1 and -α7 have similar functions and that other cellular factors are involved which are not specifically bound and transported by these isoforms. Further analyses are needed to identify importin-α specific and unspecific host cell factors within the polymerase interacting network and their contribution on host dependent polymerase activity and pathogenicity. Other groups have further postulated based on structural and biological data that importin-α proteins might play a role in polymerase assembly [15], [16], [29]. PB2 E627K was shown to modulate vRNP complex formation in mammalian cells [10] consistent with our findings here. Our studies suggest that enhanced binding of importin-α1, -α5 and -α7 to human-like vRNPs is mediated by NP. This is further supported by the fact that the same importins show highest affinity for NP when expressed alone, in the vRNP context or in infection (Figure S3). However, our studies cannot distinguish whether enhanced importin-α binding to human-like vRNPs is due to defects in assembly or a subsequent consequence that less efficiently NP is bound to the polymerase complex. Purified human-like vRNPs from importin-α silenced cells did not show altered PB2/NP ratios (Figure S5) suggesting that differential importin-α binding to human-like vRNPs is likely to present a consequence than a cause of defective vRNP assembly. However, we cannot exclude from these studies the role of importins in polymerase or vRNP formation. Clearly, future investigation is needed to distinguish between all these possibilities and whether these functions are regulated by host adaptive signatures. It would also be interesting to analyze whether the PB2 D701N substitution also affects transport independent functions mediated by importins since the functional substitution described for PB2 627K and 701N [19] might suggest alternative transport independent pathways which are affected by both mammalian signatures.
The differential role of importins was further confirmed with recombinant virus containing human-like PB2 627K which displayed dependency on importin-α1 and -α7 in human lung cells. Interestingly, importin dependency was lost with avian-like PB2 627E substitution. However, the general inhibitory activity of importin-α3 was restricted to vRNPs and was not observed in the viral context, since silencing of importin-α3 did not affect human- or avian-like virus growth kinetics in human lung cells. Specificity to individual importin-α isoforms is likely due to competition of multiple cellular and viral proteins for nuclear import [5], [20]. Upon infection, the restricting activity of importin-α3 for vRNPs might be overcome by another viral factor besides the polymerase complex. Previously, dependency on importin-α3 was mainly observed with HPAIV. However, the role of importin-α3 on polymerase activity of HPAIV was not studied [5]. Clearly, further studies are needed to identify the factor responsible for overcoming the restricting activity of importin-α3 observed with the mammalian influenza virus strain used in this study.
Remarkably, the regulatory role of importin-α7 observed in vitro could also be verified in vivo. Importin-α7−/− mice were less susceptible to human- but not avian-like virus infection (Table S1, Figure S8). Accordingly, systemic spread was mainly restricted to the lung in human-like influenza virus infected importin-α7−/− mice compared to WT mice where infection of the brain was observed (Figure S6 and S7). In contrast, no difference in survival, virus titres or lung pathology has been observed with avian-like virus in wildtype or importin-α7−/− mice (Table S1, Figure S6, S7 and S8). Surprisingly, viremia has been observed for human- and avian-like viruses (Figure S6). However, this did not necessarily lead to virus replication in the brain with the avian-like virus suggesting that PB2 627K is needed for efficient systemic spread. This further supports that PB2 627K and the importin-α7 gene are important for efficient systemic spread and extrapulmonary infection of human-like influenza viruses.
In summary, these findings strongly suggest that importin-α1 and -α7 isoforms play an important role in host adaptation of several mammalian viruses with either PB2 627K or PB2 701N adaptive positions [5]. However, the primary role of importin-α1 and -α7 in nuclear transport was not pivotal for PB2 E627K mediated host adaptation suggesting that importins are multifunctional proteins modulating viral polymerase activity by novel but yet unknown mechanisms. Further, our observations highlight the impact of importin-α isoforms in interspecies transmission of influenza viruses. Therefore, targeting especially importin-α7 may provide a strategy with therapeutic potential against human influenza viruses.
Animal experiments were performed according to the guidelines of the German animal protection law. All animal protocols were approved by the relevant German authority (Behörde für Stadtentwicklung und Umwelt Hamburg). Mice were humanely killed upon ≥25% weight loss according to the guidelines of the German animal protection law.
Importin-α7−/− mice [5], [30] and wildtype littermates in the C57BL/6J genetic background were bred and housed at the animal facility of the Heinrich-Pette-Institute, Leibniz Institute of Experimental Virology in Hamburg, Germany. Wildtype (n = 20) and importin-α7−/− (n = 20) mice were intranasally infected with 105 p.f.u. (∼20-fold MLD50) of WSN-PB2627K or 5×106 p.f.u. (∼10-fold MLD50) of WSN-PB2627E. Survival and weight loss was monitored for 14 days. Five animals were sacrificed on day 3 and 6 post infection (p.i.). Lungs and brains were removed and blood was obtained by heart-punctuation. Viral titers were determined by plaque assay.
MLD50 studies were performed using serial dilutions of WSN-PB2627K (103, 104 and 105 p.f.u.) or WSN-PB2627E (105, 106 and 5x106 p.f.u.) in wildtype and importin-α7−/− mice using additional mice (n = 5–10) per group and per dose.
Cell lines of human embryonic kidney cells (293T), human alveolar adenocarcinoma cells (A549) and chicken fibroblasts (DF1) were grown in DMEM (Dulbecco's modified Eagle's medium, PAA) supplemented with 10% fetal calf serum (PAA), 1% penicillin/streptomycin (PAA) and 1% L-Glutamine (PAA) at 5% CO2 and 37°C. Recombinant A/WSN/33 viruses WSN-PB2627K and WSN-PB2627E were rescued using the pHW2000 based 8-plasmid system as described previously [31].
All transfections were performed using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. Vector constructs used were pHW2000-(WSN-PB2-627K, WSN-PB1, WSN-PA, WSN-NP, WSN-HA, WSN-NA, WSN-M, WSN-NS) kindly provided by H.-D. Klenk (Institute of Virology, Marburg, Germany), pHW2000-(WSN-PB2-627K-FLAG, WSN-PB2-627E, WSN-PB2-627E-FLAG, NP-FLAG), pPol-I-NP-Luc-human [3], pPol-I-NP-Luc-chicken (kindly provided by M. Schwemmle, Institute of Virology, Freiburg, Germany), pRL-TK (PROMEGA), pcDNA-importin-α1-FLAG, pcDNA-importin-α3-FLAG, pcDNA-importin-α4-FLAG, pcDNA-importin-α5-FLAG and pcDNA-importin-α7-FLAG. The pcDNA-importin-α-FLAG constructs were described previously [5]. Constructs pHW2000-WSN-PB2-627K-FLAG and pHW2000-WSN-NP-FLAG constructs were generated by standard PCR techniques attaching a FLAG M2 tag to the C-terminus of the coding sequence. pHW2000-WSN-PB2-627E and pHW2000-WSN-PB2-627E-FLAG constructs were generated by site-directed mutagenesis.
Primary antibodies used for Western blot analysis, cell fractionation and immunofluorescence assays include mouse anti-FLAG (Sigma), rabbit anti-importin-α1 (Abcam), goat anti-importin-α3 (Abcam), goat anti-importin-α4 (Abcam), rabbit anti-importin-α5/α7 (kindly provided by E. Hartmann, Institute of Biology, Lübeck, Germany), mouse anti-importin-β1 (BD Transduction Laboratories), rabbit anti-GAPDH and rabbit anti-LSD1 (Cell signaling), mouse anti-PB2 (kindly provided by J. Ortín, CSIC, Madrid, Spain), rabbit anti-PA (kindly provided by G.G. Brownlee and E. Fodor, University of Oxford, United Kingdom), rabbit anti-FPV-serum [4], rabbit anti-NP (Abcam). Secondary HRP-conjugated antibodies used were anti-mouse-HRP, anti-rabbit-HRP and anti-goat-HRP (Sigma). For immunofluorescence, donkey anti-mouse-Cy3 and donkey anti-rabbit-Cy2 secondary antibodies were obtained from Jackson ImmunoResearch.
293T cells were transfected with siRNA designed against human importins (α1, α3, α4, α5 and α7) as described previously [5]. Allstars negative control siRNA (QIAGEN) was used as a control. 48h after silencing, cells were transfected with pHW2000 vector constructs encoding PB2-627K or PB2-627E, PB1, PA and NP with reporter constructs pPol-I-NP-Luc encoding firefly and for normalization pRL-TK (Promega) for renilla luciferase. Luciferase activity was measured 24 h after transfection according to the manufacturer's instructions. The vRNP reconstitution assay has been performed according to previously validated experimental settings described before [3]. Successful siRNA mediated silencing of human importin-α isoforms was confirmed using Western blot analysis.
All immunoprecipitations were performed using EZview Red ANTI-FLAG M2 affinity gel (Sigma) and eluted using a 3x FLAG peptide (Sigma) according to the manufacturer's instructions. As expression levels of PB2 627K or 627E can vary in the vRNP context [10], we have adjusted co-precipitated PB2 levels when expressed as vRNPs by serial dilution. Quantification of co-immunoprecipitation products was performed with the Bioimager Image Quant LAS 4000 at non-saturated levels. Relative amounts of co-immunoprecipitated products associated with PB2, vRNPs or trimeric polymerase complexes (PB1, PB2 and PA) containing PB2 627K were set to 100%.
Experimental setting was performed as described before [4]. Briefly, 293T cells were infected with an MOI of 2 and subcellular fractions were analyzed 6h p.i. Nuclear and cytoplasmic fractions were obtained using the NE-PER Kit (Pierce) according to the manufactures instructions. Protein concentration of each fraction was measured by BCA Protein Assay (Pierce). Subcellular distribution of viral and cellular proteins was analyzed by Western blot analysis. Equal amount of protein was loaded for each fraction. Fractionation purity was controlled using specific subcellular markers, such as GAPDH as a cytoplasmic protein and LSD1 as a nuclear protein.
293T cells were silenced for importin-α1, -α3 or -α7 using siRNA as described before [5]. Cells were co-transfected with plasmids encoding PB2-627K-FLAG or PB2-627E-FLAG, PB1, PA, NP as well as with pPol-I-NP-Luc 48h after siRNA transfection and seeded on glass cover slips for 24h. Cells were fixed with 3% paraformaldehyde in PBS for 10 min at room temperature and permeabilized with methanol for 5 min at −20°C. Permeabilized cells were blocked with 5% donkey serum (Abcam) in PBS for 1 h and stained with antibodies directed against PB2-FLAG and NP. Cellular DNA was stained with DRAQ5 (Cell signaling). All images were taken on a confocal laser scanning microscope (Zeiss 510 Meta CLSM) in multitrack mode with x63/1.4 oil Plan-Apochromat objective. Zeiss Confocal Microscopy Software (Release 3.29) was used.
A549 cells were silenced for importin-α1, -α3 or -α7 using siRNA as previously described [5] and infected with recombinant A/WSN/33 virus containing either PB2 627K or PB2 627E at MOI 0.001. Allstars negative control siRNA (QIAGEN) was used as a control. Supernatants were taken 72 hours and 96 hours post infection (p.i.). Virus titers were determined as plaque forming units (p.f.u.) by plaque assay as described before [5].
For histopathological examination, lungs of infected wildtype and importin-α7−/− mice were sectioned on day 6 p.i. and treated as described previously [32]. Viral antigens were stained using anti-FPV-serum and ZytoChem Plus HRP-DAB Kit Broad Spectrum (Zytomed) according to the manufacturer's instructions.
For statistical analysis of experimental data, mean values +/− standard deviation (SD) were calculated and p-values were obtained according to student's t-test for paired data. For comparison of animal survival rates Geham-Brelow-Wilcoxon test was used. Statistical significance was defined as p<0.05 (*p<0.05, **p<0.01, ***p<0.001).
|
10.1371/journal.pgen.1003314 | The Archipelago Ubiquitin Ligase Subunit Acts in Target Tissue to Restrict Tracheal Terminal Cell Branching and Hypoxic-Induced Gene Expression | The Drosophila melanogaster gene archipelago (ago) encodes the F-box/WD-repeat protein substrate specificity factor for an SCF (Skp/Cullin/F-box)-type polyubiquitin ligase that inhibits tumor-like growth by targeting proteins for degradation by the proteasome. The Ago protein is expressed widely in the fly embryo and larva and promotes degradation of pro-proliferative proteins in mitotically active cells. However the requirement for Ago in post-mitotic developmental processes remains largely unexplored. Here we show that Ago is an antagonist of the physiologic response to low oxygen (hypoxia). Reducing Ago activity in larval muscle cells elicits enhanced branching of nearby tracheal terminal cells in normoxia. This tracheogenic phenotype shows a genetic dependence on sima, which encodes the HIF-1α subunit of the hypoxia-inducible transcription factor dHIF and its target the FGF ligand branchless (bnl), and is enhanced by depletion of the Drosophila Von Hippel Lindau (dVHL) factor, which is a subunit of an oxygen-dependent ubiquitin ligase that degrades Sima/HIF-1α protein in metazoan cells. Genetic reduction of ago results in constitutive expression of some hypoxia-inducible genes in normoxia, increases the sensitivity of others to mild hypoxic stimulus, and enhances the ability of adult flies to recover from hypoxic stupor. As a molecular correlate to these genetic data, we find that Ago physically associates with Sima and restricts Sima levels in vivo. Collectively, these findings identify Ago as a required element of a circuit that suppresses the tracheogenic activity of larval muscle cells by antagonizing the Sima-mediated transcriptional response to hypoxia.
| Cells in multicellular animals must adapt to changing environmental conditions in order to ensure survival of the larger organism. One key challenge they face is fluctuation in the availability of dissolved oxygen. As cells get low on oxygen, they respond by turning on a program of gene expression that helps them survive. The key to this program is a protein, called HIF-1α in humans and Similar (or Sima) in the fruit fly Drosophila melanogaster, that is kept inactive in normoxia but is activated in hypoxia. The mechanisms responsible for this switch are not completely understood. In this study, we present genetic and molecular evidence that a component of the protein degradation machinery called Archipelago is required to keep Sima inactive in developing muscle cells and that genetically removing Archipelago makes these cells “think” they are hypoxic. This finding and the data that support it provide new insight into genetic circuits that cells use to control their response to changing oxygen levels and suggest that defects in oxygen homeostasis may contribute to cancerous disease states associated with loss of the human equivalent of Archipelago called Fbw7.
| Metabolically active tissues require an adequate supply of dioxygen (O2) for metabolic production of ATP by aerobic glycolysis and as a necessary substrate in a variety of enzymatic reactions (reviewed in [1]). Consequently, cells in metazoan organisms have evolved a conserved hypoxia-response mechanism that senses low O2 (or hypoxia) and modulates cellular metabolism and signaling in response to this environmental challenge. Activation of this adaptive mechanism results in changes in transcription that allow organisms to adapt to O2 conditions that might otherwise be incompatible with normal development and homeostasis (reviewed in [2]). In most metazoans, these changes include elevated expression of factors involved in oxygen-independent ATP production, increased expression of oxygen-carrying hemoglobin-like molecules and increased branching of O2-carrying tubular organs, the net effect of which is to reduce overall O2 demand and increase O2 delivery.
Molecular mechanisms through which changes in O2 concentration alter metabolism and drive increased tubular branching are conserved through the metazoan tree to include invertebrates like the fruit fly Drosophila melanogaster (reviewed in [3]). A key element of this mechanism is the hypoxia-inducible factor-1 [4]–[7] (HIF-1 or Drosophila HIF [dHIF] in flies), which is a heterodimeric transcription factor composed of an oxygen-regulated HIF–1α subunit and a constitutively expressed HIF–1β subunit. In Drosophila these subunits are respectively encoded by the similar (sima) [8], [9] and tango (tgo) [10]–[12] genes. The HIF-1/dHIF heterodimer is required for cellular adaptation to hypoxic conditions [4]–[7] and is regulated mainly at the level of HIF–1α stability [reviewed in 13]). In normoxic conditions, HIF–1α is hydroxylated at conserved proline residues by the 2-oxoglutarate/Fe(II)-dependent prolyl-4-hydroxylase family member HIF prolyl hydroxylase (HPH) [14], [15]. Prolyl-hydroxylation of HIF–1α facilitates binding with the von Hippel Lindau (VHL) E3-ubiquitin ligase subunit, and subsequent polyubiquitination and proteasome-dependent degradation of HIF-1α [16]–[20]. Drosophila Sima is controlled by a well-conserved version of this pathway involving the HPH homolog fatiga (fga), and the Drosophila VHL homolog, dVHL [14], [21]–[24]. Because the HPH enzymatic activity is dependent upon the availability of oxygen [14], [15], the HPH/VHL pathway effectively functions as a sensor of cellular oxygen levels, allowing HIF–1α/Sima stabilization only in hypoxic conditions and preventing HIF activity in normoxic cells [reviewed in 2]. Mutations in the VHL gene stabilize HIF-1α and are associated with a dominantly inherited hereditary cancer syndrome in humans that predisposes to a variety of malignant and benign tumors of the eye, brain, spinal cord, kidney, pancreas, and adrenal glands [25]. Excess HIF-1α can promote several important aspects of cancer biology, including the metabolic switch to anaerobic glycolysis characteristic of tumor cells [i.e. the Warburg effect; 26], neoangiogenesis, and increased tumor metastasis [reviewed in 13], [27], [28].
The invertebrate response to hypoxia mirrors key features of the mammalian hypoxic response [3], [29], [30]. Hypoxia stabilizes Sima and induces expression of genes that include homologs of mammalian HIF targets, such as lactate dehydrogenase (LDH) [31]. Hypoxic treatment of flies also produces physiological changes reminiscent of the mammalian hypoxic response [32], including altered metabolism and reduced oxygen consumption [33]–[36]. Adult Drosophila respond to hypoxia by entering into state of stupor characterized by low or undetectable neurological activity that allows them to tolerate extended periods of low oxygen [34], and recovery from this state is dependent upon genes necessary for survival in low-oxygen conditions [31]–[33], [35]. Hypoxia also induces a neoangiogenesis-like process in Drosophila involving increased branching of the tracheal system, an open network of interconnected, epithelial tubes that duct gases in and out of the animal [reviewed in 37]. Drosophila larvae reared in chronic hypoxia show increased branching of cells at the tip of each tracheal branch termed ‘terminal tip’ cells, whereas those raised in chronic hyperoxia show a reciprocal decrease in the extent of terminal branch elaboration [22], [38]. This increased larval tracheal branching in low O2 involves the FGF receptor homolog breathless (btl) [39] and the FGF ligand branchless (bnl) [40]: hypoxic exposure results in a sima-dependent increase in expression of btl in tracheal cells and bnl in peripheral oxygen-deficient tissues [22], [38]. Bnl then acts on tracheal terminal tip cells, which express Btl [41], [42], to induce fine tubular extensions that project toward Bnl-expressing cells. These terminal branches serve as the primary site of gas exchange between the tracheal system and internal tissues. When the oxygen demand is met, Bnl and Btl expression decreases, thereby limiting hypoxia-induced tracheal growth. This oxygen responsiveness allows for growth of tracheal terminal branches specifically to localized areas of hypoxia in order to shape the mature tracheal architecture and to increase oxygen-delivery capacity in hypoxic conditions.
In addition to the oxygen-dependent HPH/VHL pathway, mammalian HIF-1 is regulated by VHL-independent mechanisms that are incompletely understood [43], [44]. Recent studies have linked HIF–1α turnover to phosphorylation by the GSK3ß kinase and subsequent binding of the ubiquitin ligase subunit Fbw7 [45], [46], which is a sequence and functional ortholog of the Drosophila Archipelago (Ago) protein. Intriguingly Drosophila Ago binds and stimulates turnover of the Trachealess protein (Trh), which is a Sima/HIF-1α sequence homolog, in embryonic tracheal cells [47]. Genetic data show ago and dVHL also coregulate oxygen-sensitivity in the developing embryonic tracheal arbor [48].
In light of these connections, we have tested the requirement for ago in oxygen-sensitive stages of larval tracheal development and find evidence that ago is an antagonist of dHIF during the larval stage. Genetic manipulations that reduce ago function within post-mitotic larval muscle cells lead to a sima-dependent increase in the branching of nearby terminal cells. This phenotype is not suppressed by a trh allele that suppresses branch defects in ago mutant embryonic tracheal cells [47], but rather correlates with elevated expression of the Sima-induced gene bnl expression in larval muscle cells and genetic dependence on bnl. At an organismal level, reducing ago activity results in constitutive expression of some dHIF target genes in normoxia, increases the sensitivity of others to mild hypoxic stimulus, and allows adult flies to recover more rapidly from hypoxic stupor than normal flies. Significantly, non-cell autonomous effects of ago and dVHL alleles on terminal branching are synergistic, suggesting that the Ago and dVHL proteins co-regulate dHIF. Consistent with this, Ago protein can be found in a complex with Sima in larval extracts and loss of Ago elevates Sima levels in peripheral tissues. Collectively these findings define an important role for Ago as a required antagonist of the Sima-dependent hypoxic response during the larval stage of Drosophila development.
Heterozygosity for a null allele of ago sensitizes the Drosophila embryonic tracheal system to mild hypoxia [48]. To determine whether ago is also involved in hypoxia responsiveness in the subsequent larval stage, it was necessary to generate an allele of ago that allowed development beyond the late embryonic lethality associated with ago null alleles [49]. This was achieved by transposase-mediated imprecise excision of EP(3)1135, a P-element located 16 base pairs (bp) upstream of the ago genomic locus (Bloomington Drosophila Stock Center [BDSC]) that behaves genetically as a weak ago hypomorph. Excision of EP(3)1135 produced a 603 bp deletion removing the first exon of the ago-RC transcript (Figure 1A–1B) that was designated agoΔ3–7. The effect of agoΔ3–7 on patterns of ago transcription was determined by quantitative real-time PCR (qRT-PCR). Of the three predicted ago transcripts (ago-RA, ago-RB, and ago-RC) only RA and RC are detected in whole larvae (Figure 1C). Consistent with the location of the deletion in the agoΔ3–7 allele, the ago-RC transcript is specifically absent in agoΔ3–7 mutant larvae while expression of the ago-RA transcript is unaffected. Notably, the ago-RA and RC transcripts display inverse expression patterns: ago-RA is approximately 3-fold more abundant than ago-RC in imaginal discs and larval brain and ventral nerve cord, but ago-RC is 3-fold more abundant than ago-RA in filleted larval body wall preparations (Figure 1D). The agoΔ3–7 allele is thus a tissue- and transcript-specific allele that primarily reduces ago expression in peripheral tissues such as body wall muscle.
Approximately 49% of agoΔ3–7 homozygotes or trans-heterozygotes in combination with the null alleles ago1 and ago3 die as pupae (Table 1) and the remainder die as late 2nd and 3rd instar larvae (data not shown). Those that live to late 3rd instar show tracheal phenotypes (Figure 2 and Table 2). The most prevalent of these phenotypes is an approximate doubling of the number of cytoplasmic branches elaborated from multiple subtypes of terminal cells, including those found along the lateral trunk that serve to oxygenate the ventrolateral body wall muscles: LH cell terminal branching increases from 20.4±0.64 branches (n = 33) in control larvae to 39.5±1.59 branches (n = 34) in agoΔ3–7/1 larvae (p = 2.6×10−16) (Figure 2A–2C and Table 2), and LG cell terminal branching increases from 19.6±0.54 branches (n = 33) in control larvae to 36.8±1.94 branches (n = 31) in agoΔ3–7/1 larvae (p = 9.5×10−13) (Figure 2C). Notably, the magnitude of these increases in terminal cell branch number is similar to that seen in larvae grown in hypoxic conditions [22], [38]. Loss of ago function also causes additional tracheal branch phenotypes in approximately 25% of larvae, including the appearance of terminal branch tangles (Figure 2D) and the development of ‘ringlet’-shaped ganglionic branches (Figure 2F), which resemble phenotypes seen in hypoxic larvae or those in which Sima is activated by genetic disruption of the Fga/dVHL regulatory pathway [22]. Given the transcript- and tissue-specific nature of the agoΔ3–7 allele, these tracheal phenotypes support the hypothesis that ago has a non-autonomous role in in restricting terminal branching.
Although the agoΔ3–7/1 larval phenotypes are reminiscent of hypoxia-induced tracheal growth, they do not exclude the possibility that an earlier developmental requirement for ago (e.g. in the embryo) affects later branching events in the larva. To test the temporal requirement for ago in regulating tracheal terminal branching patterns, a dominant negative ago transgene (UAS-agoΔF) [47], [50] was combined with the hs-Gal4 driver to produce animals in which ago activity could be antagonized at later developmental stages by application of a heat-shock. Whereas control and hs>agoΔF larvae show similar LH cell branch number prior to transgene induction (22.2±0.89 branches [n = 27] vs 21.7±0.69 branches [n = 24]), administration of a transient heat-shock to hs>agoΔF larvae is sufficient to drive an increase in terminal branching throughout the tracheal system (effects on LG and LH cells quantified in Figure 3A and 3B). LH cell branching is increased 24 hrs post heat-shock in hs>agoΔF larvae (40.2±1.48 branches [n = 24]; (p = 3.0×10−14 relative to no heat-shock) but remains unchanged in control larvae (22.6±0.67 branches [n = 24]). Thus animals that complete embryonic and early larval development with wt ago activity can be induced to undergo excess branching by transient expression of an ago dominant-negative allele.
Excess terminal branch phenotypes in hs>agoΔF and agoΔ3–7 animals may reflect a requirement for ago in either tracheal or non-tracheal cell types. To directly test whether ago activity is required in non-tracheal tissue to restrict branching, the agoΔF transgene was driven with the 5053A-Gal4 driver (5053A>agoΔF), which is expressed specifically in ventrolateral body wall muscle 12 (VLM12) and has been used to study non-cell autonomous tracheogenic activity of the Btl/Bnl pathway [38]. The VLM12 muscle expresses endogenous, nuclear Ago protein (Figure 4C–4D) and is normally tracheated by the LF and LH cells (Figure 4A). The 5053A>agoΔF genotype approximately doubles the number of LF and LH tracheal branches that terminate on VLM12 (Figure 4B) relative to either the adjacent muscle (VLM13) or to control larvae expressing a nuclear-localized GFP (nlsGFP) (5.11±0.16 branches [n = 54] in control vs 9.54±0.29 branches [n = 50] in agoΔF, p = 4.67×10−24) (Table 3). This degree of excess branching produced by muscle-specific expression of the agoΔF transgene is similar to that produced by organism-wide depletion of the Ago-RC isoform with the agoΔ3–7 genomic allele. These combined genetic data provide evidence that Ago is required within larval body wall muscle cells to restrict the post-embryonic branching of nearby tracheal terminal cells.
ago mutations lead to tissue-specific activation of factors normally degraded by the SCF-Ago ubiquitin ligase, including the proliferative proteins CycE and dMyc in larval imaginal discs [49], [50] and the transcription factor Trachealess in tracheal cells [47]. Although the expression patterns of these proteins in body wall muscle are not well defined, we wished to test whether ectopic expression of CycE, dMyc, Trh, or the SCF-Fbw7 target Notch [reviewed in 51] was even capable of conferring a non-cell autonomous tracheogenic activity to VLM12. To this end, each of these factors was individually overexpressed using the 5053A-Gal4 driver (Table 3). Muscle-specific expression of trh or Notch failed to stimulate excess terminal branch growth. The inability of trh to affect tracheal recruitment to VLM12 contrasts with its ability to phenocopy ago mutant phenotypes in the embryonic trachea [47] and further suggests that the ago larval tracheal role is from separable from its embryonic role. Muscle-specific expression of dMyc also had no effect on the degree of terminal cell branching, despite a 28.5% increase in the 2-dimensional size of the VLM12 muscle (Table 4). Notably, increased tracheation of VLM12 driven by agoΔF occurs without an increase in the size of the VLM12 muscle, which is consistent with no role for post-mitotic growth in this phenotype (Table 4). 5053A>cycE does increase terminal branch number, although to a lesser degree than agoΔF. However, CycE protein levels are not obviously affected by expression of agoΔF (Figure S1), suggesting that deregulated CycE is an unlikely cause of the non-autonomous effect of ago alleles on terminal cell branching.
The similarity of ago mutant terminal branching phenotypes to those induced by hypoxia suggests that ago may antagonize the dHIF pathway. To test the genetic relationship between ago and sima in larval tracheal branching, the sima07607 loss-of-function allele [23] was introduced into the 5053A>agoΔF and agoΔ3–7/1 genetic backgrounds. Heterozygosity for sima (i.e. sima07607/+) dominantly suppressed the agoΔF VLM12 phenotype (Table 3) by decreasing terminal branch number from 9.54±0.29 (n = 50) to 6.34±0.27 branches (n = 53, p = 4.36×10−12), and also suppressed the excess and overlapping terminal branching seen in agoΔ3–7/1 larvae (Table 2 and Figure 5A–5B; white arrow in 5A indicates a ringlet-shaped ganglionic branch) from 39.5±1.59 (n = 34) to 29.0±1.48 branches per LH cell (n = 34, p = 7.45×10−6). In addition, the sima07607 allele dominantly delayed the lethal phase of both agoΔ3–7 homozygotes and agoΔ3–7/1 or agoΔ3–7/3 trans-heterozygotes (Table 1). Reciprocally, ectopic expression of sima in the VLM12 muscle (5053A>sima) increased tracheal recruitment in normoxic conditions (Table 3). Muscle cells are thus distinct from ectodermal cells, which do not recruit branching following overexpression of sima [22].
To more directly assess dHIF activity in ago mutant animals, the transcription of the Drosophila LDH gene (dLDH) was measured in the body wall muscle of agoΔ3–7 and control larvae. LDH is a well-validated HIF target in vertebrates and invertebrates, and HIF-responsive elements from the LDH promoter have been used as the basis of HIF activity reporters in many different systems including Drosophila [e.g. 24]. This analysis showed a 27.3-fold increase in dLDH transcription in ago mutant larval body wall muscle preparations but no equivalent upregulation in steady-state levels of the sima mRNA (Figure 5C). Sima-driven expression of the FGF ligand bnl is a key element of the hypoxic response among non-tracheal cells [22], [38], [52]. The bnlP1 loss-of-function allele dominantly suppressed both the 5053A>agoΔF VLM12 phenotype (Table 3), from 9.54±0.29 (n = 50) to 6.54±0.28 branches (n = 54, p = 5.99×10−11) and the agoΔ3–7/1 excess branching phenotype (Table 2), from 39.5±1.59 (n = 34) to 28.4±1.80 branches per LH cell (n = 29, p = 1.75×10−5). In parallel, qRT-PCR detected an ∼50% upregulation of bnl transcription in body wall muscle of agoΔ3–7 larvae relative to control muscle (Figure 5C). Previous studies using a genomic duplication of the bnl locus have demonstrated that a similar 50% increase in bnl gene-dosage is sufficient to elicit excess tracheal terminal cell branching [38]. Thus reduced ago function in body wall muscle is associated with ectopic expression of the dHIF target dLDH, increased levels of the bnl mRNA, and a genetic dependence on sima and bnl.
The data above suggests that ago alleles might exhibit functional interactions with components of the Fga/dVHL pathway, which controls Sima stability and activity in vivo [21]–[23], [53], [54]. A previously characterized dVHL RNAi knockdown transgene (dVHLi) [48]) was used with the 5053A-Gal4 driver to reduce dVHL expression in VLM12. Consistent with the role of dVHL upstream of sima, the 5053A>dVHLi genotype showed an increase in terminal branching relative to a non-specific RNAi control (Figure 6A, and Table 3) (5.28±0.20 branches [n = 40] in Adf1i control vs 7.48±0.21 branches [n = 89] in dVHLi, p = 1.27×10−9, Figure 6). The dVHLi and agoΔF transgenes were then co-expressed with 5053A-Gal4 to determine their ability to enhance VLM12 tracheogenic activity (Figure 6C–6D). The 5053A>agoΔF,VHLi compound genotype shows a synergistic increase in the number of branches that terminate on VLM12 (Table 3), but also leads to a phenotype not seen in either individual genotype: whereas expression of agoΔF or dVHLi individually increase terminal branching of LF and LH onto VLM12, the agoΔF+dVHLi combination also recruits ectopic branches from the LG lateral terminal cell (as seen in the two different focal planes of a single agoΔF+dVHLi-expressing VLM12 muscle; Figure 6C–6D) which normally bypasses VLM12. This ectopic LG recruitment phenotype occurs in approximately 10% of agoΔF+dVHLi VLM12 muscles and is also observed upon 5053A-Gal4 driven expression of bnl [38] or sima (data not shown). Thus dVHL and ago are individually required to restrict the ability of muscle cells to recruit new branch growth, and combined reduction of ago and dVHL activity leads to increased tracheogenic signals emanating from body wall muscle.
To further define the relationship between ago and dVHL in terminal branching, transgenes expressing each factor were tested for rescue of VLM12-branching phenotypes produced by reducing the function of the other (Table 3). Expression of wild type dVHL led to a 66% suppression of the agoΔF branching phenotype (p = 6.55×10−12); reciprocally, over-expression of wild type ago showed a 54% suppression of the dVHLi branching phenotype (p = 2.73×10−4). Thus, each gene can to some degree ameliorate non-autonomous branching phenotypes produced by loss of the other in the VLM12 segment.
In view of the genetic and molecular links between ago, sima, dLDH, and dVHL, the organism-wide transcriptional response to hypoxia was examined in ago mutant animals. Drosophila respond to varying degrees of hypoxia by driving transcription of distinct sets of target genes at differing oxygen concentrations, including those involved in metabolic adaptation and survival in low oxygen [31], [32]. A subset of hypoxia-inducible genes was selected for this analysis based on their differential transcription in hypoxic adult Drosophila [31] and predicted links to known mechanisms of the hypoxic response. These included dLDH, which plays a role in the metabolic switch to high flux glycolysis [reviewed in 55], [56], lysyl oxidase (lox), a HIF target in mammalian cells that plays a role in hypoxia-induced changes in cell adhesion [57], [58] and vascular remodeling [59], and dHIG1 (CG11825), the Drosophila homolog of Hypoxia Induced Gene-1 (HIG1), which is induced by HIF and promotes cell survival [60]. qRT-PCR analysis was carried out for each of these genes under conditions of decreasing environmental oxygen (21%, 5%, 0.5%) in whole control larvae or whole agoΔ3–7 larvae (Figure 7A–7B). We find that each of these genes is differentially induced in hypoxia in a manner consistent with findings in adult Drosophila [31] and can be ectopically induced by the agoΔ3–7 allele. dLDH is minimally transcribed in normoxic control larvae, and with progressively higher transcription as the oxygen level falls (1.6 and 2.7-fold increases in 5% and 0.5% O2 respectively, Figure 7A), confirming that dLDH transcription increases with increasing dHIF activity. In agoΔ3–7 homozygous animals, dLDH expression is increased 8.1-fold in whole normoxic larvae (this lower fold induction in the whole larva relative to the ∼27-fold enrichment seen in dissected body wall muscle in Figure 5C is presumably a reflection of the tissue-specific nature of the agoΔ3–7 allele), and is increased approximately 14-fold in agoΔ3–7/1 larvae relative to control larvae at both 5% and 0.5% O2 (Figure 7B, top panel). Thus ago restricts dLDH expression activity across a broad range of oxygen concentrations. The lox gene is normally only up-regulated in whole control larvae by strong hypoxia (4.4-fold induction at 0.5% O2; Figure 7B). The agoΔ3–7 allele leads to a 2.2-fold increase in lox transcription in normoxia, and lox transcription reaches near maximal levels at 5% O2; the 3.4-fold induction seen in ago mutants in 5% O2 is not significantly different from that seen in control larvae at 0.5% O2 (Figure 7B, middle panel). This pattern suggests that the lox promoter is induced by levels of dHIF activity achieved in moderate hypoxic conditions, and that this threshold is more easily reached in ago mutants. The dHIG1 gene displays a more exaggerated version of the lox response pattern: dHIG1 mRNA levels are only induced strongly (19.9-fold) in whole control larvae by 0.5% O2 (Figure 7A); the agoΔ3–7 allele is not sufficient to drive ectopic dHIG1 transcription in normoxic conditions but it is sufficient to sensitize the dHIG1 promoter to reduced O2 levels such that maximal dHIG1 expression is now achieved at a ten-fold higher O2 concentration than normal (Figure 7B, bottom panel). In addition to dLDH, lox, and dHIG1, three other genes also induced by hypoxia, hairy, amy-p and thor genes [31], are also moderately up-regulated in normoxic agoΔ3–7 mutant larvae (Table S1). Reducing ago activity is thus sufficient to alter the threshold required to drive expression of multiple hypoxia-inducible genes.
Adult Drosophila respond to prolonged periods of oxygen deprivation by entering into a state of hypoxic stupor characterized by inactivity and reduced oxygen consumption [34]. Many mutations have been identified that slow this hypoxic recovery [32], [33], [35], but few mutations have been described that enhance it. Under our standard laboratory conditions, control adult flies enter stupor after approximately fifteen to twenty minutes in a 0.5% O2 environment and remain unconscious until re-oxygenation. We assayed control +/+ adults, agoΔ3–7/+ adults, and adults trans-heterozygous for the agoΔ3–7 allele and the ago hypomorphic allele EP(3)1135 (BDSC) for recovery time following acute hypoxia (1 hour at 0.5% O2) (Figure 7C). agoΔ3–7/EP(3)1135 flies display no obvious developmental phenotypes and enter into hypoxic stupor at the same rate as control flies (data not shown); however, they recover significantly faster than either control +/+ or agoΔ3–7/+ adults. Linear regression analysis indicates that the time for 50% recovery is reduced from 4.5±0.75 minutes in control +/+ flies, to 1.4±0.16 minutes in agoΔ3–7/EP(3)1135 flies (p = 0.0015). The agoΔ3–7/EP(3)1135 population also reaches 100% recovery after 10 minutes of re-oxygenation, whereas neither the control +/+ or agoΔ3–7/+ populations achieved 100% recovery by the end of the 15 minute measurement period (data not shown). Thus, the genetic evidence of a role for ago as a regulator of dHIF-regulated branching in the larval tracheal arbor is paralleled at the organismal level by an enhanced transcriptional sensitivity to hypoxia and an increased ability of flies to recover from a transient hypoxic challenge.
To test the molecular relationship between Sima and Ago, Sima levels were assessed in two ways: by immunoflourescent staining of VLM12 muscles expressing the UAS-agoΔF transgene and by Western blotting of lysates of agoΔ3–7 larvae (Figure 8). Fluorescence microscopy confirms that a previously described anti-Sima antibody [24] detects high levels of transgenically expressed Sima in the VLM12 nuclei of 5053A>sima muscles, and that endogenous Sima is not readily detectable by this method of analysis in the nuclei of adjacent non-transgenic muscles (Figure 8A). Following expression of the agoΔF dominant-negative transgene (5053A>agoΔF), a fraction of VLM12 nuclei accumulate anti-Sima reactive epitopes (see arrows, Figure 8B). This same anti-Sima antibody detects elevated levels of an ∼110 kD molecular weight band in agoΔ3–7 filleted 3rd instar pelts relative to wt control pelts (Figure 8C). This ∼110 kD band is absent in lysates of sima07607 larvae (Figure 8D, lane 1 vs. 2), and is specifically enriched in precipitates of an anti-Ago polyclonal antibody from lysates of hypoxic larvae (Figure 8D, lane 5). Collectively, these molecular data indicate that Ago can associate with Sima in larval lysates, and that Ago limits Sima levels in developing tissues.
The selective stabilization of the Sima/HIF–1α transcription factor in hypoxia plays a key role in the response of metazoan organisms to low oxygen concentrations by its ability to induce a program of hypoxia-specific gene expression [reviewed in 2]. Evidence suggests that in Drosophila, Sima plays a dual role in the post-mitotic growth of tracheal terminal branch cells toward hypoxic peripheral tissues by acting within both the ‘signaling’ hypoxic peripheral cells and in the ‘responsive’ terminal tips cells [reviewed in 37].
Our data implicate the Ago-SCF ubiquitin ligase as a required regulator of Sima during hypoxia sensing in peripheral cells, but do not rule out an additional role for Ago within tracheal terminal tip cells which contributes to their ectopic branching in ago mutant larvae (see below). Phenotypes produced by muscle-specific expression of an ago dominant-negative allele, or by a genomic allele that specifically affects ago expression in peripheral tissues, are phenocopied by overexpression of Sima (this study) or the FGF homolog Bnl [38]. These non-cell autonomous effects of ago alleles on terminal branching are accompanied by a strong induction in peripheral tissues of the dHIF target dLDH, and can be dominantly suppressed by an allele of sima. ago alleles induce expression of a set of dHIF-inducible hypoxia-response genes in normoxia that includes dLDH, and this is paralleled at the organismal level by an enhanced ability of ago mutant flies to the recover from hypoxic stupor. ago alleles are thus among the first genetic alterations shown to enhance the recovery of adult Drosophila from hypoxic exposure. Within larval muscle, ago appears to inhibit sima in parallel to dVHL, which targets the Sima/HIF–1α protein for constitutive degradation in normoxia [reviewed in 3]. Consistent with this, we find evidence that Ago can associate with Sima and limits its levels in vivo. Collectively these data significantly expand the known role of Ago in organism development by demonstrating that it is required in an apparently novel pathway that collaborates with dVHL to inhibit Sima-regulated hypoxic gene expression in peripheral tissues.
Though the work presented here focuses on the ‘tracheo-attractant’ effects of reducing ago expression in body wall muscle, this may be just one manifestation of roles Ago plays in controlling hypoxia-regulated gene expression. Indeed, reducing ago function has a quantitatively stronger effect on terminal branching than a genomic duplication of the bnl locus [38], suggesting either that ago also act within tracheal cells to limit branching [as in 47], [48] or that a larger set of dHIF target genes contribute to the effect. Consistent with this latter hypothesis, normoxic ago mutant larvae display ectopic induction of hypoxia-responsive metabolic genes such as dLDH, lox, hairy, amy-p and thor. Based on this profile, it appears that ago mutant larvae reared in normoxia elevate expression of bnl but also engage a metabolic switch to high-flux glycolysis that is characteristic of hypoxic cells [32], [33], [35], [36], [61]. Future studies will be required to assess the full effect of these transcriptional changes on the behavior of terminal tracheal cells and the tissues into which they project.
In wild type animals, the transcriptional response of cells to hypoxia is graded such that different target genes are induced across a range of environmental O2 concentrations [31]. In ago mutants, this differential induction is largely abolished such that expression of genes such as lox and dHIG1 is virtually indistinguishable at 5% and 0.5% O2. Thus, ago appears to be required both for inhibition of hypoxia-inducible genes in normoxia and for the graded expression of hypoxia-inducible genes under variable levels of oxygen deprivation. We hypothesize that this graded sensitivity is normally a product of the interaction between the Ago and Fga/dVHL regulatory mechanisms. The HPH/VHL pathway has been demonstrated to act in a graded manner, such that it degrades HIF–1α efficiently in normoxia, but is progressively less efficient as the oxygen concentration drops [62]. This leads to a gradient of HIF activity that is presumably required for the differential induction of target genes. We hypothesize that ago acts in parallel to dVHL to dampen Sima/HIF-1 activity across a range O2 concentrations, and that Ago may function as a dHIF regulatory mechanism at very low O2 concentrations in which the HPH/dVHL pathway is hypothesized to be inactive [62]. Thus, the absence of Ago allows a mild hypoxic stimulus (∼e.g. 5% O2) to be translated into levels of dHIF-dependent gene expression that would normally only result from much stronger hypoxic exposure. The data presented here support this prediction, with the end result that the transcriptional response profile of hypoxia-response genes in ago mutant larvae is shifted toward induction by more mild stimuli.
The molecular mechanism(s) underlying the genetic relationship between ago and sima in tracheal branching appears to involve a physical association of their encoded proteins that modulates Sima levels. Given that Ago is a ubiquitin ligase specificity factor, these data are consistent with a model in which Ago supports Sima polyubiquitination and turnover. Recent studies have identified the human Ago ortholog Fbw7 as a HIF–1α interacting factor and have proposed that Fbw7 promotes HIF-1α turnover following GSK3ß phosphorylation in cultured cells [45], [46]. Phenotypic predictions made by this molecular model appear to be confirmed by the ago terminal cell branching phenotypes documented here. Intriguingly, RNAi depletion of GSK3ß/shaggy or a proteasome subunit in VLM12 also elevates the number of tracheal branches that terminate on this muscle (Table S2). However Fbw7 is implicated in the proteolytic destruction of two transcription factors, the Notch intracellular domain (NICD) [reviewed in 51] and sterol-regulatory enhancer binding protein (SREBP) [63], that indirectly modulate HIF-dependent hypoxic gene expression in eukaryotic cells [64], [65]. Ago could thus theoretically influence hypoxic gene expression via these paths as well. Future biochemical studies will be required to clarify the full range of Ago molecular targets that contribute to its role in hypoxic gene expression.
The well-studied anti-proliferative role of ago is conserved in its human ortholog Fbw7, which is mutationally inactivated in a wide spectrum of primary human cancers [reviewed in 51]. Some cancer cells engage a program of gene expression that supports a switch to high-flux glycolysis (a phenomenon termed ‘Warburg effect’ [26]) and are more resistant to transient hypoxia than normal cells [reviewed in 1]. Both of these properties can now, to some degree, also be associated with ago loss in Drosophila. In view of the functional conservation between SCF-Ago and SCF-Fbw7 in degradation of shared oncogenic targets [49], [50] and the proposed role of Ago/Fbw7 in Sima HIF-1α turnover [this study and 45], [46], our data raise the interesting possibility that sensitization to mild hypoxia may be a feature of Fbw7 mutations in vertebrates as well. If so, then tumor suppressive properties of Fbw7 may derive in part from its established anti-proliferative role and in part due to modulation of HIF-regulated angiogenic and metabolic pathways.
The FRT80B and w1118 strains were used as wild type controls. The ago1 and ago3 alleles have been previously described [49]. The agoΔ3–7 allele was identified as an imprecise excision of the agoEP(3)1135 transposon. Alleles used in this study: bnlP1, sima07607, agoEP(3)1135 (all from the Bloomington Drosophila Stock Center) and 1-eve-1 [66]. The following transgenes were also used: UAS-agoΔF and UAS-ago [47]; UAS-CycE, UAS-dMyc, UAS-N, UAS-nlsGFP (all from the Bloomington Drosophila Stock Center), UAS-trh [67], UAS-sima [24], UAS-dVHL [21], UAS-Adf1RNAi and UAS-sggRNAi (Vienna Drosophila RNAi Center), UAS-dVHLRNAi [48], hs-Gal4, and 5053A-Gal4 (from the Bloomington Drosophila Stock Center). Statistical comparisons were made using Student's t-Test (Microsoft Excel) with the indicated significance levels.
Hypoxia treatments were performed in a sealed Modular incubator chamber (Billups-Rothenberg Inc., Del Mar, CA) with separate gas intake and exhaust openings. Internal O2 concentration was measured with an electronic O2 sensor (OX-01, RKI Instruments, Inc., Union City, CA). To assay recovery from hypoxia, 5–7 day old adult flies were put into plain glass tubes in groups of 9–15. The flies were then placed into the hypoxia chamber at 0.5% O2 for one hour and then removed to normoxia. Following hypoxic treatment, >99% of the flies (178 of 179) had fallen into hypoxic stupor. Recovery time was defined as the time required for each individual fly to resume walking following re-oxygenation.
Total RNA was isolated from dissected third instar larval body wall muscles. cDNA was reverse-transcribed using random hexamer primers (Invitrogen) with Superscript II Reverse Transcriptase (Invitrogen). dVHL and ß-tubulin transcripts were then amplified with gene-specific primers. For quantification of mRNA levels, total RNA was isolated from whole third instar larvae or dissected larval tissues and reverse transcribed as described above. Levels of Arp87c, ago-RA, -RB and -RC, dLDH, sima, bnl, lox, hairy, dHIG1, thor and amy-p were then assayed with gene-specific primers using the SYBR green method of quantitative real-time PCR on a Roche LightCycler 480 machine. Transcript abundance was normalized to levels of Arp87c as in [31].
To image the larval tracheal system, third instar larvae were dissected in cold PBS and fixed in 4% paraformaldehyde. Air-filled tracheal branches were imaged using bright-field microscopy. and assembled using Photomerge (Adobe Photoshop CS).
Third instar larvae were dissected in cold PBS, fixed in 4% paraformaldehyde and incubated with guinea pig anti-CycE (1∶500) or rabbit anti-Sima (1∶1000). Secondary antibodies (anti-guinea pig conjugated to Cy3 or anti-rabbit conjugated Cy5) were used as recommended (Jackson ImmunoResearch). To assess Sima protein levels in third instar larvae, larval pelt extracts were prepared in sample buffer and resolved on 7.5% SDS-PAGE prior to Western blotting with rabbit anti-Sima (1∶1000) [24] and developed with anti-rabbit HRP (1∶1000; Jackson ImmunoResearch). Whole larval extracts were immunoprecipitated with guinea pig anti-Ago polyclonal sera (1∶1000) [47] prior to immunoblotting with anti-Sima antibody.
|
10.1371/journal.pntd.0000240 | Altered Patterns of Gene Expression Underlying the Enhanced Immunogenicity of Radiation-Attenuated Schistosomes | Schistosome cercariae only elicit high levels of protective immunity against a challenge infection if they are optimally attenuated by exposure to ionising radiation that truncates their migration in the lungs. However, the underlying molecular mechanisms responsible for the altered phenotype of the irradiated parasite that primes for protection have yet to be identified.
We have used a custom microarray comprising probes derived from lung-stage parasites to compare patterns of gene expression in schistosomula derived from normal and irradiated cercariae. These were transformed in vitro and cultured for four, seven, and ten days to correspond in development to the priming parasites, before RNA extraction. At these late times after the radiation insult, transcript suppression was the principal feature of the irradiated larvae. Individual gene analysis revealed that only seven were significantly down-regulated in the irradiated versus normal larvae at the three time-points; notably, four of the protein products are present in the tegument or associated with its membranes, perhaps indicating a perturbed function. Grouping of transcripts using Gene Ontology (GO) and subsequent Gene Set Enrichment Analysis (GSEA) proved more informative in teasing out subtle differences. Deficiencies in signalling pathways involving G-protein–coupled receptors suggest the parasite is less able to sense its environment. Reduction of cytoskeleton transcripts could indicate compromised structure which, coupled with a paucity of neuroreceptor transcripts, may mean the parasite is also unable to respond correctly to external stimuli.
The transcriptional differences observed are concordant with the known extended transit of attenuated parasites through skin-draining lymph nodes and the lungs: prolonged priming of the immune system by the parasite, rather than over-expression of novel antigens, could thus explain the efficacy of the irradiated vaccine.
| Schistosoma mansoni is a blood-dwelling parasitic worm that causes schistosomiasis in humans throughout Africa and parts of South America. A vaccine would enhance attempts to control and eradicate the disease that currently relies on treatment with a single drug. Although a manufactured vaccine has yet to generate high levels of protection, this can be achieved with infective parasite larvae that have been disabled by exposure to radiation. How these weakened parasites are able to induce protective immunity when normal parasites do not, is the question addressed by our experiments. We have used a technique of gene expression profiling to compare the patterns in normal and disabled parasites, over the period when they would trigger an immune response in the host. We found that only a handful of genes were differentially expressed, all of them diminished in the disabled parasite. However, a more sensitive technique to examine groups of genes revealed that those involved in nervous system and muscle function were depressed in the disabled parasites. We suggest that reduced mobility of these larvae permits them longer contact with the immune system, thus enabling a strong protective immune response to develop.
| The radiation-attenuated schistosome (RA) vaccine remains the most effective way of inducing high levels of protective immunity against Schistosoma mansoni in rodent and primate hosts (reviewed by Coulson) [1]. However, an effective recombinant vaccine based upon it, for use in humans, has thus far proved elusive [2]. Few differences have been reported between irradiated and normal larvae apart from an altered morphological phenotype at the lung stage of development [3] that produced subtle differences in motility. This accorded with a key feature of the vaccine that attenuated larvae must undergo a truncated migration, as far as the lungs, to prime the immune system [4]. Furthermore, extensive parasite tracking [5] and immunological investigations [6] have revealed the lung schistosomulum to be the principal target of immune effector responses in the murine host. The requirement for CD4+ T cells [7] means that antigens must be released by, or exposed on, the surface of target larvae for processing and presentation by accessory cells to trigger such effector responses.
Targets of protective immunity have historically been identified by screening crude antigen preparations and expression libraries with sera from putatively immune hosts [8],[9]. In the schistosome context such screens have, in the main, produced a catalogue of abundant cytoplasmic proteins that one would not ordinarily expect to be secreted or surface-exposed and thus available to the immune system. Indeed, the abundance and antigenicity of cytoplasmic proteins appears to pose a major obstacle to identifying truly protective antigens. Abundant transcripts can dominate the content of cDNA libraries; equally, highly expressed proteins may mask attempts to identify bona fide vaccine candidates using proteomics [10] or immunoproteomics [11]. Clearly alternative approaches are needed to pinpoint antigens relevant to protection in this model and the sequencing of the schistosome transcriptome [12] and genome (www.GeneDB.org) now provide unparalleled opportunities for rapid progress using post-genomic techniques [13].
We have previously constructed a microarray comprising cDNAs derived from normal lung stage schistosomula and used it to identify genes highly expressed in the migrating parasite relative to six other life cycle stages [14]. We found genes encoding six membrane, six membrane-associated and five secreted proteins that were preferentially expressed at the lung or skin and lung stage. However, when considered in isolation it is difficult to predict which of these proteins, if any, will make suitable vaccine candidates. Their site of expression in the complex parasite body is unknown and some are hypothetical proteins with no ascribed function except at the motif or domain level. We now report use of the same lung stage array to pinpoint transcripts differentially expressed between normal and irradiated parasites cultured to the lung stage. This experiment was designed to identify the molecular changes underlying the altered phenotype, primarily using Gene Set Enrichment Analysis (GSEA) to delineate groups of genes with associated functions, which could explain the enhanced immunogenicity of the irradiated larvae.
A Puerto Rican isolate of S. mansoni was maintained by passage through NMRI strain mice and Biomphalaria glabrata snails, the animal work being approved by the Biology Department Ethics Committee, University of York. The microarray [14] was screened with mRNA from schistosomula, derived from mechanically transformed cercariae and grown in vitro for four, seven or ten days [15]. The times were chosen on the basis of previous parasite tracking [4] and lymphadenectomy experiments [16]. Attenuated schistosomula begin to accumulate in the lymph node and lung at day four, reaching a plateau in both sites at day seven [4]. Excision of skin-draining lymph nodes at, or prior to, day ten has a major ablative effect on subsequent protection [16] The cercariae were obtained by exposing snails with a patent infection to a bright light. Prior to culture one half of each cercarial shed was exposed to 200 Gray of radiation from an X-ray source at Cookridge Hospital, Leeds.
The microarray (ArrayExpress A-SGRP-2/E-TABM-408) containing approximately 6000 features printed in duplicate (accession numbers AM042715-AM048613), the hybridisation protocol, and array scanning were as described in Dillon et al. (2006). The array represents 3088 unique sequence contigs and singlets, encompassing an estimated 44% of the lung worm transcriptome [12]. At each of the day four, seven and ten sampling times total RNA was extracted from parallel cultures of normal and irradiated schistosomula with Trizol (Invitrogen) according to manufacturer's instructions. Each total RNA was labelled with Cy3 or Cy5 dyes (Perkin Elmer), without amplification, before hybridisation to the array at 20μg per channel [14]. Analysis of the normal and irradiated treatments, in pairs, at three time points encompassed twelve slides, comprising three biological replicates per treatment and one technical replicate (i.e. one of the biological replicates was split and repeated in order to control for experimental error). Dye swaps were balanced across treatments to limit bias resulting from differential dye incorporation and intensity, i.e. 50% of irradiated samples were labelled with Cy3 and 50% with Cy5. One sample from day ten failed to label so only 3 slides in total contributed to that time point.
The quantative dataset obtained using the GenePix 4000B instrument (Axon Instruments Inc.), was analysed with the GenePix Pro software and the R language for statistical computing (www.r-project.org) [17]. Specifically, the data was processed with the microarray analysis tools available from the Bioconductor Project, a tool for the analysis and comprehension of genomic data (www.bioconductor.org) [18]. The background was subtracted from array data using a Bayesian model-based method [19]. Array data were normalized using the LIMMA component (Linear Models for Microarray Data) of the Bioconductor package [20] with printtip loess to correct for spatial and other artefacts generated during the printing process. (Loess is a locally weighted polynomial regression; see LIMMA documentation.) Linear models were applied and significance statistics generated using empirical Bayesian methods to assess differential gene expression. This has the effect of borrowing information from the ensemble of genes to aid with inference about each individual gene [20]. An observation was classed as significant if it exceeded a natural log-odds (lods) cutoff of 3. To determine the effects of radiation, irrespective of sampling time, normal and irradiated results were pooled and reanalysed as a two way comparison. Detailed description of the methods used can be found in the LIMMA documentation: http://bioconductor.org/packages/2.1/bioc/vignettes/limma/inst/doc/usersguide.pdf.
As each EST is duplicated on the array, mean red and green values for the 6528 probes were generated from background-subtracted red and green fluorescence values. The LIMMA function “normalizeQuantiles” *was applied to these mean fluorescence values to normalize between arrays. Thus each quantile of each EST is adjusted to its mean across all arrays, irrespective of channel, normalising the data by ensuring the signal intensities within each treatment have the same empirical distribution. In those instances in which two or more ESTs on the array were members of the same Sm contig, the mean normalised values were taken, resulting in a single value for each contig. The normalised signal intensities were combined into tables containing all Sm contigs and singlets, with their GO/Protein analyst annotation (as described in Dillon et al. 2006), test channel signals and reference channel signals, prior to submission to the GSEA package. GSEA statistically assesses whether expression of groups of genes correlates with a given phenotype, and requires those groups to contain 15 or more members to function [21]. It quantifies the enrichment of individual members at the top and bottom of a ranked list of gene expression. The enrichment score (ES) is calculated by parsing the ranked gene list for members of a single category, and increasing a running-sum statistic when one of those genes is encountered or decreasing that sum if it is not. The enrichment score is then normalized by adjusting for the number of genes in a category and the GSEA package estimates the significance of each normalized enrichment score (NES) by calculating a false discovery rate (FDR). Gene sets were deemed to be enriched when the FDR ≤ 0.25, this apparently relaxed cut-off being used because the primary goal of GSEA , as specified by Subramanian et al. (2005), is to generate hypotheses rather than exclude every last false positive. The FDR is calculated by comparing the tails of the observed and null distributions for the NES. The null is produced by randomly assigning phenotype labels and producing a reordered gene list; this is done 1000 times to generate a null ES for each set. The LES is defined as the core grouping of genes contributing to the enrichment score; this generally represents approximately 30–50% of genes in an enriched category [21].
Three ESTs deemed to be differentially expressed, using the LIMMA package of Bioconductor, plus one on the threshold of significance were chosen for validation of array predictions by real time PCR analyses. The ESTs and primers used are outlined in Table S1. The Primer Express package (Applied Biosystems) was used to design primers to the four ESTs and the 18S ribosomal RNA control. A dissociation plot was performed for each primer to determine specificity. Comparable amplification was confirmed and assays performed in triplicate, on an ABI 7300 PRISM instrument using SYBR green dye, according to the manufacturer's instructions. All data were normalized to the lowest level of expression as determined by real time PCR.
The LIMMA package analysis of changes in single genes across the three time points highlights only seven significant differences between irradiated and normal parasites (Table 1). In all cases genes reaching our stringent statistical cut-off are conspicuously down-regulated in the irradiated parasite. Two differentially regulated transcripts encode proteins destined for the plasma membrane. One of these is the previously characterised Sm25 (also known as Gp18–22) and the other codes for a hypothetical protein. A third transcript encoding Tetraspanin D (Sm-TSP-2) , known to be present at the tegumental surface [22], is down-regulated at all three time points. Of the remaining genes revealed, JF-2 codes for a membrane-associated cytoskeletal component thought to link actin filaments to the plasma membrane, cdc2 is a key control enzyme of the cell cycle and two code for hypothetical proteins (Table 1).
The level of expression of the four selected ESTs was determined using real time RT-PCR and compared with that estimated from the array hybridisations (Figure S1). Plotting the data as a histogram highlights the broad level of agreement between the two techniques, and shows that in contrast to previous work [14], differences in sensitivity are not as pronounced. This is likely due to the smaller variations in expression measured. A scatter plot of the same data (data not shown) demonstrates that the two methods exhibited high concordance, with a correlation coefficient R = 0.80.
The expression of genes, grouped by biological function or subcellular location, correlating with a specific phenotype, was assessed using the GSEA package. A heat map recording the differential expression of the 1769 unique features on the array is presented in Figure 1A. A symmetrical distribution of expression profiles, can be discerned whereby two thirds of the genes are visibly associated with a phenotype. Approximately one third are up-regulated in the irradiated parasite (the red-dominated top left corner) and a different third in the normal parasite (the red-dominated bottom right corner); the remaining central third display no obvious pattern. A list of genes ranked by the intensity of their expression was derived from the heat map (Figure 1B) and used to produce a graphical plot of the running sum statistic (Figure 1C, E and G). This running sum statistic increases every time a member of a given gene set (i.e. a GO category) is encountered in the ranked gene list and decreases when it is not encountered. Where no correlation occurs between a gene set and the N or I phenotype the genes appear randomly in the ranked list producing a plot of the running sum statistic that fluctuates either side of zero (e.g. Figure 1C). The running sum for expression of the gene sets that correlate with the irradiated phenotype (e.g. Figure 1E) is skewed to the left by the abundance of numerous members in that region of the ranked gene list (Figure 1F). Conversely, correlation with the normal phenotype is skewed to the right (e.g. Figure 1G and H). The leading edge subset (LES) represents the core of genes most strongly associated with the N or I phenotype (Figure 1E and G).
The most prominent enriched gene sets at day four include ‘protein modification’ in the irradiated parasite and ‘RNA-directed DNA polymerase activity’ (root GO term is ‘Molecular Function’, which is identical to the ‘Biological process’ category ‘RNA-dependent DNA replication‘) in the normal parasites (Table 2). The former category is noteworthy for containing the E1-3 ubiquitinating enzymes in its LES (Table S2) while the latter appears to consist primarily of retrotransposon transcripts (at least 17 members, data not shown). The ‘GTP binding’ LES associated with the irradiated phenotype contains no fewer than 17 ras/rab/rac small G-protein homologues, together with a stimulatory and an inhibitory heterotrimeric G-protein alpha subunit (Table S2). A fifth ‘calcium ion binding’ category, also associated with the irradiated phenotype, possesses numerous transcripts encoding proteins of disparate motor or structural function including EF hand-containing proteins such as Sm22.6, myosin, at least two annexins and severin (Table S2).
At day seven, only the irradiated phenotype shows gene set enrichment (Table 2). Gene categories associated with ‘metabolism’, ‘mitochondrion’ and ‘electron transport’ are overrepresented and an analysis of LES overlap reveals commonalities between the three subsets of enriched genes. The genes shared are specifically involved in the respiratory electron transport chain. A number of cytochrome subunits, NADH metabolising and antioxidant thioredoxin enzymes all contribute to the enrichment score of the three gene sets (Table S3). Protein synthesis and degradation also appear to be prominent processes in the day seven irradiated parasite. The categories ‘protein biosynthesis’, ‘protein folding’, ‘proteolysis’, ‘ribosome’, ‘structural constituent of ribosome’, ‘cysteine-type peptidase activity’ and ‘isomerase activity’ encompassing genes encoding translation initiation factors, isomerases and chaperones all correlate with the irradiated parasite phenotype (Table 2). Intriguingly, transcripts encoding extracellular proteins also appear to be enriched, although the heterogeneous nature of this LES makes it difficult to discern a biological pattern (Table S3). Nevertheless, the presence of the antigen 5 transcript, protease inhibitors and a lipoprotein receptor is noteworthy.
The transcriptional divergence of the irradiated and normal parasites is even more apparent by day ten; 32 of the 89 gene sets submitted to GSEA show enrichment correlating with one or other phenotype. For the irradiated parasite the enrichment of ‘ribosome’ components persists into day ten and ‘RNA-directed DNA synthesis’ is again comparatively depressed with respect to the normal parasite. Indeed, at this stage the irradiated parasite differs from the normal parasite in many biological systems (Table 2). Categories for ‘transcription regulation’, ‘RNA binding’ and ‘helicase activity’ are under-represented in the irradiated parasite and there is also a relative shortfall in ‘intracellular kinase signalling’ and ‘structural proteins’, specifically cytoskeletal transcripts. The comparative paucity of receptor-encoding transcripts is particularly striking in the irradiated parasite, as is the general dearth of transcripts from the gene sets ‘endoplasmic reticulum’, through the ‘golgi’ to the ‘plasma membrane’. An analysis of the LES of the ‘receptor activity’ category reveals an overlap with other gene sets diminished in their own right, including ‘ion channel activity’ and ‘G-protein coupled receptor signalling’. The functional overlap reveals that a significant proportion of these transcripts are neuroreceptors or channels, including acetylcholine, purinergic, nicotinic, glutamate and aspartate receptors plus voltage and ligand-gated ion channels (Table S4).
Examining the pooled data for differential enrichment of categories, in normal versus irradiated, emphasises the apparent importance of protein synthesis and degradation in the irradiated parasite. Categories associated with protein metabolism, including ‘protein folding’ are prominent as is the ‘cysteine-type peptidase activity’ GO set, containing a number of cathepsins, and a ‘cytosol’ set that contains proteosome activators and some 20S proteosome components in its LES (Table S5). While Golgi-related transcripts do not meet the FDR cut-off, the overlap between the ‘golgi’, ‘GTP binding’ and ‘small GTPase mediated signal transduction’ LES together with deficiencies in the ‘ER’ category is noteworthy (Table S5). Although the comparative paucity of ‘receptor activity’ transcripts in the irradiated parasite is not obvious at days four and seven, the receptor activity is depressed at all time points when irradiated versus normal parasites were compared (Table 2). Analysis of the category ‘ion channel activity’ comprising transcripts encoding receptors associated with ion flux across membranes is also diminished in the irradiated parasite. Signalling cascades, particularly ‘kinase activity’ (Table 2) are also less abundant in the irradiated parasite. The kinases may well interact with the ‘cell adhesion’ and ‘cytoskeleton’ categories contributing to the observed differences (Table 2) but overlap analysis does not indicate shared genes in their respective LES.
Lung stage schistosomula of S. mansoni are a validated target of protective immunity induced in the murine host by exposure to RA cercariae. However, attempts to identify the antigens responsible, a key step in the development of a recombinant vaccine, have met with limited success [1],[8]. Microarrays offer a route to antigen identification by pinpointing subtle differences in gene expression between irradiated and normal worms, irrespective of transcript abundance. Characterising the underlying transcriptional differences should highlight changes at the parasite-host interface that explain why irradiated larvae can elicit protective immunity when normal larvae do not. In addition, by shifting the focus away from antibody-based technologies, microarrays may identify genes encoding non-immunogenic proteins that are nevertheless fundamental to parasite migration and development.
Analysis of the normal and irradiated parasite transcriptomes at day four, seven and ten revealed only seven genes that showed significant differences in expression. All were down-regulated as a result of radiation. Proteome Analyst predicted two as plasma membrane proteins (Sm25; hypothetical protein). A third, a tetraspanin (Sm-TSP-2, CD63-like, tetraspanin D), mispredicted as lysosomal, is known to be exported to the tegument surface plasmamembrane [22], as is Sm25 [23],[24]. Biotinylation studies on adult worms indicated that tetraspanin D may play a role in maintaining tegumental membrane structure and organisation and could be accessible to the immune system [22]. Indeed, this particular tetraspanin, identified using a signal sequence trap [25], elicited protective immunity when the major extracellular loop was used to vaccinate mice [26]. Conversely Sm25, or its decorating glycans, may actually protect the parasite by subverting the host immune response as, despite eliciting high antibody titres, the recombinant protein does not protect vaccinated animals [27]. On the basis of membrane association and immunofluorescence studies it has been suggested that the actin binding protein JF-2 may be available at the tegument surface [28]. However, a sizeable proportion of patients infected with S. japonicum possess antibodies to JF-2 [28], yet continual chemotherapy is still required to limit the impact of reinfection [29]. This observation argues that JF-2 normally confers little or no resistance and may simply be another cytoplasmic protein albeit one associated with plasma membranes [30]. Cdc2, the final protein with an ascribed function, is a crucial cell cycle control enzyme. While the down-regulation of a single gene should not be over interpreted, suppressed levels of the cdc2 protein may reflect the inability to re-enter the cell cycle [31]. Migrating schistosomes are in a semi-quiescent metabolic state (Lawson and Wilson, 1980) with no cell division taking place [32]. However, they are primed to enter cell cycle upon reaching the portal vein and beginning to blood feed [33]. Thus, down-regulation of cdc2 may be part of the explanation why irradiated parasites never mature.
It is clear from numerous studies on the RA vaccine (reviewed by Coulson 1997) that attenuated parasites must persist in the host for 1–2 weeks to elicit effective protection. Furthermore they must also migrate beyond the skin to its draining lymph nodes, and to the lungs. As anticipated, large transcriptional changes were not evident four days or more after the radiation insult, since the acute stress response has long subsided by that time [34]. Therefore, the ability to detect small coordinated changes, using the GSEA package developed by Subramanian et al. was particularly important as a means of dissecting out the longer-term effects of radiation exposure. Schistosomula undergo marked phenotypic changes while resident in the skin soon after penetration, which include remodelling of the tegument surface and ablation of penetration glands [35]. Subsequently, mid-body spines are lost and the larval body elongates to facilitate intravascular migration beyond the lungs [36]. At day four, approximating to the skin stage, it was difficult to detect meaningful differences in transcript abundance, suggesting that the delayed effects of radiation were very subtle. However, RNA-directed DNA synthesis, indicated by retrotransposon transcription, was more prominent in the normal parasite. Why this should be depressed in the irradiated parasite is unclear but could reflect long term suppression by DNA repair mechanisms [37]. The ‘protein modification’ category provided an early indication of enhanced protein metabolism in the irradiated parasite. By day seven the protein metabolism categories specified by GSEA revealed a more pronounced effect but this distinction was diminished by day ten. Despite the lack of obvious morphological differences between early normal and irradiated parasites (Mastin et al., 1983) our data are consistent with the observations of Wales et al (1992) that protein synthesis is temporarily inhibited by irradiation. It seems likely that body remodelling has been delayed so the enhanced protein metabolism may reflect a catch-up process relative to the normal parasite. In a similar vein the switch to anaerobic respiration [38] may be retarded as evidenced by the enrichment of energy metabolism categories at day seven, in the irradiated parasite.
By day ten the divergence between normal and irradiated parasites was greatest, with the majority of highlighted gene sets down-regulated in the latter. We consider that these represent biological processes damaged beyond recovery by the now-distant radiation event. The decreased prominence of categories involving intracellular signalling (e.g. ‘G-protein coupled receptor signalling pathway’) may indicate a reduced ability to respond to external developmental cues; the down-regulation of cdc2, already noted, should be viewed in this context. In addition, deficiencies in structural categories such as ‘cytoskeleton’ may further impede the irradiated parasite's capacity for locomotion. This apparent inability to detect and respond appropriately to the surroundings is further reinforced by the comparative paucity of receptor transcripts, especially those encoding components of neurotransmitter pathways. All these categories identified by GSEA accord with the visible phenotype revealed by SEM studies [3]. Although the irradiated parasite is in most respects morphologically similar to the normal parasite, elongating and losing mid-body spines [39], it nevertheless displays abnormal constrictions of circular muscle fibres in the body wall, resulting in uncoordinated movement [3]. It is this compromised locomotion that leads to the persistence of irradiated parasites in the host lymph nodes and lungs for five weeks or more [4],[39]. In the lymph nodes the parasites drive lymphocyte proliferation [40] and in the lungs they act as a long-term stimulus to recruit lymphocytes that arm that organ against challenge parasites [41].
The reason that RA parasites in general elicit protective immunity when normal parasites do not has long been the subject of speculation and investigation [42]. Our study strongly indicates that the up-regulation of specific gene products to provide elevated immune stimulation is not the key. Indeed an expressed fragment of the tetraspanin gene that we detected as down-regulated by single gene analysis, was recently shown to have protective potential in the mouse [26]. This underlines our thesis that even if gene expression is reduced, the extended stay of attenuated parasites in the skin draining lymph nodes may still result in enhanced immune priming against exposed antigens. Equally once the host has been primed by the vaccine, antibody or cell mediated effector responses could act early upon the incoming parasite, after cercaria-schistosomulum transformation has been completed; from previous microarray experiments we already know that tetraspanin D is strongly expressed in the two day old schistosomulum [14].
We cannot rule out that parasites in vivo respond differently to some host factor, not present in vitro, by up-regulating specific genes as suggested by ex vivo experiments [43]. However, the subtle nature of differences between normal and irradiated parasites leads us to believe that changes in protein expression are poor indicators of potential antigenicity; it is likely that the accessibility rather than abundance of an antigen is the important factor. In this context, retarded development increasing the duration of immune stimulation appears to be the salient feature. In the long term, the radiation insult compromises the transcription of schistosome genes involved in neuromuscular activity and ultimately cell cycle progression. In this respect schistosomes are particularly suited to deliver a prolonged stimulus as they undertake a protracted migration from skin to portal system, lasting 7–21 days after penetration (i.e. irradiation). Only when blood feeding and cell division begin in the liver [44] will DNA strand breaks prove lethal. This priming by larvae is quite distinct in both location and antigen load from the continuous priming over months to years provided by adult worms and their eggs. Furthermore, recent studies in the baboon model have shown that protective responses elicited by the irradiated vaccine are dissociated from both responses to chemotherapy and an ongoing chronic infection [45].
Exposure to irradiated metazoan and protozoan parasites has been widely used to study protective immunity, as the basis for vaccine development, but we believe this is the first attempt to interrogate the transcriptome of such a parasite. In addition to schistosomes, protective immunity is induced by radiation-attenuation of the nematodes Dictyocaulus and Ancylostoma spp. and the protozoa, Plasmodium, Eimeria and Theileria spp. [46]–[50]. Given the efficacy of radiation-attenuated parasites as vaccines, the findings of this study should provide pointers to the phenotypic changes that account for the success of these other parasites as inducers of protective responses. |
10.1371/journal.pgen.1006224 | Downstream Antisense Transcription Predicts Genomic Features That Define the Specific Chromatin Environment at Mammalian Promoters | Antisense transcription is a prevalent feature at mammalian promoters. Previous studies have primarily focused on antisense transcription initiating upstream of genes. Here, we characterize promoter-proximal antisense transcription downstream of gene transcription starts sites in human breast cancer cells, investigating the genomic context of downstream antisense transcription. We find extensive correlations between antisense transcription and features associated with the chromatin environment at gene promoters. Antisense transcription downstream of promoters is widespread, with antisense transcription initiation observed within 2 kb of 28% of gene transcription start sites. Antisense transcription initiates between nucleosomes regularly positioned downstream of these promoters. The nucleosomes between gene and downstream antisense transcription start sites carry histone modifications associated with active promoters, such as H3K4me3 and H3K27ac. This region is bound by chromatin remodeling and histone modifying complexes including SWI/SNF subunits and HDACs, suggesting that antisense transcription or resulting RNA transcripts contribute to the creation and maintenance of a promoter-associated chromatin environment. Downstream antisense transcription overlays additional regulatory features, such as transcription factor binding, DNA accessibility, and the downstream edge of promoter-associated CpG islands. These features suggest an important role for antisense transcription in the regulation of gene expression and the maintenance of a promoter-associated chromatin environment.
| Gene transcription is regulated by the coordinated interaction of genetic, epigenetic and trans-acting factors. The chromatin environment at gene promoters, including positioned nucleosomes that may display functional histone modifications, is a key regulator of gene expression, contributing to transcriptional activation and repression. In addition to sense-strand transcription of gene sequences, antisense transcription is prevalent at gene promoters. Often resulting in a short-lived non-coding RNA transcript, the function of antisense transcription is poorly understood. Using next-generation sequencing techniques, we characterized transcription in human breast cancer cells and found extensive correlations between antisense transcription and the chromatin environment at promoters. We found that downstream antisense transcription initiates from between regularly positioned nucleosomes and that those nucleosomes between sense and downstream antisense transcription start sites display histone modifications associated with active gene promoters. Chromatin remodelers and other protein complexes responsible for creation and maintenance of the promoter chromatin environment associate with this same region, suggesting an important role of antisense transcription in the regulation of gene expression.
| The promoter region is intimately tied to the transcription of genes, providing the initial site of transcriptional machinery binding and assembly. Comprised of DNA elements in defined spatial arrangements [1], promoters also display key genomic features that facilitate gene regulation. Active promoters present a nucleosome-deprived region that allows for the association of RNA polymerase II (Pol II) and transcription factors [2]. Active promoters also possess distinct histone marks associated with gene expression [3].
Divergent transcription has emerged as a common feature of mammalian promoters [4–7]. In divergent transcription, an additional transcription event initiates upstream and antisense of a nearby gene promoter. Though divergent transcription at promoters may result in two different protein-coding transcripts in opposite orientations, often a short-lived non-coding RNA is transcribed anti-sense of a gene [8–10]. In divergent transcription at mammalian promoters, antisense transcription initiates at the upstream antisense transcription start site (uaTSS). The position of the uaTSS intersects with distinguishing features associated with promoters, with uaTSSs falling at the border of nucleosome-depleted regions where transcription factor binding may be enriched [11]. Additionally, uaTSSs tend to broadly coincide with the upstream edge of promoter-associated CpG islands [11].
In addition to divergent transcription, convergent transcription has been observed at genes in a variety of systems, ranging from fly and yeast to mammals [12–14]. Convergent transcription is possible in a variety of different gene structures, including non-overlapping genes on opposite strands and genes with internal promoters [13]. Promoter-proximal convergent transcription is common; analysis identified convergent transcription at roughly one quarter of queried genes [15]. Here, we describe genetic and epigenetic features at downstream antisense transcription start sites (daTSSs) associated with promoter-proximal convergent transcription in human T47D/A1-2 cells. We find that daTSSs coincide with the downstream edge of promoter-associated genomic features, such as promoter-associated histone marks. Though convergent transcription has been suggested as a repressive feature [15], we find that genes with observable daTSSs do not display lower gene expression in T47D/A1-2 cells. Despite this, coincidence of daTSSs with a variety of transregulatory factors, such as transcription factors and chromatin remodelers, suggests an intimate connection between antisense transcription and gene regulation.
To characterize transcription initiation in human T47D/A1-2 cells, Start-seq was performed on nascent RNA transcripts [16]. In a Start-seq experiment, RNA is isolated from the nucleus and selected for short size, allowing characterization of nascent RNA transcripts that may be subsequently degraded and that otherwise could not be analyzed in a traditional RNA-seq experiment. Cap-sensitive degradation ensures that the 5’ end of each read in a Start-seq experiment corresponds to a TSS with single nucleotide resolution. Gene TSSs and uaTSSs were identified using previously described methods [11]. daTSSs were called in a method analogous to that used for uaTSS identification. In brief, for each gene TSS called, a search window from the gene TSS to 2 kb downstream was defined. For each search window with reads exceeding an FDR-defined significance threshold (5 reads, observing aligned library depth), a daTSS was called at the position with greatest read density on the opposite strand relative to the gene TSS. Stringent filtering was used to ensure that an identified daTSS could not be a miscalled TSS of another gene or a uaTSS of an alternative start site.
Observable genes with antisense transcription display overlap between identified TSSs and the 5´ end of Start-seq reads on both sense and antisense strands (in black and in red; Fig 1A and 1B, respectively). Identified gene TSSs are consistent with RNA-seq read density, and both gene TSSs and antisense TSSs display overlap with Pol II ChIP-seq reads (in blue and red, respectively; Fig 1A). Over 10,391 observed gene TSSs, 5,519 uaTSSs (53% of gene TSSs) and 2,956 daTSSs (28%) were identified (Fig 1B). Over all observed gene TSSs, both a uaTSS and a daTSS were identified at 1,815 genes, indicating a statistically significant association between the two events and implying potential cooperation between upstream and downstream antisense transcription (two-sided Fisher’s exact test: p-value < 10−6). To ensure reproducibility, a biological replicate was prepared from the same cell-line, and similar read densities were found at the identified TSSs (S1A Fig). The overall rate of identification is consistent with previous approaches using other experimental methods [7,11,15]. Start-seq read counts are greatest at gene TSSs, with lower counts at uaTSSs and daTSSs (averages of 776, 467, and 276 reads at gene TSSs, uaTSSs, and daTSSs, respectively). Given that the same read threshold was used to call all classes of TSS and that stringent filtering was used to limit miscalling of antisense TSSs, we anticipate that our rate of identification underestimates the prevalence of antisense transcription. daTSSs were also identified in mouse macrophage cells considering observed gene TSS positions found in previous work [11]. 4,921 daTSSs were identified over 12,229 observed gene TSSs, or roughly 40% of genes (S2 Fig). The difference in the identification rate between human T47D/A1-2 cells (28%) and mouse macrophage cells (40%) may be indicative of variability in the landscape of antisense transcription across organisms and cell types. Heatmaps centered on gene TSSs and sorted by the distance from the observed daTSS show that human T47D/A1-2 and mouse macrophage calls are coincident with enriched Pol II ChIP-seq signal (Fig 1B; S2B Fig), supporting calls as bona fide Pol II-dependent transcription events. There is a lack of stranded RNA-seq coverage immediately downstream of uaTSSs and daTSSs (S3A Fig), indicating that transcripts originating from these sites are not present at steady-state in the cytosol and are short-lived.
daTSSs identified in T47D/A1-2 cells were investigated in other cell lines using other experimental data types associated with nascent transcription (MCF-7, GM12878, K562, HeLa S3, and HEK239T cells; S3B Fig). These experiments include a variety of sequencing-based approaches designed to interrogate different biological phenomenon, including direct Pol II-DNA interactions and nascent Pol II-associated transcription. Despite wide ranging differences in technology and cell line, we found an enrichment of nascent RNA- or Pol II-associated signal at daTSS positions. Given the conservation of daTSSs across cell lines, we leveraged a variety of publically available data from different cell lines to annotate daTSSs. Additionally, though we found that a majority of daTSSs are preserved, some do not display signal in other cell lines (S3C Fig). Of the 2,956 daTSSs identified in T47D/A1-2 cells, 1,985 (67%) and 1,966 (67%) display GRO-cap signal in K546 and GM12878 cells, respectively. We examined these genes to see if they displayed cell-specific characteristics. daTSSs with no GRO-cap signal in either K546 or GM12878 cells and that may be considered T47D/A1-2 specific (605 genes; 21%) show an enrichment in categories associated with breast cancer (“breast or ovarian cancer”: p-value = 1.28 x 10−5; “mammary tumor”: p-value = 2.10 x 10−5). This enrichment suggests cell-line specificity in downstream antisense transcription.
We examined sequence content at all observed TSSs to both further characterize identified TSSs as associated with Pol II-dependent transcription events and to compare across the three observed TSS classes. Generally, we find that sequence content is similar across all three classes. All TSSs show enrichment for GC content (Fig 2A). Consistent with previous observations, we find enriched GC content upstream of T47D/A1-2 uaTSSs [11]. In addition, we see enriched GC content downstream of daTSSs. Taken together, the antisense TSSs coincide with apparent boundaries of GC content enrichment. The positioning of CpG islands is generally consistent with this observation. Observed uaTSSs and daTSSs broadly coincide with the upstream and downstream edges of promoter-associated CpG islands, respectively (Fig 2B) [17]. However, the alignment between CpG island boundaries of daTSSs is not absolute. Of those 2,392 daTSSs whose associated gene TSS overlaps a CpG island, 1,325 (55%) are within 250 bp of the downstream edge of the CpG island. Thus, observation of a daTSS is not dependent on the presence and relative location of a promoter-proximal CpG island.
Considering a narrow 100-bp window around each TSS (Fig 2A), we observe sequence patterns that are preserved across each class of TSS. Roughly 25 bps upstream of the TSS, we identify an area enriched for TA content. Centered on the TSS itself, there is a distinct sequence pattern with pyrimidine-purine dinucleotide at its center. De novo motif discovery reveals enriched sequences similar to TATA box and Inr binding motifs at these regions (Fig 2D) [20]. Though these motifs are present near each class of TSS, subtle differences in relative enrichment of Inr-like motifs (found at 24%, 17%, and 41% of gene TSSs, uaTSSs, and daTSSs, respectively; Fig 2D) may reflect differences in regulation. The occurrences of these specific motifs are consistent with general patterns found in searches performed with Pol II-associated sequence motifs (Fig 2C). Regardless of TSS class, an enrichment of Pol II-associated motifs [18] is observed upstream of TSS positions. GC-box occurrences dominate the enriched motifs (S3D Fig) with fewer motifs identified at daTSSs.
We additionally investigated sequence conservation at daTSS positions. Over promoters displaying antisense transcription, we examined PhyloP conservation scores calculated from sequence alignments across placental mammals [19]. Positive PhyloP scores indicate enhanced sequence conservation immediately upstream of daTSSs (Fig 2E). This is consistent with observations seen at both gene TSSs and uaTSSs [11]. Positive scores are also evident between gene TSSs and daTSSs, implying sequence conservation in this region (Fig 2E). Evolutionary pressure to maintain sequence at daTSSs suggests functional significance for downstream antisense transcription.
We next sought to examine the connection between downstream antisense transcription and the expression of associated genes. Recent studies have proposed that promoter-proximal downstream antisense transcription represses expression of upstream genes [15]. We examined the expression of genes displaying downstream antisense transcription in two distinct ways. We first compared the extent of transcription initiation at those genes with and without observed daTSSs (categories “w/ daTSSs” and “w/o daTSSs”; Fig 3A). We restricted this comparison to those genes without observed uaTSSs to limit the effect of upstream antisense transcription on observed trends (N = 4,872). The difference between genes with and without daTSSs is not statistically significant (Wilcoxon test: p-value = 0.37). We also find that the extent of downstream antisense transcription as measured by Start-seq read counts at daTSS positions shows no correlation with read counts at gene TSSs (genes with no observed uaTSSs; Spearman test: rho = 0.017, p-value = 0.56).
We next compared steady-state transcript levels using RNA-seq-derived FPKM values. The results for RNA-seq analysis are comparable to those derived from Start-seq (Fig 3B). Again, comparisons were restricted to genes without observed uaTSSs. The difference in FPKM values between genes with and without daTSSs is not statistically significant (Wilcoxon test: p-value = 0.89). In summation, downstream antisense transcription is not associated with lowly expressed genes in human T47D/A1-2 cells. These results suggest convergent transcription near gene promoters is not inhibitory.
These results are distinct from those reported in recent a study describing promoter-proximal downstream antisense transcription as mark of lowly expressed genes [15]. This conclusion was based upon differences in the gene body density of Net-Seq reads between genes displaying only upstream antisense or downstream antisense transcription. When we use a similar categorical separation of genes, we find that genes displaying only uaTSSs show greater steady-state transcription levels than genes displaying only daTSSs (Wilcoxon and Kolmogorov-Smirnov tests: p-values < 10−6; S4A Fig). However, these differences appear attributable not to the presence of downstream antisense transcription but to the absence of upstream antisense transcription. Considering genes without daTSSs (N = 7,435), those genes with uaTSSs display significantly higher steady-state transcription than genes without uaTSSs (Wilcoxon test: p-value < 10−6; S4B Fig). These results suggest that a complete understanding of the effects of downstream antisense transcription requires deconvolution from that of upstream antisense transcription.
An examination of promoter proximal pausing suggests that downstream antisense transcription may be coordinated with transcription of associated genes. We find that Ccnt2 (or CycT2), part of the P-TEFb complex involved in the regulation of Pol II elongation, selectively associates to the area between gene TSSs and daTSSs and shows statistically significant enrichment compared to equivalent regions at genes without daTSSs (Wilcoxon test, p-value <10−6; Fig 3C). This is coincident with enrichment of components of NELF and DSIF (Wilcoxon tests, p-values < 10−6; Fig 3C). PRO-seq experiments measuring strand-specific association of elongating Pol II also show a sense-strand enrichment of Pol II near the daTSS (Fig 3C), further supporting a connection between Pol II pausing and antisense transcription. These results do not necessarily imply that genes displaying downstream antisense transcription are more paused. Pausing itself is a highly regulated event in gene transcription with connections to signal-dependent gene expression [21]. Pol II-dependent transcription antisense of genes may contribute to signal-dependent response of paused genes. Consequently, while downstream antisense transcription does not appear to correlate with steady-state transcription levels, antisense transcription could affect signal-dependent gene expression changes.
We find that the daTSS position coincides with the downstream edge of a chromatin environment displaying promoter-associated features. To investigate the interplay between antisense transcription and nucleosome positioning, we performed MNase-seq on T47D/A1-2 cells. As evidenced in the resulting data, nucleosomes are regularly positioned relative to all three classes of identified TSSs. Like gene TSSs, MNase-seq read density is consistent with “+1” nucleosomes placed immediately downstream of daTSSs and uaTSSs, though MNase-seq peaks are less sharp when centered on daTSS positions (Fig 4A). The daTSS location respects the regular positioning of promoter-proximal nucleosomes. A histogram of gene TSS-daTSS distances is anti-correlated with average MNase-seq density (Fig 4B), implying that daTSSs fall between regularly-spaced nucleosomes oriented at gene TSSs. This pattern is reproduced in MNase-seq data from other breast epithelial cells (MCF-7 cells; S1B Fig) [22]. The observed MNase-seq density is consistent with Pol II ChIP-seq density at daTSS positions (Fig 4A). Downstream of observed daTSS positions, we find a regular pattern of Pol II ChIP-seq read density that is attributable to Pol II initiating at gene TSSs and that mirrors the MNase-seq data. As previously described [11], we find that regions between gene TSSs and associated uaTSSs are nucleosome depleted. This is clear in heatmaps of MNase-seq density (Fig 4C, left). In comparison, the region between gene TSSs and daTSSs displays a regular pattern of positioned nucleosomes (Fig 4C, left). When quartiled by TSS-daTSS distance, TSSs with greater distances show less distinct downstream MNase-seq peaks, perhaps indicating less regular or more transient nucleosome associations at these positions (Fig 4C, right). However, the location of MNase-seq peaks downstream of gene TSSs seem to be similar across genes with and without daTSSs (S4C Fig), indicating that the location of the gene TSS predominantly influences nucleosome positioning.
Given that nucleosomes are regularly positioned at identified TSSs, we sought to characterize histone modifications in those regions. We find that histone modifications at those nucleosomes positioned between the gene TSS and daTSS are distinct from proximal regions. ChIP-seq data in HMEC cells [23] show an enrichment of histone marks associated with active promoters. When compared to equivalent positions at genes without daTSSs, H3K27ac and H3K3me3 modifications show significant enrichment by Wilcoxon test (p-values of 1.17 x 10−3 and 3.94 x 10−2 for H3K27ac and H3K4me3 modifications, respectively; Fig 4D). There is tendency for histone modification enrichment to end at the daTSS, as clearly seen in profiles for histone variant H2A.Z (Wilcoxon test: p-value < 10−6; Fig 4D). This same enrichment is not seen for H3K4me1 and H3K36me3 modifications, associated with enhancers and actively transcribed regions, respectively (Wilcoxon test: p-values of 0.53 and 0.44, respectively). Consistent with the lack of nucleosomes in this region, we do not see the same enrichment between the gene TSS and uaTSS (S5 Fig).
Given the observed histone modification profile, we next sought to characterize the relationship between the association of transregulatory factors and antisense transcription at gene promoters. The association of transcription factors is enriched at open regions of DNA. We characterized accessible regions of DNA in T47D/A1-2 cells using FAIRE-seq. Like gene TSSs and uaTSSs, FAIRE-seq reveals an open genomic region at daTSS positions (Fig 5A) [24]. This observed density is consistent with FAIRE-seq results reported in a previous study (S1C Fig) [24] and recapitulates observations made using DNase-seq [15]. Following characterization by FAIRE-seq, we performed protein motif searches in similar areas to characterize the potential of these areas to interact with DNA-binding proteins. Analysis of known vertebrate motif occurrences shows a depletion of protein-binding motifs between gene TSSs and daTSSs and an enrichment of motifs immediately upstream of daTSSs (Fig 5B) [18]. Consistent with these areas being open and enriched for protein-binding motifs, ChIP-seq data [23] reveal that the daTSS coincides with the binding of trans-regulatory factors. Transcription factors were found to associate with open regions at both the gene TSS and the daTSS (Fig 5C, top). Comparisons with equivalent positions at genes without daTSSs show a significant enrichment of transcription factor-associated signal at daTSS positions (Wilcoxon test: p-values < 10−6; Fig 5C, top, and S2 Table). We present selected transcription factors known to associate at gene promoters and broadly participate in a number of signal-dependent pathways. However, the coincidence of p300, a known co-activator of numerous additional transcription factors [25], implies potential interaction with many others at daTSS positions (Fig 5C, top-right). In contrast, chromatin remodelers were found to associate in the area between the gene TSS and the daTSS with a significant enrichment of associated signal at daTSS positions (Wilcoxon test: p-values < 10−6; Fig 5C, bottom, and S2 Table). Though this area displays high GC content, this alone does not explain the enrichment of chromatin remodelers; randomly selected genomic regions matched for GC content show diminished signal relative to these regions (Wilcoxon test: p-value < 10−6). Like daTSSs, a variety of trans-regulatory factors associate to uaTSS positions (S6, S7 and S8 Figs). However, differences in regions between gene and antisense TSSs distinguish these two classes of TSS. Unlike daTSSs, transcription factors associate to the nucleosome-depleted areas between gene TSSs and uaTSSs while chromatin remodelers do not. This likely reflects differences in nucleosome occupancy between nucleosome-rich TSS-daTSS regions and nucleosome-deprived TSS-uaTSS regions (Fig 4). Factors involved in the deposition of histone marks (CHD1-A and Sap30) and in the positioning of nucleosomes (SWI/SNF-associated factors) may contribute to the distinct chromatin environment seen at this region, with CTCF potentially contributing to the definition of this region (Fig 5C, bottom-right). This suggests a mechanism similar to that observed in yeast where antisense transcription may contribute to a chromatin environment that ultimately impacts gene expression [26].
Our analyses indicate that downstream antisense transcription proximal to gene promoters is common in mammals. Its coincidence with a number of different regulatory features suggests that antisense transcription borders the chromatin environment characteristic of promoters and may possess a regulatory role. Previous studies have characterized convergent transcription as a repressive feature of genes [27]. The promoter-proximal convergent transcription described here is a narrow subset of convergent transcription, where downstream antisense transcription initiates at or within 2 kb of a gene TSS. Based on our analysis, we see little evidence for repression of associated genes by downstream antisense transcription. Considered categorically, comparisons between all genes and those displaying observable daTSSs fail to show significant differences in levels of transcription initiation and steady-state expression levels (Fig 3). Ultimately, the interplay between antisense transcription and gene expression will be complex, as coincidence of daTSSs positions with other promoter-associated features suggests interplay with other regulatory pathways (S9 Fig).
Active enhancers are often present and transcribed within intronic regions of gene bodies. Given the position of daTSSs downstream of gene TSSs, there is a question as to whether downstream antisense transcription is predominantly the consequence of canonical enhancer activity within gene introns. The lack of a direct connection to increased expression levels suggests that downstream antisense transcription is not associated with active enhancers regulating the nearby gene. Though a majority of daTSSs are positioned within introns (74%), a similar proportion remain overlapped with introns when gene models are randomly shuffled (75%), indicating that daTSSs are not significantly enriched within introns (S3 Table). There is also a conspicuous lack of signal attributable to enhancer-associated H3K4me1 at observed daTSSs, though there is an apparent association of p300 and H3K27ac marks (Figs 4D and 5C).
Rather than describing a functionally distinct element, e.g. proximal enhancers, downstream antisense transcription seems to be a feature of promoters themselves. Along with antisense transcription upstream of gene TSSs, downstream antisense transcription may be an intrinsic feature at many mammalian promoters (Fig 6). There appears to be a connection between antisense transcription and promoter-specific features at genetic and epigenetic levels (for an overview of features at each class of TSS, see S9 Fig). daTSSs respect the positioning of promoter-proximal nucleosomes, with observed daTSSs falling within valleys of MNase-seq read density (Fig 4B). Downstream antisense transcription may affect nucleosome organization at promoters; genes with larger distances between gene TSSs and daTSSs display less prominent nucleosome-associated peaks in MNase-seq data (Fig 4C). daTSSs also coincide with the binding of transregulatory factors. In particular, regions between gene TSSs and daTSSs show association of chromatin remodeling factors (Fig 5C). These factors potentially contribute to the chromatin environment bordered by gene TSSs and daTSSs and distinguished by enrichment of promoter-associated histone marks. It is not clear whether these functions are attributable to generated transcripts or transcription itself, though produced transcripts are not apparently stable (S3A Fig). Recent studies suggest that non-coding RNAs generated near promoters participate in the establishment of nucleosome occupancy [28].
Though both are proximal to promoters, daTSSs and uaTSSs exist in distinct epigenetic environments (S9 Fig). daTSSs and uaTSSs display fundamentally different relationships with nucleosomes in the promoter region. uaTSSs are an apparent boundary of the nucleosome depleted region at promoters [11] while daTSSs initiate from between regularly oriented nucleosomes downstream of gene TSSs (Fig 4). The different relationships with nucleosomes seem to inform in part the differences observed with other epigenetic features. Nucleosome depletion near uaTSSs allows for transcription factor association between uaTSSs and gene TSSs while the presence of nucleosomes likely prevents transcription factor association between daTSSs and gene TSSs (Fig 5; S6 Fig) [11]. Likewise, the presence of nucleosomes gives functional relevance to the association of chromatin remodelers observed between daTSSs and gene TSSs but not between uaTSSs and gene TSSs (Fig 5; S7 Fig).
Despite differences in epigenetic features, tendency for association with transregulatory factors, and capacity to produce stable RNA transcripts, all three classes of TSS described in this work display similarities in sequence content, including enrichment for GC content and Pol II-associated sequence motifs (Fig 2). As such, antisense transcription appears to be encoded in genetic sequence. This connection between sequence content and epigenetic features provides the compelling suggestion that antisense transcription encoded by sequence may direct the positioning of nucleosomes and deposition of histone marks. Antisense transcription may also participate in signal-dependent modulation of epigenetic content where activation of sequence-encoded antisense TSS precedes nearby changes in chromatin structure. In this way, the collection of transcription initiation-associated sequence motifs near promoters may define regulatory potential for a given gene. This connection to sequence also provides a means to interrogate antisense transcription function. Future studies with selective mutation of associated sequence motifs may elucidate the function of antisense transcription and its coincidence with promoter-associated features. Directed mutagenesis could also establish the extent of the effect of antisense transcription on the chromatin environment at promoters.
We characterized downstream antisense transcription initiating near gene promoters in human T47D/A1-2 cells. daTSSs fall between regularly positioned nucleosomes downstream of gene TSSs. Histones within this region are enriched for marks closely associated with active promoter regions, such as H3K4me3 and H3K27ac modifications. Chromatin remodeling complexes show enriched binding upstream of observed daTSS positions, suggesting that antisense transcription contributes to the establishment and maintenance of a promoter-specific chromatin environment. Downstream antisense transcription is common to many human promoters, and daTSSs correlate with the downstream edge of promoter-associated chromatin features. Coincidence of daTSSs with these features suggests interplay between antisense transcription and regulatory pathways.
T47D/A1-2 cells were cultured in DMEM containing 10% FBS. Prior to RNA isolation, cells were cultured in medium supplemented with 5% charcoal dextran-treated serum for at least 24 hours. The A1-2 cell line is a previously described derivative of the T47D breast cancer cell line that overexpresses rat glucocorticoid receptor (GR) and contains a stably-integrated MMTV luciferase reporter gene [29]. Short capped RNA was isolated from T47D/A1-2 cells as previously described [16]. Libraries were generated using the Illumina TruSeq small RNA kit. Two independent replicates were performed. A primary dataset was generated from combined Illumina HiSeq and MiSeq runs. This data set was used in all TSS calling and downstream analysis. A secondary validation data set with fewer reads was generated on an Illumina MiSeq run. This data set was used to validate reproducibility of read density at called TSSs (S1A Fig). For this and other in-house sequencing experiments, libraries were prepared and sequenced by the NIH Intramural Sequencing Center (NISC).
Start-seq reads were first filtered by quality score; reads with an average Sanger score less than 20 were removed from analysis. Following quality filtering, Cutadapt (version 1.2.1) was used to remove adapter sequences [30]. Alignment of Start-seq reads was performed using Bowtie (version 0.12.8) to hg19 or mm9 genome assemblies [31]. From each uniquely mapped Start-seq fragment, the 5’ end was taken forward into the TSS calling procedure.
Identification of TSSs was performed based on methods described previously [11,16], TSS identification was guided by RefSeq annotation (retrieved 05/09/2014) [32]. From the reference annotation, a list of non-redundant TSSs was taken from all mRNA RefSeq IDs (“NM_”). 2000-nt search windows were created about each RefSeq TSS. If search windows overlapped and had the same common gene name, those search windows were merged. For other overlapping search windows, boundaries were defined as the midpoint between associated TSSs. The intersection of search windows and 5’ Start-seq ends was then determined.
TSSs were called within each window in which the strand-specific 5’ end read count at any given nucleotide position met or exceeded a threshold of 5 reads. This threshold was determined in a previously described method [33]. In short, the FDR was estimated based upon the distribution of Start-seq reads across the genome and a background model where the probability of finding a given number of aligned reads by chance is given by a sum of Poisson probabilities. The read threshold was selected to allow less than 1 expected false positive by this measure.
In those windows where a single nucleotide position met or exceeded the read threshold, a gene TSS was called. The calling method aims to select as the TSS the position with the highest read counts in window region with the highest read density. To accomplish this, two potential TSSs were first determined. The first was the position with the most aligned 5’ ends across the entire window. For the second, the search window was divided into 200-nt bins at every 10 nt across the search window. At the 200-nt bin with the most overlapped 5’ ends, the second potential TSS was called as the position with the most aligned 5’ ends. Of the two putative sites, the TSS closest to the associated annotated RefSeq TSS was selected.
Following gene TSS identification, uaTSSs and daTSSs were found. For uaTSSs and daTSSs, search spaces were defined in antisense orientation as 1 to 1000 nt upstream and 1 to 2000 nt downstream, respectively. uaTSSs and daTSSs were called at the position with the most aligned 5’ ends within the search window if the count at any single nucleotide position met or exceeded a threshold of 5 reads. To ensure that identified daTSSs were not simply mis-called gene TSSs or uaTSSs, additional criteria were used to filter daTSS calls. A daTSS call was filtered if (1) within 1000 nt upstream of the daTSS there was a genomic position with Start-seq reads greater than or equal to 10% of reads at the associated gene TSS on the same strand as the gene TSS, implicating the daTSS as a potential uaTSS for an uncalled gene TSS or (2) it was within 1000 nt of an annotated TSS on the same strand, implicating the daTSS as a potential gene TSS.
Prior to alignment, Pol II ChIP-seq reads from MCF-7 cells [23] (see also Supplemental Table 1) were filtered based on quality; reads with an average Sanger quality score less than 20 were removed from analysis. Following quality filtering, Cutadapt was used to remove adapter sequences [30]. Alignment was performed using Bowtie [31]. Fragment lengths were estimated using Homer (version 4.6) [34]. From each uniquely mapped fragment, the fragment was extended based on the estimated length, and the fragment center was subsequently found.
Considering each TSS group separately, the nucleotide composition of each position in a -1000 to +999 window was determined and reported as a percentage. Logo plots were generated using Web Logo 3 considering sequences in a -5 to +5 window about identified TSSs [35].
To identify occurrences of Pol II-associated and known vertebrate motifs, FIMO (version 4.10.0) was used considering a p < 0.0001 significance cutoff and a 0-order Hidden Markov Model from promoter regions as background [36]. Publically available position weight matrices from JASPAR were used in motif identification [18]. The JASPAR POLII database (2008 version; 13 motifs) and vertebrate motifs in the JASPAR CORE database (2014 version; 205 motifs) were used for Pol II-associated and known vertebrate motif identification, respectively. Across all promoter regions, on the order of 105 Pol II-associated and 107 known vertebrate motifs were identified. Motifs and additional information are available at http://jaspar.genereg.net/.
De novo motif discovery was performed using MEME (version 4.10.0) with default parameters [37]. Sequence windows from -35 to -20 and from -5 to +5 relative to TSS positions were used. Sequences from each TSS class were combined prior to motif analysis. Logo plots were generated using Web Logo 3 after aligning identified motifs [35].
CpG island heatmaps reflect the intersection of annotated CpG islands retrieved from UCSC Genome Browser [17] with TSS-centered windows.
Sequence conservation heatmaps were generated using phyloP scores from placental mammal alignments retrieved from UCSC Genome Browser [19]. Each position in the heatmap represents the average score over all positions in a 40-bp bin for which phyloP scores were available.
GRO-cap data from K562 and GM12878 cells [7] were compared to daTSSs identified in T47D/A1-2 cells. If a given site did not have GRO-cap signal in either K562 or GM12878 cells, that daTSS was considered T47D/A1-2 specific. The list of genes with T47D/A1-2-specific daTSSs was applied to Ingenuity Pathway Analysis [38] considering experimentally observed associations over mammalian tissues and cell lines.
Nuclei were harvested from cultured T47D/A1-2 cells and digested for 5 minutes at 37°C with a range of MNase (Worthington) concentrations. Reactions were stopped by the addition of EDTA and then treated with RNase and proteinase K. Digested DNA was isolated by phenol/chloroform extraction and ethanol precipitation. Libraries were prepared using an Illumina TruSeq sample preparation kit and sequenced on an Illumina HiSeq for paired-end 50 base reads.
Publically available MNase-seq data [22] and data generated in this work were prepared in the same way. Prior to alignment, MNase-seq reads were filtered based on quality; reads with an average Sanger quality score less than 20 were removed from analysis. Following quality filtering, Cutadapt was used to remove adapter sequences [30]. Alignment of MNase-seq read pairs was performed using Bowtie [31]. From each uniquely mapped fragment, the fragment center was found. Any report of MNase-seq coverage only considers the fragment-center position.
FAIRE-seq data were collected as described previously [24]. Publically available FAIRE-seq data [24] and data generated in this work were prepared in the same way. Prior to alignment, FAIRE-seq reads were filtered based on quality; reads with an average Sanger quality score less than 20 were removed from analysis. Following quality filtering, Cutadapt was used to remove adapter sequences [30]. Alignment of FAIRE-seq reads was performed using Bowtie [31]. Aligned FAIRE-seq reads were then de-duplicated using Picard (version 1.118) [39]. FAIRE-seq fragment lengths were estimated using Homer [34]. For each uniquely mapped fragment, the fragment was extended based on the estimated length, and the estimated fragment center subsequently found. Reported FAIRE-seq coverage only considers the fragment-center position.
Over biological triplicates, total RNA was harvested using an RNeasy Kit (Qiagen) with on-column DNase treatment. RNA quality was validated by Bioanalyzer (Agilent). Paired-end strand-specific poly-A enriched libraries were sequenced on an Illumina HiSeq 2500 for 125 base paired-end reads.
Prior to alignment, RNA-seq reads were filtered based on quality; reads with an average Sanger quality score less than 20 were removed from analysis. Following quality filtering, Cutadapt was used to remove adapter sequences [30]. Insert lengths were estimated by transcriptome alignment using Bowtie [31]. Sequence alignment was then performed using TopHat (version 2.0.4) [40]. Following de-duplication by Picard [39], alignments from individual replicates were merged. FPKM values were calculated using Cufflinks (version 2.2.1) [41].
For publically available data sets (with the exceptions of MCF7 Pol II ChIP-seq data and of FAIRE-seq and MNase-seq validation data sets), read coverage files were retrieved from public depositories (S1 Table). Given that reported read densities were considered, the data processing of the original authors was effectively observed. deepTools was used to generate matrices describing the intersection of read coverage with TSS-centered genomic windows observing strand specificity when appropriate [42]. These matrices were then used to generate heatmaps.
Heatmaps consider 40-bp/40-nt bins over TSS-centered windows. Unless otherwise noted, each position in a heatmap gives the number of reads or other features overlapping with that bin. Heatmap images were generated using Partek (version 6.6) [43]. Two-dimensional plots, unless otherwise noted, consider 10-bp/10-nt bins and report average values across all TSSs considered. To test enrichment of ChIP-seq signal at daTSS positions, ChIP-seq coverage was found in a 100-bp window about all identified daTSSs. Equivalent regions were found at genes without daTSSs by selecting a 100-bp window shifted downstream of TSSs by the median observed TSS-daTSS distance (507 nts). The significance of enrichment was then calculated by Wilcoxon test comparing the two groups.
To generate the gene TSS-centered panels in S5, S6, S7 and S8 Figs, uaTSS- and daTSS-centered plots were first reflected across uaTSS and daTSS positions, respectively, to orient these plots relative to gene TSSs. These plots were then translated upstream or downstream by the median distances observed between gene TSSs and uaTSSs or between gene TSSs and daTSSs across calls made in T47D/A1-2 cells.
The data sets supporting the results of this article are available in the GEO repository, GSE74308.
|
10.1371/journal.pntd.0006089 | Analysis of the interactome of Schistosoma mansoni histone deacetylase 8 | Histone deacetylase 8 from Schistosoma mansoni (SmHDAC8) is essential to parasite growth and development within the mammalian host and is under investigation as a target for the development of selective inhibitors as novel schistosomicidal drugs. Although some protein substrates and protein partners of human HDAC8 have been characterized, notably indicating a role in the function of the cohesin complex, nothing is known of the partners and biological function of SmHDAC8.
We therefore employed two strategies to characterize the SmHDAC8 interactome. We first used SmHDAC8 as a bait protein in yeast two-hybrid (Y2H) screening of an S. mansoni cDNA library. This allowed the identification of 49 different sequences encoding proteins. We next performed co-immunoprecipitation (Co-IP) experiments on parasite extracts with an anti-SmHDAC8 antibody. Mass spectrometry (MS) analysis allowed the identification of 160 different proteins.
SmHDAC8 partners are involved in about 40 different processes, included expected functions such as the cohesin complex, cytoskeleton organization, transcriptional and translational regulation, metabolism, DNA repair, the cell cycle, protein dephosphorylation, proteolysis, protein transport, but also some proteasome and ribosome components were detected. Our results show that SmHDAC8 is a versatile deacetylase, potentially involved in both cytosolic and nuclear processes.
| Using a target-based strategy to develop new drugs for the treatment of schistosomiasis we had earlier identified Schistosoma mansoni histone deacetylase 8 (SmHDAC8) as essential for parasite development and survival in the mammalian host. Selective inhibitors of this enzyme show promise as lead compounds for drug development. However, the biological role of SmHDAC8 has not been established. We identified the potential partner proteins and the processes in which it is involved by combining two methods: yeast two-hybrid screening and co-immunoprecipitation with mass spectrometry. These approaches yielded complementary sets of potential partner proteins that are actors in about 40 different cellular processes. These include known roles for the human counterpart of SmHDAC8, like interactions with the cohesin complex and the cytoskeleton, as well as novel interactions such as with proteasomes and ribosomal proteins. Our results emphasize the implication of SmHDAC8 in both nuclear and cytosolic processes and the versatility of this enzyme, suggesting why it is vital to the parasite and a promising drug target.
| Schistosomiasis is a neglected tropical parasitic disease of major public health importance [1–3] caused by blood flukes of the Schistosoma spp., with 258 million people requiring treatment worldwide, and 780 million at risk of infection (http://www.who.int/mediacentre/factsheets/fs115/en/). There is currently no effective vaccine against schistosomiasis [4] and the treatment of the disease relies on a single drug, praziquantel (PZQ). Because of the intensive use of PZQ, field observations show the appearance of schistosome strains resistant to PZQ [5, 6]. Thus, the development of new drugs is imperative.
We have used a “piggy-back” strategy that consists in identifying orthologues of proteins already targeted in other pathologies, like cancer. Among these, we have chosen to target enzymes involved in epigenetic processes [7] and in particular, histone deacetylases (HDAC), which are among the most studied epigenetic targets. Schistosoma mansoni possesses three class I HDACs (SmHDAC1, 3 and 8) and four class II HDACs, two class IIa and two class IIb enzymes [8, 9]. We have shown [10] that Trichostatin A (TSA), a pan-inhibitor of HDACs, induces hyperacetylation of histones, deregulates gene expression and induces the death of schistosome larvae and adult worms in culture. Schistosome HDACs are therefore promising targets for the development of new drugs against schistosomiasis, especially SmHDAC8 [11, 12]. Indeed, it is the only schistosome class I HDAC for which the structure of its catalytic pocket differs significantly from that of the human orthologue [13], allowing the development of selective inhibitors that are toxic for schistosome larvae (apoptosis and death) and adult worms (changes in the reproductive organs, separation of worm pairs and arrest of egg laying) in culture [13–15]. Moreover, transcript knockdown of SmHDAC8 leads to markedly reduced parasite viability and fecundity [13] suggesting that this enzyme is essential to parasite growth and development.
Human HDAC8 (hHDAC8) catalyzes the deacetylation of lysine residues within histone and non-histone proteins but its biological role has long remained elusive [11]. The two best-characterized non-histone substrates are the estrogen-related receptor (ERRα) and the structural maintenance chromosome 3 protein (SMC3). hHDAC8 interacts directly with ERRα in vivo and deacetylates ERRα in vitro, increasing its DNA binding affinity [16]. In the case of SMC3, a member of the cohesin complex, hHDAC8 is involved in its deacetylation that allows the recycling of the cohesin complex, and hHDAC8 mutations are linked with the Cornelia de Lange syndrome [17, 18]. Two recent studies using more systematic approaches identified novel hHDAC8 substrates. The first [19] detected seven proteins all of which are nuclear proteins, including SMC3, and this approach failed to identify histones or ERRα as substrates. The second study [20] detected 19 novel hHDAC8 substrates, not all of which were nuclear. Among them, the cohesin complex component SMC1A, but also all the protein substrates identified by Olson [19] were predicted by this approach. Clearly, each approach provides different information about the hHDAC8 substrates.
In order to characterize the protein partners of the 11 human HDACs, Joshi et al. [21] performed a global study of their interactions by establishing T-lymphoblast cell lines stably expressing Enhanced Green Fluorescent Protein (EGFP)-tagged hHDACs combined with proteomics and functional studies to identify hHDAC-containing protein complexes. They showed that in these cells hHDAC8 interacts with 15 proteins: four related to the cell cycle including SMC1A and 3, three related to protein and ion transport and 8 with other/unknown functions. However, although these proteins are members of hHDAC8-containing complexes, the methodology used cannot distinguish whether or not they interact directly with the enzyme. Moreover, human acetylome analysis [22] interestingly reveals that among these partner proteins, SMC1A, SMC3, SA2, SEC16A and NUP98 are acetylated but only SMC1A and SMC3 have been identified as hHDAC8 substrates [19, 20].
SmHDAC8 shows major structural differences compared to hHDAC8, most obviously in the presence of insertions within the catalytic domain that form loops at the surface of the protein [13] and therefore represent potential surfaces for protein-protein interactions. We therefore sought to determine whether SmHDAC8 interacts with the same or different proteins than hHDAC8. Since it is not possible to overexpress a tagged “bait” protein in schistosomes we decided to use two different methods to identify protein partners, yeast two-hybrid screening (Y2H) and co-immunoprecipitation experiments coupled to mass spectrometry (Co-IP/MS). The former technique characterizes direct protein-protein interactions and the latter identifies proteins that may interact directly or are members of protein complexes that interact with the target protein. They therefore yield different results, but in combination give an overall picture of the cellular processes in which SmHDAC8 participates. Our results show that SmHDAC8 is a versatile deacetylase, potentially involved in both cytosolic and nuclear processes, and contribute to the understanding of its status as a therapeutic target.
A yeast two-hybrid (Y2H) S. mansoni adult worm (6 week-old male and female worms) cDNA library that consists of the Gal4-activation domain (Gal4-AD), amino acids 768–881, fused with S. mansoni adult worm cDNA was used. The cDNA library was constructed according to the manufacturer's instructions (Matchmaker Library Construction and Screening Kit, Clontech) using the pGADT7 plasmid containing the LEU2 reporter gene. The cDNA library was transformed into Saccharomyces cerevisiae AH109 strain containing HIS3/ADE2/LacZ reporter genes, under the conditions recommended by the supplier (Yeast Protocols Handbook, Clontech). The cDNA library was screened with a bait construct corresponding to the Gal4-DNA binding domain (Gal4-DBD) fused with the full-length coding sequence of SmHDAC8 (EF077628) [8], amplified using oligonucleotides SmHDAC8 Fw (5'-GCTCGAATTCATGTCTGTTGGGATCG-3') and SmHDAC8 Rw (5'- ACCTCGAGGATCCCATACCAGTTAAATTATA-3'), and cloned into the EcoRI and BamHI restriction sites of the pGBKT7 vector bearing the TRP1 reporter gene. The S. cerevisiae Y187 strain containing the LacZ reporter gene was transformed with the bait construct and mated with the AH109 strain overnight. After incubation, diploid yeasts were plated on selective medium lacking adenine, histidine, leucine and tryptophan and the plates were incubated at 30°C for 5 days. Positive clones were confirmed both by restreaking on selective medium and by a liquid LacZ assay (Yeast Protocols Handbook, Clontech).
Each selected positive clone was cultivated in medium lacking leucine. Cells were harvested by centrifugation (2,000 g, 20 min) and disrupted with glass beads. Plasmid extraction was performed with the Nucleospin plasmid kit (Macherey-Nagel) according to the manufacturer’s instructions. Plasmids extracted from yeast cells were then transformed in Subcloning Efficiency DH5α competent cells (Invitrogen). Prior to sequencing (Eurofins Genomics), about 400 individual clones were pre-screened by digestion with the restriction enzymes HindIII and EcoRV and electrophoresis on agarose gels in order to select clones with unique restriction profiles and weed out duplicates.
After extraction of plasmids, each clone was transformed in the AH109 strain to confirm that the interaction was robust. The S. cerevisiae Y187 strain was transformed with the bait construct SmHDAC8 pGBKT7 and mated with the AH109 strain overnight. After incubation, diploid yeasts were plated first on selective medium lacking leucine and tryptophan and then on another higher stringency selective medium lacking adenine, histidine, leucine and tryptophan and the plates were incubated at 30°C.
Sequence analyses (correction and alignment) were performed using Sequencher software (Gene Codes Corporation). Identification and functional annotation of SmHDAC8 interactors was performed using Blast 2GO software [23].
All animal experimentation was conducted in accordance with the European Convention for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes (ETS No 123, revised Appendix A) and was approved by the committee for ethics in animal experimentation of the Nord-Pas de Calais region (Authorization No. AF/2009) and the Pasteur Institute of Lille (Agreement No. A59-35009).
A Puerto Rican strain (NMRI) of S. mansoni is maintained in the laboratory using the intermediate snail host Biomphalaria glabrata and the definitive golden hamster host Mesocricetus auratus. S. mansoni adult worms were obtained by hepatic portal perfusion of hamsters infected six weeks previously [24].
Purified recombinant SmHDAC8 (a kind gift from M. Marek and C. Romier, IGBMC, Strasbourg, France [13]) was used to generate rat polyclonal antiserum. Male Lou Rats were injected i.p. with 50 μg of SmHDAC8 with alum adjuvant in a total volume of 500 μL three times at two-week intervals. The rats were bled two weeks after the final injection. The monospecificity of the rat antiserum was controlled after SmHDAC8 immunoprecipitation from S. mansoni protein extract (see section below) and western blotting (S1 Fig).
Three independent experiments were performed as follows. Adult worms (50 couples) were suspended in 500 μL of lysis buffer (20 mM Tris HCl pH 7.4, 50 mM NaCl, 5 mM EDTA, 1% Triton and protease inhibitors), crushed with a Dounce homogenizer and sonicated ten times for 30 s (maximum power, Bioruptorplus, Diagenode). After centrifugation, at 10,000 g for 10 min at 4°C, immunoprecipitation of SmHDAC8 was performed using the Pierce Crosslink Immunoprecipitation Kit (Thermo Scientific) according to the manufacturer’s instructions. Briefly, the protein lysate (500 μL) was pre-cleared by incubation with 20 μL of IgG from rat serum crosslinked to protein-L Agarose beads (Thermo Scientific) for 2 h at 4°C on a rotator. Then, pre-cleared lysate was collected after centrifugation, at 1,000 g for 1 min at 4°C, and incubated overnight at 4°C on a rotator, with 1 μL of anti-SmHDAC8 antibodies or 1 μL of IgG from rat serum as a control, bound to protein-L Agarose beads.
Protein samples were denatured at 100°C in 5% SDS, 5% β-mercaptoethanol, 1 mM EDTA, 10% glycerol, and 10 mM Tris pH 8 buffer for 3 min, and subsequently fractionated on a 10% acrylamide SDS-PAGE gel. Electrophoretic migration was stopped when the protein sample had entered 1 cm into the separating gel. The gel was labelled briefly with Coomassie Blue, and five bands, containing the whole sample, were cut out. Digestion of proteins in the gel slices was performed as previously described [25].
Separation of the protein digests was carried out using an UltiMate 3000 RSLCnano System (Thermo Fisher Scientific). Peptides were automatically fractionated onto a commercial C18 reversed phase column (75 μm × 150 mm, 2 μm particle, PepMap100 RSLC column, Thermo Fisher Scientific, temperature 35°C). Trapping was performed during 4 min at 5 μL/min, with solvent A (98% H2O, 2% ACN (acetonitrile) and 0.1% FA (formic acid)). Elution was carried out using two solvents A (0.1% FA in water) and B (0.1% FA in ACN) at a flow rate of 300 nL/min. Gradient separation was 3 min at 5% B, 37 min from 5% B to 30% B, 5 min to 80% B, and maintained for 5 min. The column was equilibrated for 10 min with 5% buffer B prior to the next sample analysis.
Peptides eluted from the C18 column were analyzed by Q-Exactive instruments (Thermo Fisher Scientific) using an electrospray voltage of 1.9 kV, and a capillary temperature of 275 °C. Full MS scans were acquired in the Orbitrap mass analyzer over the m/z 300–1200 range with a resolution of 35,000 (m/z 200) and a target value of 5.00E + 05. The ten most intense peaks with charge state between 2 and 4 were fragmented in the HCD collision cell with normalized collision energy of 27%, and tandem mass spectra were acquired in the Orbitrap mass analyzer with resolution 17,500 at m/z 200 and a target value of 1.00E+05. The ion selection threshold was 5.0E+04 counts, and the maximum allowed ion accumulation times were 250 ms for full MS scans and 100 ms for tandem mass spectrum. Dynamic exclusion was set to 30 s.
Raw data collected during nanoLC-MS/MS analyses were processed and converted into *.mgf peak list format with Proteome Discoverer 1.4 (Thermo Fisher Scientific). MS/MS data were interpreted using search engine Mascot (version 2.4.0, Matrix Science, London, UK) installed on a local server. Searches were performed with a tolerance on mass measurement of 0.2 Da for precursor and 0.2 Da for fragment ions, against a composite target decoy database (25,970 total entries) built with the S. mansoni Uniprot database (taxonomy id 6183, 12,861 entries) fused with the sequences of recombinant trypsin and a list of classical contaminants (124 entries). Up to one trypsin missed cleavage was allowed. For each sample, peptides were filtered out according to the cut-off set for protein hits with one or more peptides longer than nine residues, an ion score >30, an identity score >6, leading to a protein false positive rate of 0.8%.
The aim of our study was to identify SmHDAC8 partners in order to better apprehend its biological role and function. We used here two strategies to characterize the SmHDAC8 interactome.
The yeast two-hybrid (Y2H) screening using SmHDAC8 revealed a large number of positive clones, of which we sequenced 137 after the prescreening step. After manual correction and assembly using Sequencher, we characterized 49 different sequences, which were then identified and functionally annotated using Blast 2GO software (S1 Table).
Three independent co-immunoprecipitation (Co-IP) experiments were performed, using an anti-SmHDAC8 antibody (named IP1, IP2, and IP3). As a control, we performed Co-IP with a rat IgG antibody alone in each experiment. Mass spectrometry (MS) of the Co-IP proteins identified 1,500 different proteins (S2 Table). A significant degree of variation between the proteins identified in each experiment was noted, for which several reasons can be invoked. The three parasite protein extracts were each obtained from a pool of S. mansoni adult worms of both sexes and not from homogeneous cell cultures. Variations in protein expression between worm batches could induce differences between the three Co-IP/MS experiments. Moreover, S. mansoni is a complex multicellular parasite and protein quantities can vary between the different cellular types within a given individual as well as between different worms. Therefore, to take account of this relative variability between each extract, we chose to pool the results obtain for the three Co-IP/MS experiments IP1, IP2 and IP3. The possibility that the observed variability may have been due to non-specific interactions of our anti-SmHDAC8 antibody with other proteins can be discounted. A single band corresponding to the molecular weight of SmHDAC8 was detected on western blots of the immunoprecipitated material (S1 Fig) and the only HDAC detected by MS in the immunoprecipitates was SmHDAC8 (S2 Table). Of the 1,500 proteins for which peptides were detected we selected only those that fulfilled three criteria: (i) at least three peptides in the Co-IP experiment, (ii) with no more than two peptides in the control and (iii) with a spectral count ratio between Co-IP SmHDAC8 and control of greater than 3. After, this selection step we obtained 160 different proteins that were considered good candidates as SmHDAC8 partners (S2 Table).
Among the proteins identified by these two approaches, four are common between Y2H and Co-IP/MS: the Proliferation-associated protein 2G4, 38kDa (PA2G4, n°G4LXR6), Cathepsin-B1 (SmCB1, n°Q8MNY2), putative NADH-ubiquinone oxidoreductase (n°G4VK53) and microsomal glutathione S-transferase 3 (GST-3, n°G4VH65) (Fig 1).
Among these proteins, two illustrate the previously characterized roles in HDAC-dependent processes and/or the potential involvement of acetylation. (i) PA2G4 is highly conserved in eukaryotes. The human member of this family, ErbB3 binding protein 1 (Ebp1) was identified as a putative downstream member of an ErbB3-regulated signal transduction pathway [26]. More particularly, the C-terminal region (300–372) of Ebp1, which is important for transcriptional repression, was shown to bind HDAC2 and inhibitors of HDACs significantly reduced Ebp1-mediated transcriptional repression [27, 28]. Like its human counterpart, the interaction between PA2G4 and SmHDAC8 is mediated by its C-terminal moiety, corresponding to the fragment cloned in the Y2H screen. (ii) SmCB1 is an essential gut-associated peptidase that digests host blood proteins as a source of nutriments (note that we also detected, but only with the Co-IP/MS experiments, cathepsin D and L2 which are also part of the gut peptidase network) and is also a drug target because enzyme inhibition induced severity phenotypes in the parasite [29]. Moreover, SmCB1 is a promising vaccine candidate because its administration elicits protection against S. mansoni challenge infection in mice and hamsters [30, 31]. A direct link with HDACs has never been reported. Nevertheless, acetylome analysis from Schistosoma japonicum reveals that Cathepsin-B (n°Q7Z1I6) is acetylated on K241 and sequence alignments between SjCB1 and SmCB1 show that it is conserved.
The lack of convergence between Y2H screening and Co-IP/MS is perhaps not entirely surprising since they represent very different methods for interactome studies. Both strategies have weaknesses. In Y2H screening, direct interactions between two proteins are detected, but in some cases, these may be non-specific due to the juxtaposition of proteins or fragments that are never in contact within the cell. In the case of Co-IP/MS, some proteins cannot be identified because they are present in low quantities in the parasite extract and some proteins identified may be members of immunoprecipitated complexes and are not direct partners. The results obtained can therefore be considered as complementary and, taken together, provide an overall picture of the cellular processes in which SmHDAC8 is involved.
For some of the partner proteins identified only by Y2H and not by Co-IP/MS screening we carried out independent experiments to verify the interaction with SmHDAC8. For instance, the binding of SmCtBP, SmMBD2, tensin and actin-1 proteins to SmHDAC8 was verified by candidate-specific Y2H experiments. As expected, SmHDAC8 indeed interacted with SmCtBP and SmMBD2, as well as the other two proteins (Fig 2). These results are in agreement with available data for the human orthologues. (i) The human C terminal binding protein (CtBP) family members appear to mediate transcriptional repression in a histone deacetylase (HDAC)-dependent manner [32]. Some human class I HDACs, HDAC 1, 2 and 3 [33–35] and class II HDACs (HDAC4 and 5) [36] are present in the CtBP1 nuclear protein complex, but the possible involvement of hHDAC8 was not investigated. (ii) The methyl-CpG binding domain protein 2 (MBD2) binds to methylated DNA and represses transcription through the recruitment of NuRD co-repressor complex [37]. The MBD2-NuRD complex contains a histone deacetylase core composed of HDAC1/2, RbAp46/48 and MTA2 [38–41]. The MBD2 protein possesses, among others, an intrinsically disordered region (IDR) able to recruit RbAp48, HDAC2 and MTA2 [42]. More particularly, it is possible that HDACs bind directly to the MBD2 IDR, because the human HDAC interactome study reveals a specific interaction between MBD2 and HDAC1/2 [21]. The Database of Protein Disorders (DisProt, www.disprot.org) [43] predicts an IDR domain located between the MBD and the C-terminal coiled-coil domains of SmMBD2, which may be involved in SmHDAC8 binding.
Focusing on the biological processes predicted with the blast2GO software for each protein identified, among the 49 partners identified with the Y2H screening (S1 Table), seven were of unknown function, although one encoded a peptide including an EGF-like domain and an IgG-like domain. One further sequence corresponded to a hitherto unannotated gene. The remaining proteins are involved in 21 different biological processes (S1 Table), and those represented by the largest numbers of different sequences are cytoskeleton organization, transcription regulation, metabolism, transport, cell cycle regulation, DNA repair and chromatin remodeling (Fig 3).
The 160 proteins identified by Co-IP/MS (S2 Table) are involved in 33 different biological processes, and those represented by the most different sequences are metabolic process, cytoskeleton organization, proteasome, proteolysis, translation regulation, transport, protein dephosphorylation and stress response (Fig 3).
These results show that some of these processes (22%) are common between our different experiments, and in particular metabolism and cytoskeleton organization (Fig 3). Moreover, they suggest that SmHDAC8 is a versatile deacetylase involved in both cytosolic (cytoskeleton organization, ribosome, proteolysis or proteasome) and nuclear (transcription regulation, DNA binding and repair, chromatin remodeling or cohesin complex) processes. This contrasts with the current knowledge of partners and substrates proteins of human HDAC8 (hHDAC8) that emphasizes nuclear functions. The global interactome study [21] that identified 15 partner proteins and the recent systematic studies of hHDAC8 substrates [19, 20] mainly identified nuclear partner proteins involved in the cohesin complex, transcriptional regulation and chromatin remodeling. Some extranuclear proteins were identified as partners, such as the Ca2+-dependent phospholipid binding protein Copine III [21] or substrates, such as the elongation factor 1 EF1α1 [20] but these were in the minority.
Here, the involvement of SmHDAC8 in extranuclear functions is illustrated by the direct interaction between SmHDAC8 and proteins involved in cytoskeleton organization like tensin, actin-1, actin 5c and Rho1 GTPase (S1 and S2 Tables). As previously mentioned, the binding of tensin and actin-1 proteins to SmHDAC8 was verified by candidate-specific Y2H experiments. As expected, SmHDAC8 indeed interacted with SmActin-1 and SmTensin (Fig 2). Interestingly, Waltregny et al. have shown that hHDAC8 interacts with smooth muscle alpha actin (but not beta-actin) to regulate cell contractility [44, 45]. Proteomic analyses have shown that all three human actin isoforms (alpha, beta and gamma) are acetylated [22, 46]. Moreover, the actin-associated proteins cortactin [47] and anillin [20] are hHDAC8 substrates. Several regulatory proteins of actin polymerization in human are also acetylated (gelsolin, CapZ, profilin and the Arp2/3 complex) [22] and we found some orthologues in our SmHDAC8 Co-IP/MS analyses (gelsolin, septin and a subunit of the Arp2/3 complex, S2 Table). Actin dynamics are also controlled by small GTPases of the Rho family, like RhoA, notably in the formation of focal adhesions and stress fibers [48–50]. RhoA was not identified by Joshi et al. [21] as interacting with hHDAC8 and does not seem to be acetylated in human [22], although it was found to be acetylated in Schistosoma japonicum [51]. Here, our Y2H screen shows the direct interaction between SmHDAC8 and SmRho1, the orthologue of human RhoA. Hence, it is possible that binding of SmHDAC8 to SmRho1 could participate in the control of the SmRho1 pathway.
Among all the hHDAC8 substrates and partners identified [19–21], only Sec proteins, involved in endoplasmic reticulum protein secretory pathways were also found to interact with SmHDAC8, although these are different and phylogenetically distinct proteins: Sec16A in human and Sec1 and Sec61β in S. mansoni (S2 Table).
Strikingly, we found no orthologues of the major hHDAC8 substrates/partners already characterized among the protein partners of SmHDAC8. This is exemplified by members of the cohesin complex, which is notably responsible for the correct separation of sister chromatids into two daughter cells during mitosis [52]. The tripartite cohesin ring is formed by two SMC (structural maintenance of chromosome) proteins SMC1A and SMC3, and the α-kleisin subfamily protein RAD21 [53–55]. Both SMC3 and SMC1A were found in the hHDAC8 interactome [21], as is SA2 (an accessory protein that binds to hRAD21). In their study, hRAD21 was also detected, but at a non-significant level. Unexpectedly, we did not identify any of the cohesin complex members present in S. mansoni using our Co-IP/MS strategy contrary to Joshi et al. One possible explanation is that they used cells overexpressing EGFP-tagged hHDACs that generated a higher quantity of hHDAC8 protein than is present under physiological conditions. Correct MS identification depends strongly on the quantity of immunoprecipitated protein, and unfortunately, we are unable to increase SmHDAC8 expression in S. mansoni. However, and interestingly, our Y2H screen shows a direct interaction of SmHDAC8 with SmRAD21. In order to confirm the Y2H screening result we carried out a candidate-specific Y2H experiment with the SmRAD21 clone and, as expected, SmHDAC8 interacts with SmRAD21, suggesting that SmHDAC8 forms an integral part of the cohesin complex and may have a central role in cell division, transcriptional regulation and DNA repair.
In conclusion, the protein partners of SmHDAC8 identified by our Y2H library screen and Co-IP/MS experiments are all orthologues of proteins not previously identified as substrates or partners of human HDAC8. While we did not necessarily expect to detect substrates using these methodologies, it is striking that none of the partner proteins detected by Joshi et al. were among the SmHDAC8 partners. We propose that the principal reason for this is that the methodology used in the previous study, with tagged HDACs (as bait to pull down protein complexes) expressed in a T-lymphoblast cell line, may address a limited subset of potential partners. In contrast, our Co-IP/MS study was performed on adult worm protein extracts and S. mansoni is a complex parasite encompassing a variety of cell types.
These Co-IP/MS approaches do not necessarily identify the proteins to which the HDACs physically bind because some proteins identified can be part of immunoprecipitated complexes. This is illustrated by the cohesin complex members identified by Joshi et al. (SMC1A, SMC3, and SA2) and that do not include RAD21, which only appeared as a non-significant binder. This is the reason why we also performed an Y2H screen using SmHDAC8 in order to identify proteins that bound directly to SmHDAC8. Our Y2H shows that SmHDAC8 physically binds to SmRAD21 and we suggest that the other cohesin complex components are pulled down by this interaction. Against this, it can be argued that the Y2H screen can identify non-specific interactions due to the juxtaposition of proteins or fragments that are never in contact within the cell. Again, this is why we have followed up the results of the initial screen with more detailed investigations of the interactions of two of the most interesting candidates, SmRAD21 and SmRho1, allowing us to confirm that both are bona fide partners of SmHDAC8.
Despite the fact that we did not identify the same protein partners as Joshi et al., it would be unwise to assume that all those identified in the Y2H screen and in the Co-IP/MS experiment are specific for the schistosome enzyme. It is probable that a number of their human orthologues will interact with hHDAC8 when investigated individually. Nevertheless, given the structural differences between the schistosome and human enzymes, and particularly the unstructured loops at the surface of SmHDAC8, encoded by insertions in the catalytic domain sequence, it can be assumed that specific partners of the latter are present. Some of the proteins we identified have no human orthologues and have unknown functions. Further work will determine which of the partner proteins are schistosome-specific, how they interact with SmHDAC8 and the role of these interactions within the parasite.
|
10.1371/journal.pbio.1002015 | A Nutrient-Driven tRNA Modification Alters Translational Fidelity and Genome-wide Protein Coding across an Animal Genus | Natural selection favors efficient expression of encoded proteins, but the causes, mechanisms, and fitness consequences of evolved coding changes remain an area of aggressive inquiry. We report a large-scale reversal in the relative translational accuracy of codons across 12 fly species in the Drosophila/Sophophora genus. Because the reversal involves pairs of codons that are read by the same genomically encoded tRNAs, we hypothesize, and show by direct measurement, that a tRNA anticodon modification from guanosine to queuosine has coevolved with these genomic changes. Queuosine modification is present in most organisms but its function remains unclear. Modification levels vary across developmental stages in D. melanogaster, and, consistent with a causal effect, genes maximally expressed at each stage display selection for codons that are most accurate given stage-specific queuosine modification levels. In a kinetic model, the known increased affinity of queuosine-modified tRNA for ribosomes increases the accuracy of cognate codons while reducing the accuracy of near-cognate codons. Levels of queuosine modification in D. melanogaster reflect bioavailability of the precursor queuine, which eukaryotes scavenge from the tRNAs of bacteria and absorb in the gut. These results reveal a strikingly direct mechanism by which recoding of entire genomes results from changes in utilization of a nutrient.
| Ribosomes translate mRNA into protein using tRNAs, and these tRNAs often translate multiple synonymous codons. Although synonymous codons specify the same amino acid, tRNAs read codons with differing speed and accuracy, and so some codons may be more accurately translated than their synonyms. Such variation in the efficiency of translation between synonymous codons can result in costs to cellular fitness. By favoring certain coding choices over evolutionary timescales, natural selection leaves signs of pressure for translational fidelity on evolved genomes. We have found that the way in which proteins are encoded has changed systematically across several closely related fruit fly species. Surprisingly, several of these changes involve two codons both read by the same tRNA. Here we confirm experimentally that the anticodons of these tRNAs are chemically modified—from guanine to queuosine—in vivo, and that the levels of this modification in different species track the differences in protein coding. Furthermore, queuosine modification levels are known to change during fruit fly development, and we find that genes expressed maximally during a given developmental stage have codings reflecting levels of modification at that stage. Remarkably, queuosine modification depends upon acquisition of its precursor, queuine, as a nutrient that eukaryotes must obtain from bacteria through the gut. We have thus elucidated a mechanism by which availability of a nutrient can shape the coding patterns of whole genomes.
| Because the genetic code maps the 64 possible nucleotide-triplet codons to only 20 amino acids and three stop signals, proteins can be coded in multiple ways in the genome using different sets of synonymous codons. Despite specifying the same amino-acid sequence, the particular coding employed can alter fitness, sometimes dramatically [1], resulting in the highly nonrandom codings found in extant genomes [2],[3]. Although selection has long been thought to only weakly shape variation in synonymous codings, very recent evidence from Drosophila indicates a much stronger potential role for selection [4].
Selection acts in a host of different ways to constrain the evolutionarily viable set of protein codings, with most constraints imposed on aspects of gene expression. The charged tRNA molecules that physically embody the genetic code, bearing a triplet anticodon on one end and an amino acid at the other, read codons with differing speed and accuracy [5],[6] arising from their cellular abundances and kinetic properties.
Recent work has uncovered an ever-multiplying panoply of potential mechanisms by which codon choice alters fitness. Codon choice influences the stability of mRNA secondary structures [7],[8] and reduced stability associates with higher protein production, consistent with higher rates of translational initiation [9],[10]. Slowly translated codons may induce ribosomal pauses necessary for proper protein folding and targeting [11], or regulate the entry of ribosomes into coding sequences in ways which limit jamming [12]. Adding to the complexity are mechanisms which constrain synonymous codon choice due to pressures on other processes, such as mutational biases and selection for efficient splicing [13].
All of these effects remain limited in their ability to explain the biased use of certain codons over their synonyms at the genome scale [14]. Two mechanisms remain dominant: selection on translational speed, and selection on translational accuracy.
Across widely diverged bacterial species, shorter generation times correlate with increases in total tRNA and ribosomal RNA (rRNA) copy number and elevated preferential usage of particular codons in high-expression genes [15]–[17]. These trends constitute evidence for selection acting to speed ribosomal transit across transcripts. Increased speed reduces the density of ribosomes on transcripts, thus raising the proportion of unbound ribosomes, which accelerates the translation initiation rate and, finally, overall protein production rate [18]. Consequently, selection for increased growth rate favors coding sequences that cause rapid elongation rates.
Evidence for selection on speed remains sparse in multicellular organisms [19], and recent work has failed to find systematic codon-dependent ribosome velocity differences correlated to codon usage [19]–[21]. Selection on speed may be of reduced importance for animals [17], whose developmental processes sharply reduce the coupling between fitness and the cell doubling rate.
By contrast, natural selection to improve translational accuracy has been demonstrated in organisms ranging from bacteria to humans [22]–[24]. Amino acid errors at the ribosome, estimated to occur in roughly one out of every five average-length proteins [25], may cause loss of function [22] or cytotoxic misfolding [24],[26]. Consequently, coding sequences which reduce such errors, and reduce their impact on folding and function, will be favored by selection. Selection against mistranslation-induced misfolding suffices to generate major patterns of accuracy-driven codon usage observed from bacteria to humans [24].
Akashi introduced a clever method to isolate selection on translational accuracy [22], which has since been widely applied [23],[24],[27]. Akashi's test quantifies the tendency of particular codons, such as those corresponding to abundant tRNAs, to be found encoding amino acid sites that are sensitive to substitution, such as those conserved over evolutionary time, where errors in translation are likely to be most costly [22].
The use of tRNA abundance estimates to predict which codons will be most efficiently translated has become commonplace. A standard approach predicts tRNA abundances from modestly correlated but more readily measurable genomic tRNA gene copy numbers [12],[28],[29], and designates codons “optimal” or “preferred” if they are predicted to be read by the most-abundant tRNAs. However, tRNAs are heavily chemically modified, often in the anticodon [30], making assignments of which tRNA reads which codon nontrivial. As a relatively well-known example, a eukaryotic tRNA with a genomically encoded anticodon 5′-AGC-3′ might be naively predicted to bind and read the alanine codon GCU more readily than the synonym GCC. Instead, such tRNAs generally have their 3′-adenine modified post-transcriptionally to inosine (I) by tRNA-adenosine deaminases, yielding 5′-IGC-3′, which binds GCC more strongly than GCU [31]. Accounting for these modifications substantially improves the correlation between genomic codon usage and levels of corresponding tRNA [17],[32].
Codons corresponding to the most-abundant tRNAs are often assumed to be read more rapidly and more accurately. Consideration of the kinetics of translation, however, indicates that this need not be true [17]: codons read by high-abundance tRNAs may also be misread by high-abundance near-cognate tRNAs, reducing their accuracy [33].
These studies reveal the surprising richness of selection on protein coding across the tree of life. Both speed and accuracy selection play substantial roles, although major questions remain about the relative strength of selection on these traits [20]. Developing a mechanistic answer to the question of what determines genome-wide protein coding requires synthesizing translation kinetics, tRNA biochemistry, mutational processes, gene expression, population genetics, organism life-history traits, and systems-level pressures on organism fitness. Changes in coding between orthologous proteins are easy to find, but only between organisms that have diverged on many if not all of these contributing factors.
Here we report the discovery and mechanistic illumination of whole-genome recoding restricted to the well-studied drosophilids (Drosophila melanogaster and its relatives), in which codon choice has been thought to be highly conserved [34]. We develop a novel measure for selection on translational accuracy, based on Akashi's insight, which reveals a large-scale, phylogenetically coherent reversal in the relative accuracy-driven fitness benefit of multiple codons over their synonyms. To explain this reversal, we hypothesize that levels of a known tRNA modification in the anticodon, guanine to queuosine, change across species. We detect this quantitative change in queuosine modification directly in tRNAs of four species by electrophoretic separation, finding that modification levels vary exactly opposite published predictions. We then predict, and verify, that because queuosine modification levels change throughout D. melanogaster development, the accuracy-driven codon usage of genes expressed at different developmental stages should covary with the modification level much as they do across the phylogeny. We propose a kinetic model to explain how changes in queuosine modification suffice to reverse the relative accuracy of synonymous codons, while preserving their relative speed. Surprisingly, queuosine modification is known to be largely determined by intake of the precursor nutrient queuine, which animals solely acquire from bacteria, providing a remarkably simple pathway for nutrient availability to alter genome-scale protein coding.
Akashi's insight allows use of equilibrium frequencies of codons, and their conservation across species, to estimate the average population-scaled selective advantage due to a codon change that is attributable to translational accuracy selection (see Methods). The essential procedure, closely following Akashi's, is to estimate the population-scaled difference in fitness between a synonymous change at a conserved site (where the same amino acid is preserved in all 12 drosophilid species) compared to the same change at a variable site (all other sites) within the same coding sequence. We term this scaled selective advantage the Akashi selection score. Because it derives entirely from within-gene information, this score is unaffected by between-gene differences in expression level, function, and so on; to the extent that variation in these features affects selection for accuracy, the Akashi selection score will tend to underestimate overall selection on accuracy. However, it gains specificity for the mechanism of selection, a unique and powerful feature.
We determined Akashi selection scores for 5,182 genes present in 1∶1∶‥∶1 orthologs across 12 sequenced drosophilid species (Figure 1A), an analysis covering 13.7 million conserved and 8.2 million variable sites in D. melanogaster with similar numbers in the other species. The Akashi selection scores are symmetric, in that the benefit gained from changing one codon to another is equal to the cost of reversing this change. To facilitate comparisons between species, we orient the codon changes such that the D. melanogaster scores are all positive, indicating an accuracy benefit in changing one codon into another, and sort the rows from most-beneficial to most-deleterious (Figure 1A) or by amino acid (Figure S1). For D. melanogaster, codons conferring the greatest accuracy benefit in our mechanistically specific analysis closely match those reported to be “favored” or “optimal” in previous analyses (Table S1), confirming previous results using a structural approach to detect selection on accuracy [23].
Codon usage across 12 sequenced species of drosophilids has been examined in multiple studies, always with the conclusion that codon usage is largely stable, with D. willistoni the notable exception [34]–[36]. If codon usage were largely stable, our data would show nearly uniform benefit preserved from melanogaster across the phylogeny to virilis. But instead, a strikingly different cross-species portrait of codon usage emerges compared previous studies. The Akashi selection scores of many codon changes shift from beneficial to deleterious, and D. willistoni forms part of a coherent trend which spans the Drosophila subgenus (Figure 1A; Table S2). Among the amino acids showing the strongest and most consistent shifts, evolutionarily conserved aspartates, histidines, asparagines, and tyrosines are more often encoded by the codon NAC (where N = G, C, A, or U, respectively) in the melanogaster subgroup, exemplified by D. melanogaster itself, and by NAU in the Drosophila subgenus, exemplified by D. virilis (Figure 1B). The remaining two-codon C/U-ending families, encoding cysteine and phenylalanine, shift modestly or not at all (Figure 1B). These results therefore expose a previously unreported cross-species shift in codon usage linked specifically to selection on translational accuracy.
Because tRNAs bearing a single genomically encoded anticodon read both codon synonyms in all 12 species (Table S3) [34],[37], changes in tRNA gene copy number or tRNA gene expression cannot explain the reversals. However, in D. melanogaster, as in most other organisms, the anticodons of the tRNAs that read these four amino acids are partially modified in vivo by tRNA-guanine transglycosylase (TGT) from guanosine (G) to queuosine (Q) in the 5′ (wobble-binding) position (Figure 2A) [38].
In eukaryotes, Q modification of tRNA depends on scavenging the precursor nutrient queuine from the anticodons of bacterial tRNAs [39] obtained either by feeding or from gut microbiota [40]. Despite decades of study, the function of the queuosine modification remains unclear, as do its biochemical effects. In normally fed D. melanogaster, levels of Q modification vary over the course of development [38],[41], dropping to their lowest point in third-instar larvae and peaking in adults [38]. The restriction of Q modification to the anticodon strongly suggests a role in translation. In vitro, Q-modified tRNA bound C-ending triplets more stably than U-ending triplets [31]. Structural studies report that Q-tRNA and G-tRNA have similar codon-recognition properties [42], as might be naively expected because the ribose moiety differentiating Q from G does not involve the codon-recognizing portion of the nucleoside [43] (Figure 2B). Drosophila tRNAHis injected into Xenopus oocytes translates NAU more than NAC when Q-modified and NAC when unmodified [44]. Q-modified tRNA has a higher apparent affinity for ribosomes in bacterial and eukaryotic systems [44],[45].
In pioneering work, Powell and colleagues hypothesized that elevated Q modification might explain the unusual codon usage they observed in D. willistoni under the assumption that Q-modified tRNA preferentially reads U-ending codons [46]. Measurements of TGT gene expression levels as proxies for Q modification levels produced ambiguous results [47].
We therefore employed a method to quantify Q modification levels directly starting from total RNA. Cis-diol moieties, such as the 3′ ribose of every tRNA, slow migration through gels composed of polyacrylamide covalently linked with N-acryloyl-3-aminophenylboronic acid (APB) [48]. Consequently, queuosine's additional ribose moiety (Figure 2B) slows Q-tRNA migration relative to G-tRNA, producing two bands on an APB gel [48]. This differential migration can be eliminated by oxidizing the ribose cis-diols with periodate, producing a single faster-running band [48]. We confirmed these expected effects by Northern blotting of total RNA from D. melanogaster with a probe specific to tRNATyr (Figure 3A). Subsequent quantification of Q modification in D. melanogaster tRNAHis, tRNATyr, and tRNAAsn confirmed that Q-tRNA abundance is low in third-instar larvae and rises until roughly half of these tRNAs are modified in adult flies, consistent with results of a previous study using an independent, chromatography-based method to quantify Q modification (Figure 3B) [38]. We were not able to isolate separate Q- and G-tRNAAsp bands on APB gels, likely due to a secondary mannosyl-queuosine modification [49].
We then quantified Q- and G-tRNATyr, -tRNAAsn, and -tRNAHis in third-instar larvae and adult flies of D. pseudoobscura, D. willistoni, and D. virilis, species which span the drosophilid phylogeny. Substantial differences were apparent, with D. willistoni and D. virilis showing lower levels of Q modification than D. pseudoobscura and D. melanogaster for each tRNA species. Modification levels and between-species differences were greater in adults than in larvae (Figure 3D). TGT gene expression poorly predicted modification levels (Figure S2).
Queuosine tRNA modification in adult flies, but not larvae, shows a significant and positive correlation with an accuracy-driven selective advantage favoring C- over U-ending codons, quantified by the Akashi selection score, across all codons and species (adults, Spearman r = 0.61, p<0.05, Figure 3D; larvae, r = 0.05, p = 0.86, not shown). Importantly, the relationship between Q modification and Akashi selection score in adults is positive not just in aggregate, but also within each codon family (minimum Spearman r = 0.6 for His, Asn, and Tyr, with p>0.05 due to small sample size). Indeed, variation in the characteristic level of modification in, and selection upon, each synonymous family will tend to spuriously reduce the observed relationship between Q modification and Akashi selection score. After subtracting means from each family, the overall Spearman correlation across all families is r = 0.73, p<0.01, indicating that Q modification suffices to explain more than half the variation in selection scores across species.
Based on these cross-species correlational results, we hypothesized that Q modification alters relative codon accuracies, creating a signature of selection that we can observe using Akashi selection scores. A serendipitous opportunity to test this causal hypothesis arises from the observation that levels of Q modification vary across development within a species (Figure 4A, results from White and colleagues [38]). If the level of Q modification alters relative codon accuracy, then genes expressed at their highest levels in a particular developmental stage should experience accuracy selection modulated primarily by the level of Q modification at that stage. That is, we predict that specific codon substitutions—the ones which reverse their Akashi selection scores across species as adult-stage Q modification drops—will change their selection coefficients in the same direction during D. melanogaster developmental stages where Q modification drops. To test this prediction, we determined the Akashi selection scores using non-overlapping sets of genes maximally expressed at several D. melanogaster developmental stages similar to those in the White and colleagues' study (early/late embryo, larva, pupa, adult male/female) [50],[51]. We performed statistical tests on gene sets pooled into four categories: maximal expression in embryo, larva, pupa, or adult flies.
We focused on the seven synonymous codon pairs showing the strongest changes in Figure 1: four pairs encoding amino acids read by Q-modified tRNAs, and three pairs showing shifts at least as strong (Figures 1A and 4B). As predicted, these seven pairs showed a systematic shift in Akashi selection scores from moderately positive in favor of the C-ending codon during the embryonic stage, where Q modification is elevated, to near zero or negative (favoring the U-ending codon) during the larval and pupal stage, where Q modification is lowest, rising again to strongly positive in adults, where Q modification is highest (Figure 4; Table 1). As with variation across species (Figure 3), variation in Akashi selection scores for Asp, His, Asn, and Tyr codons over development correlates with the modification levels of the corresponding tRNA species (Figure 4C, three tRNAAsn isoacceptor levels averaged) with r = 0.61, p<0.02. As noted before, differing mean levels of selection and modification in each family add unwanted noise, and also as before, subtracting means from each family yields a stronger correlation, r = 0.63, p<0.01.
Wilcoxon signed-rank tests indicated significant reductions in Akashi selection scores in genes expressed when Q is lowest (larva, pupa) compared to genes expressed when Q is highest (embryo, adult) (all four comparisons p<0.05). Comparisons between larva and pupa, and between embryo and adult, were not significant.
Most pairs individually follow the predicted pattern, though there are some exceptions. The benefit of CAC over CAU (His), for example, rises slightly from embryonic to larval stages before dropping markedly in pupa and rising again in the adult stages. Remarkably, though, we even observe reversals of relative codon accuracy selection during development: all Akashi selection scores are positive during the embryonic stage (mirroring the genome-wide average), yet more than half turn negative in the larval stage, and all but one (for Asp codons) switches sign twice between the embryo and adult stages. Asp, while always showing a benefit in favor of GAC over GAU, shows the predicted U-shaped change in selection scores mirroring the change in Q modification.
As a control, we examined the selection scores for seven codon pairs chosen to be as similar as possible to the seven shifting pairs examined above. Each of these pairs differed from one of the shifting pairs by a single nucleotide substitution, in the wobble position wherever possible (Figure 4B). These control pairs showed no significant changes across any of the developmental stages (Wilcoxon signed-rank test p>0.2 for all comparisons), demonstrating the specificity of the observed shifts in selection scores linked to Q modification.
Together, these results show the predicted shift in accuracy-driven codon usage during D. melanogaster development corresponding to changes in Q modification of tRNA. Higher levels of Q modification correspond to increasing use of C-ending over U-ending codons at sites encoding conserved amino acids. Because these predictions were made on the basis of cross-species changes in tRNA modification, and there is no other known connection between the developmental progression of D. melanogaster and the divergence of species across the phylogeny, we conclude that Q modification is likely to be the major cause of the changes in codon usage observed in both situations.
That Q modification correlates with usage of C-ending codons is perplexing because previous work in D. willistoni made the opposite prediction [47]. The idea behind that prediction was simple: assuming that G-tRNA translates C-ending codons more rapidly than U-ending codons, and observing that U-ending codons are used more frequently in D. willistoni, it makes sense to guess that Q-tRNA preferentially translates U-ending codons, and therefore that Q modification should rise in D. willistoni. So how is it possible for both G-tRNA and Q-tRNA to preferentially translate C-ending codons, but for selection to switch to favoring U-ending codons? Moreover, why do some codons that are not read by Q-modified tRNAs shift in their relative accuracy when Q modification changes?
We argue that substantial insight into these questions can be gained by examining the kinetic effects that modification may have on translational fidelity. In what follows, we present a model in which changes in Q modification alone suffice to cause the observed changes in relative accuracy, which selection would then act upon by recoding genes.
The selective changes observed in Figures 1 and 4 arise because of translational accuracy. Experimental work has established that accuracy is determined by tRNA competition [5], which can be quantified by the fraction of time a codon is translated by a cognate tRNA bearing the proper amino acid (the “right” tRNA) rather than a near- or non-cognate competitor tRNA bearing another amino acid (the “wrong” tRNA). Translation has many identifiably distinct kinetic steps, from initial binding to accommodation to proofreading to translocation, offering several places in which right and wrong tRNAs might differ and so alter their competition. Because it is not yet known in detail how queuosine tRNA modification alters any one of these steps, we concentrate on the overall rate of translation of a codon by a tRNA, which is a complex function of all kinetic steps.
The essential idea in the model detailed below is that the focal C-ending codons are translated rapidly, and mistranslated rapidly, by both right and wrong tRNAs, respectively, making their translation inaccurate unless the right tRNA is altered such that it overpowers the competitor. Q modification confers such strength. In the absence of Q modification, the U-ending codon, while read more poorly by the right tRNA, is more accurate because the competition from the wrong tRNAs is yet weaker. Thus, the presence of the modification alone is sufficient to determine which codon will be more accurately translated, and therefore favored by selection on accuracy.
To determine whether Q modification is sufficient to explain the reversal in relative codon accuracy within synonymous families, we constructed a simplified kinetic model of translation of a focal codon in which all tRNAs have the same molecular abundance (Methods) (Figure 5A; Listing S1) [52]. We focus on the asparagine codons AAU and AAC and their cognate tRNAAsn (anticodon 5′-GUU-3′, for G-tRNA, or QUU, for Q-tRNA). The competitor near-cognate tRNA is threonine tRNAThr(IGU), where I denotes inosine (see Introduction); we refer to this species as I-tRNA.
The model assumes that G-tRNA, Q-tRNA, and the near-cognate I-tRNA all have higher first-order rate constants for reading C-ending codons than for U-ending codons (Figure 5A), consistent with in vitro binding studies [31]. Q-tRNA is assumed to bind more rapidly than G-tRNA to any given codon, consistent with a higher affinity of Q-modified tRNA for ribosomes [45]. Finally, the relative rate of Q-tRNA reading C-ending over U-ending codons is assumed higher than for G-tRNA. Under these assumptions, the identity of the most accurately translated (lowest error rate) synonymous codon in a family can switch from C-ending to U-ending solely as a function of changes in queuosine modification (Figure 5B and 5C). This model generates error rates (between 10−4 and 10−3) and translation speeds (1–10 amino acids per second) matching physiological estimates (Figure 5C) [53],[54]. While the model's precise parameters are surely inaccurate, its value lies in showing that tRNA modification alone is capable of inducing an accuracy reversal under biologically plausible conditions.
The kinetic competition model offers a unique and intuitive explanation for why codons that are not normally read by a Q-modified tRNA nonetheless shift in accuracy when Q modification levels change: these codons are misread by Q-tRNA. If Q modification primarily increases tRNA affinity for ribosomes, then increased Q modification will reduce the accuracy of near-cognate codons due to misreading by Q-tRNA (Figure 5D). This accuracy reduction is detectable as a reduced Akashi selection score. Kinetically, accuracy reduction arises when we consider the inverse of the above problem: misreading of threonine ACC codons by G/Q-tRNAAsn (which would properly read AAC/AAU codons). Given the apparent codon preferences of Q-modified tRNA, we can predict that ACC and ACU will be misread more often by Q-tRNAAsn than will ACG, which is read by a separate tRNA, tRNAThr(CGU). Consistent with this prediction, ACC is deleterious relative to ACG in the melanogaster subgroup where Q modification is highest, and beneficial in D. virilis where Q modification is nearly absent (Figure 1A; Table S2). Indeed, the relative benefits of A- or G-ending codons compared to U- or C-ending synonym change similarly for six amino acids (Gly, Thr, Val, Pro, Ser, and Leu) (Figure 1A; Table S2).
Most but not all observed codon-usage shifts can be explained by this kinetic model. The major exception is isoleucine, for which the A-ending codon AUA has an accuracy benefit over AUC/AUU in virilis but a cost in melanogaster. The isoleucine codon AUA is costly relative to AUU in every developmental stage except for larva, the lowest-Q stage, where the fitness cost becomes insignificant. This change mirrors the changes in accuracy benefit of these two codons in D. virilis, the lowest-Q species in our measurements, suggesting a link to the modification which is not captured by our kinetic model.
The kinetic model predicts that, unlike accuracy, the relative speed of codons always favors C-ending codons regardless of the level of Q modification (Figure 5B and 5C). That is, speed and accuracy selection can come into conflict dependent on the modification level, where one codon is more accurately translated but less rapidly translated than its synonym. If selection for speed were strong enough for a set of genes, those genes would show little or no accuracy-driven shift. A previous analysis found that genes encoding ribosomal proteins show consistent use of C-ending codons for His/Asn/Tyr across the phylogeny, but Asp codons shift in usage from C-ending to U-ending [18]. Because selection on speed favors increased production of ribosomes, ribosomal proteins may be expected to bear strong signatures of speed selection in addition to accuracy selection, making them unusually subject to speed/accuracy conflicts. We hypothesize that in most cases the speed benefit overwhelms the accuracy cost of C-ending codons in low-Q conditions for His/Asn/Tyr, but that accuracy costs outweigh speed benefits for Asp—perhaps because Asp codon mistranslation yields products that are particularly disruptive to ribosomal assembly or function. Our hypothesis illustrates the larger principle that the outcome of speed/accuracy conflicts can be amino-acid-specific, depending upon the consequences of speed and accuracy differences for each synonymous codon.
We find that entire genomes, under pressure for both accurate and rapid translation, have been recoded to maintain translational accuracy dependent on a tRNA modification. This modification varies across development, and the coding in genes expressed at different stages depends on the stage-specific modification level. Contrary to the common assumption that certain codons are “optimal” for translational speed and accuracy, we show how particular pairs of codons can reverse their relative accuracies while preserving their relative speeds. Our results provide evidence for multiple such speed/accuracy conflicts, building on the kinetic distinction between the translational accuracy and speed of codons articulated in previous studies [17],[33]. Going further, we show that a similar modification-dependent shift occurs during the developmental process of a single species, a striking example of the plasticity of translational fidelity. These results indicate that, if a codon is to be denoted optimal for translation, it is necessary to specify what aspect of translation the codon is optimal for, and under what biological circumstances.
Many previous studies have attempted to provide explanations for why certain codons are used more frequently than others within a genome, or in particular genes. Here, we have examined related but distinct questions: why do closely related species use different codons, and use them preferentially at evolutionarily conserved sites in proteins? And how does this site-specific usage change across the developmental program? Our results do not conflict with the well-established influence of gene expression levels or tRNA abundances on codon usage bias (to choose two of several causal factors), but do indicate that existing models are incomplete in important ways. Our study provides molecular and mechanistic insights that must be incorporated into any large-scale integrated attempt to explain the evolution of codon usage within and between species.
Why were such clear, systematic, multi-species shifts in codon usage not found in previous analyses? Close examination of results in a previous study reveals that, using one analytical approach, virtually all of the same shifts we report are apparent, but were passed over in favor of other approaches to yield the conclusion that the preferred set of codons is quite constant across Drosophila [35] (cf. their Figure 2C). The approach in which results most closely match ours—analysis of relative synonymous codon usage (RSCU) in the top 10% most-biased genes as determined by their effective number of codons (ENC) [35]—has no particular mechanistic or evolutionary interpretation that differentiates it from similar approaches that gave different results. An analysis of codon usage in 69 ribosomal proteins across the 12 species reported a reversal of the most frequently used Asp codon, but not others, and argued that this change was minor and likely to be unimportant [36]. This result may reflect the restricted size or unusual constraints on the ribosomal protein-coding gene set; we argue that speed/accuracy conflicts may also explain the apparent differences between this analysis and ours. A systematic codon-usage shift in D. willistoni is well-documented [35],[47],[55], but this species appears in most analyses to be a strong outlier in its codings. Mutational biases appear to contribute to, but not fully explain, changes in codon usage in willistoni [36],[55], a conclusion our data support. A major advantage of the within-gene comparison introduced by Akashi, and exploited here, is that it controls for mutational biases that vary over large genomic regions and between chromosomes. That willistoni behaves much like related species in our analyses is consistent with the idea that mutational biases contribute to its outlier appearance, but not its codon-usage shift. Overall, it appears that previous studies have seen signs of the shifts we report, but without a mechanism-specific analytical approach, a strong control for confounding biases, and experimental knowledge of the tRNA modification, these signs failed to coalesce into a coherent picture.
Outside of the examples above, only one additional shift in codon usage has been identified in the drosophilids, a preference shift from UCC to AGC (serine) between D. melanogaster and D. virilis. This reflects small relative differences between three codons (including UCG), of six, that are all roughly equally preferred over their counterparts [35], such that the apparent change in preference is analogous to front-running athletes edging each other out rather than a fundamental change in the race. Why these changes have occurred remains unclear. Our study indicates that in terms of accuracy selection detectable using Akashi selection scores, serine codons remain quite stable, with a slight shift in the benefit of UCA relative to UCU. We identify several other accuracy-related shifts, most linked to changes in queuosine modification of tRNA, others (such as in isoleucine) less clearly so.
Our findings have implications for the recent discovery that selection on synonymous sites in the drosophilids is far stronger than previously appreciated [4]. This study concluded that standard explanations for selection on codons, such as translational speed and accuracy, could not account for this strong selection. To reach this conclusion, codons were designated optimal or non-optimal, and these assignments were assumed constant across the phylogeny (excepting D. willistoni) and over the course of development. The results here suggest all three assumptions overlook key features of codon usage in these animals: different codons can be optimal for different selective mechanisms, and the relative selective benefit of codons is not constant across the phylogeny nor across development. It may prove useful to revisit the causes of strong selective constraint on synonymous sites with a more nuanced model for how selection has acted on translation in the drosophilids.
The tRNA modification studied here, guanine to queuosine in the anticodon, has been studied for decades yet still has an unknown primary function. Given the many modifications targeting tRNA anticodons [30], we conjecture that this modification is only one of many which regulate the speed, fidelity, and possibly other aspects of translation in ways that leave evolutionary fingerprints. Our results expose multiple shifts in accuracy-driven codon usage coupled to changes in queuosine modification, many but not all of which our kinetic model can explain as consequences of the modification. We do not claim that all such shifts arise from Q modification; other factors may well contribute. However, other coordinated shifts in accuracy (such as those in isoleucine codons) may be linked to queuosine modification in ways we do not yet grasp. The parallel changes in relative accuracy of isoleucine codons between species and across development provide some evidence to suggest that our understanding of the effects of the tRNA modification is far from complete. We anticipate that further studies, population-genetic and biochemical, will deepen our understanding of the genomic upheavals exposed here.
What causes between-species and developmental variation in queuosine modification levels? Two forces are likely to be at work: regulation of Q modification by the TGT enzyme or upstream factors, and bioavailability of the precursor nutrient queuine. Several lines of evidence suggest that bioavailability provides the dominant selective force. If reduced Q modification is a regulatory effect, it should be largely independent of substrate availability. Contrary to this prediction, supplementation of free queuine to D. melanogaster third-instar larvae (the lowest-Q phase [38]) at nanomolar levels suffices to increase Q modification of tRNA several-fold [40]. Micromolar queuine supplementation leads to near-complete modification [40]. Free cellular queuine concentration is strongly positively correlated with tRNA modification level, a hallmark of a substrate-limited process [41]. Thus, substrate limitation, rather than regulation, appears to be the primary determinant of Q modification levels. Whether other species are substrate-limited for queuine like D. melanogaster remains an open question. Species-wide variation in Q modification may stem from differences in gut microbiota, consistent with the wide variation we observe in species reared on identical diets, or from host variation, such as differences in expression of the enzyme(s) responsible for liberating queuine for absorption.
Limiting queuine provides a simple explanation for the dip in Q modification during larval stages, and indeed the longstanding but poorly understood association between mitotic activity and reduced levels of queuosine tRNA modification, which is also observed in rapidly dividing cancer cells [56]. During rapid growth, such as larval development when mass (and thus tRNA content) increases more than 200-fold [40],[51], queuine intake must increase just to keep modification levels constant. If TGT is substrate-limited and the microbial sources of queuine multiply less rapidly than the growing organism, the exponential increases in tRNA abundance during rapid growth will result in transiently reduced Q modification, as observed in D. melanogaster [40].
Our results illuminate a surprising interplay between microbially acquired compounds, the fidelity of an organism's translational apparatus during development, and the evolutionary fate of its genome. Application of the general approaches introduced here to diverse taxa will likely yield more and deeper insights into this and similar novel modes of coevolutionary change.
Akashi selection scores for the 12 drosophilid species, and for genes maximally expressed at D. melanogaster developmental stages, may be accessed from the Dryad repository (datadryad.org) at doi:10.5061/dryad.1jn88 [57].
Assuming weak selection, free recombination, and evolutionary steady-state, the log proportion of codon I relative to codon J is given by ln pIC/pJC = MIJ+SIJ where MIJ is the mutational bias (the log-ratio of mutation rates from J to I) and SIJ is the population-scaled additive selective (fitness) advantage of codon J over codon I (SIJ = NesIJ with Ne the effective population size) [58].
Let the proportion of sites with codon I that encode amino acids which are conserved across all 12 species be pIC, and at unconserved (variable) sites be pIV. At sites within the same protein, then,where SIJAkashi is the Akashi selection score quantifying the population-scaled difference in selective advantage resulting from a change from reference codon J to codon I at a conserved site relative to that at a variable site in the same gene. This difference is attributable to translational accuracy. The left-hand side is the log-odds ratio for a 2×2 contingency table (conserved versus variable, codon I versus codon J) which can, given codon counts n, be estimated by ψ ˆ = ln nIC/nJC−ln nIV/nJV. Akashi pointed out that such log-odds ratios can be estimated using the Mantel-Haenszel procedure [59], which allows 2×2 tables to be computed for each gene separately and then combined into a single estimate, which, by construction, controls for all between-gene differences (such as levels of gene expression, structure, function, between-gene variation in mutational biases, and so on) which can distort other estimates of selection. With genes indexed by g and ng = nIgC+nJgV+nIgV+nJgC, the Mantel-Haenszel estimate is ψ ˆ = Σg (nIgC nJgV/ng)/Σg (nIgV nJgC/ng) with variance given by the Robins-Breslow-Greenland estimator [60]. These estimates are only approximately additive.
Selection coefficients quantify the fitness advantage of a genotype over a reference, and we take as a reference the lowest-relative-fitness codon in D. melanogaster in each synonymous family. That is, we choose the reference codon such that all synonymous changes from that codon are (in this analysis) beneficial in D. melanogaster.
We quantify Akashi selection scores for all possible pairs of synonyms (no score for single-codon families, one pair for two codons, three pairs for three codons, six pairs for four codons, and 15 pairs for six codons).
Coding sequence alignments for 12 drosophilid species were obtained for 9,855 transcripts from FlyBase [61] (ftp://ftp.flybase.org/12_species_analysis/clark_eisen/alignments/all_species.guide_tree.cds.tar.gz), and filtered to include a single transcript per gene, aligned with 1∶1 orthologs in all 12 species (ftp://ftp.flybase.org/12_species_analysis/clark_eisen/homology/GeneWise.revised.homology.tsv.gz), with a minimum fraction alignable of 50% and at least 50 codons, yielding 5,182 alignments used for all analyses.
Maximal developmental-stage expression was evaluated using published data [51], which, among the alignments above, yielded 1055, 675, 885, 502, 891, and 301 genes with maximal expression in early embryo (E0), late embryo (E3), larva (L), pupa (P), adult male (M), and adult female (F) flies, respectively. Embryo (E) genes were those with maximal expression in either E0 or E3, and adult (A) genes were those with maximal expression in M or F.
All drosophilid species were reared in bottles on standard yeast-glucose media at room temperature (approximately 23°C). RNA was extracted from third instar larvae and 2-week-old adults using the standard TRIzol (Invitrogen) protocol. For larval collection, adults were placed on fresh food for 24 hours, after which great care was used to ensure that all flies were removed. We noted when larvae first started the roaming stage. 24 hours later larvae were collected from the bottles by pouring enough 5 M NaCl to cover the media and allowing the resulting mixture to set for 5 minutes. Floating larvae were poured onto mesh and washed in water before snap-freezing in liquid nitrogen. To age adults, newly eclosed flies were transferred to fresh bottles and every few days transferred to new bottles. After two weeks the flies were snap-frozen in liquid nitrogen.
This method was based on the protocol developed by Igloi and Kössel [48]. 2.5 µg of total RNA was deacylated by incubating in 100 mM TrisHCl (pH 9) for 30 min at 37°C. The deacylated RNA was combined with an equal volume of denaturing gel loading buffer containing 8 M urea, 5% glycerol, 0.05% bromophenol blue, and 0.05% xylene cyanol. Samples were loaded onto denaturing 10% polyacrylamide gels containing 5% 3-aminophenylboronic acid (Boron Molecular) and gel electrophoresis was run at 4°C in TAE. RNA was transferred under vacuum by layering 3MW blotting paper (MIDSCI), Hybond-XL membrane (GE Healthcare), gel, and plastic wrap on a gel dryer for 2 h at 80°C. After transfer the gel was removed from the membrane by soaking in distilled water. The membrane was washed twice for 30 min each in hybridization buffer (20 mM phosphate, pH 7, 300 mM NaCl, 1% SDS), followed by incubation with 5′ 32P-labeled DNA oligonucleotide probes in the hybridization buffer for 16 h at 60°C. Membranes were washed three times for 20 min each in a solution containing 20 mM phosphate (pH 7.2), 300 mM NaCl, 2 mM EDTA, and 0.1% SDS and exposed to phosphor-imaging plates. Band intensity was quantified using software from the PhosphorImager manufacturer (Fuji Medicals).
Total RNA was first deacylated as described above. The deacylated RNA was incubated in 100 mM NaOAc/HOAc (pH 4.5) and 50 mM freshly prepared periodate (NaIO4) at room temperature for 30 min. 100 mM glucose was added and the mixture incubated for another 5 min. The mixture was run through a pre-equilibrated G25 column (GE Healthcare) to remove periodate followed by ethanol precipitation. Sample was then dissolved in the denaturing gel loading buffer.
7 µg of total RNA was incubated at 55°C for 15 min in 7% formaldehyde, 50% formamide, and 0.5× running buffer (1× running buffer is 200 mM MOPS, pH 7, 80 mM NaOAc, 10 mM EDTA). Samples were then combined with an equal volume of gel loading buffer (5% glycerol, 0.05% bromophenol blue, and 0.05% xylene cyanol), and loaded onto 0.8% agarose gels (0.8% agarose, 1× running buffer, 2% formaldehyde). After electrophoresis, the gel was washed for 15 min in distilled water and 15 min in 10× SSC. RNA was transferred by capillary blotting overnight. After transfer the RNA was cross-linked to the membrane at 70,000 µJ/cm2. The membrane was washed, hybridized, and exposed in the same manner as described for the acryloyl aminophenylboronic acid gel shift assay.
Oligonucleotide probe sequences were: tRNAHis:
5′-TGCCGTGACCAGGATTCGAACCTGGGTTACCACGGCCACAACGTGGGGTCCTAACCACTAGACGATCACGGC;
tRNATyr:5′-TCCTTCGAGCCGGASTCGAACCAGCGACCTAAGGATCTACAGTCCTCCGCTCTACCARCTGAGCTATCGAAGG;
tRNAAsn:5′-CGTCCCTGGGTGGGCTCGAACCACCAACCTTTCGGTTAACAGCCGAACGCGCTAACCGATTGCGCCACAGAGAC;
TGT mRNA:5′-CGATCCACCCARCGWATDGTVCGCTCCATRGCCTC;
Actin mRNA:5′-CTTCTCCTTGATGTCRCGNACRATTTCACGCTCAGCSGTGGTGGTGAA
We observe that the proportion of Q-tRNA rises (and G-tRNA correspondingly decreases) as C-ending codons rise in inferred accuracy relative U-ending codons. To understand this shift, we assume that (1) relevant tRNAs read C-ending codons more rapidly than U-ending codons; (2) a competitor tRNA bearing another amino acid, such that mistranslation would occur if this tRNA were accepted, also reads C-ending codons more rapidly than U-ending codons; (3) ribosomes translate cognate codons using Q-tRNA more rapidly than using G-tRNA.
We adapt a previously introduced framework to build a kinetic model [52]. Let the first-order rate constant of X-tRNA for reading NAY (Y = C or U) codons be kXY and the concentration of X-tRNA be [X]. Then, for example, the rate of translation of NAC codons by Q-tRNA is kQC[Q]. For simplicity, we model competitor tRNAs using a single “effective” tRNA with concentration [M] and reading rate constants kM*. The rate of translation of Y-ending codons is rY = kQY[Q]+kGY[G]+kMY[M], and the proportion of mistranslated Y-ending codons is ϵY = kMY[M]/rY. We denote the proportion of Q-modified tRNA as q = [Q]/[T] with total cognate tRNA concentration [T] = [Q]+[G], and further assume that the competitor tRNA is present at a concentration α times that of the cognate tRNA, [M] = α[T]. In our simulations, we choose α = 1 for simplicity. Then the error rate of a Y-ending codon (Y = C or U), as a function of the proportion of Q-modified tRNA, is ϵY(q) = kMYα/[kMYα+kQYq+kGY(1−q)]. R source code to reproduce the graphs in Figure 5 is included as Listing S1.
|
10.1371/journal.ppat.1000193 | Interferon-β Pretreatment of Conventional and Plasmacytoid Human Dendritic Cells Enhances Their Activation by Influenza Virus | Influenza virus produces a protein, NS1, that inhibits infected cells from releasing type I interferon (IFN) and blocks maturation of conventional dendritic cells (DCs). As a result, influenza virus is a poor activator of both mouse and human DCs in vitro. However, in vivo a strong immune response to virus infection is generated in both species, suggesting that other factors may contribute to the maturation of DCs in vivo. It is likely that the environment in which a DC encounters a virus would contain multiple pro-inflammatory molecules, including type I IFN. Type I IFN is a critical component of the viral immune response that initiates an antiviral state in cells, primarily by triggering a broad transcriptional program that interferes with the ability of virus to establish infection in the cell. In this study, we have examined the activation profiles of both conventional and plasmacytoid dendritic cells (cDCs and pDCs) in response to an influenza virus infection in the context of a type I IFN-containing environment. We found that both cDCs and pDCs demonstrate a greater activation response to influenza virus when pre-exposed to IFN-β (IFN priming); although, the priming kinetics are different in these two cell types. This strongly suggests that type I IFN functions not only to reduce viral replication in these immune cells, but also to promote greater DC activation during influenza virus infections.
| Influenza infection leads to a serious respiratory infection of the lung epithelium. Lying directly below the epithelial cells are immune system sentinels known as dendritic cells. These cells interact with the virus and carry parts of the virus to draining lymph nodes to activate killer T cells. In order to effectively carry out this function, DCs must perceive the presence of a virus using receptors specially adapted for this function. However, when DCs are mixed with influenza virus in the laboratory, no activation occurs because the virus produces a protein called NS1 that blocks the receptors. Yet, patients infected with influenza virus develop a strong adaptive response that leads to recovery from infection. This observation suggests that additional factors must be present that contribute to the activation of the DCs. The most likely contributor is type I interferon, a ubiquitous protein released from many cells upon exposure to virus. In this study, we mixed influenza virus with DCs in the presence of type I interferon and found that this greatly enhanced their activation. Treatment with interferon allowed the DC to bypass the block in activation mediated by the influenza NS1 protein. Our data suggest that the production of type I interferon within an infected patient may endow the DCs with the ability to fully respond to influenza virus.
| Dendritic cells (DCs) play a key role in the initiation and regulation of the immune system. They respond to various microbial stimuli by undergoing a process of activation that propels them to migrate to draining lymph nodes and endows them with the ability to efficiently activate T cells [1],[2]. The process of DC activation involves several steps including upregulation of surface markers, cytokine and chemokine secretion and the ability to leave the tissue and migrate to draining lymph nodes, and is also known as DC maturation. Depending on the nature of the stimulus maturation is signified by the up-regulation of MHC and co-stimulatory molecules, as well as the secretion of some mixture of cytokines and chemokines that may include type I interferons (IFN-α and IFN-β), IL-6, IL-12, TNF-α, IL-8, IP-10, RANTES and MIP-1β [2],[3].
In response to a viral infection, DCs can be activated by two separate pathways: a toll like receptor (TLR)-dependent and a TLR-independent pathway. The TLR-dependent pathway is made up of several different TLRs that bind specific pathogen-associated-molecular-patterns (PAMPs). TLR 3, 7/8 and 9 are the sensors for viral PAMPs recognizing double-stranded RNA (dsRNA), single-stranded RNA (ssRNA) and CpG DNA motifs, respectively [4]. These TLRs are localized to the endosome and signal via adaptor proteins to induce DC activation [5]. The TLR-independent or internal pathway primarily consists of retinoic acid-inducible gene-I (RIG-I) protein and melanoma differentiation-associated gene product (MDA-5) both located in the cytoplasm (RIG-I like receptors or RLR). RIG-I recognizes cytoplasmic uncapped 5′- tri-phosphate RNAs and MDA-5 recognizes cytoplasmic dsRNA [6].
Conventional DCs (cDCs) are considered the prototypic DCs as they are proficient at presenting antigens and activating T cells [2]. The internal pathway has been shown to play a more significant role in the activation of cDCs to RNA viruses than the TLR-dependent pathway [7],[8]. Plasmacytoid DCs are a second subset of circulating human DCs, that in contrast to cDCs, use the TLR-dependent pathways, specifically TLR7 and TLR9, for activation in response to viruses [7],[9].
Type I IFN is a critical component of the viral immune response. Its expression is highly regulated and pDCs serve as the primary producers of type I IFN in the body [10]. However, virtually all nucleated cells are capable of producing IFN and possess the IFN receptor, endowing them with the ability to respond to type I IFN [11],[12]. Type I IFN initiates an antiviral state by stimulating the transcription of over 200 IFN-responsive genes, some of which code for proteins that interfere with the ability of viruses to establish infection in the cell [13]. Important IFN response genes include MxA, IP-10, ISG54, RIG-I and PKR, among others [13],[14]. Demonstrating the in vivo importance of the type I IFN response is the observation that most successful viruses contain IFN antagonists which act to suppress the IFN pathway either at the level of IFN expression, IFN signaling or the antiviral effects of IFN-responsive proteins [15].
Influenza virus contains a potent IFN antagonist, the NS1 protein, which efficiently blocks type I IFN release from infected cells, including cDCs [16],[17],[18],[19],[20]. Moreover, the NS1 of influenza virus has been shown to block virus triggered activation of cDCs in vitro resulting in poor T cell stimulation [16],[17]. These observations are in contrast to those observed in vivo where fully mature cDCs can be identified in the draining lymph nodes of infected mice and a potent and protective immune response is generated [21]. Thus, in vivo other factors are contributing to the maturation of influenza infected DCs [22],[23]. The most likely factor contributing to the enhancement of DC maturation in vivo is type I IFN [23],[24].
Supporting this hypothesis, Pollara et. al. demonstrated type I IFN can prime cDCs to overcome a viral blockade produced during Herpes Simplex Virus (HSV) infections and Osterlund et. al. reported that pre-treating cDCs with IFN-α enhanced influenza A virus induced expression of TNF-α, IFN-α, IFN-β and IL-29 genes [23],[25]. Furthermore, mouse DCs have been shown to require type I IFN signaling in order to fully mature following infection with Newcastle disease virus (NDV) and murine cytomegalovirus (MCMV) [26],[27],[28] . Thus in addition to its antiviral effects, type I IFN may also function as an enhancer of DC maturation and may explain the discrepancy observed between the in vitro and in vivo response of cDCs to influenza virus infection. In this study, we systematically examined the influence of type I IFN on the activation profile of cDCs and pDCs in response to an influenza virus infection. We found that cDCs demonstrate a greater activation response to influenza virus when pre-exposed to IFN-β (IFN priming). Additionally, pretreatment of pDCs with IFN augments their ability to release cytokines although the priming kinetics of the two DC types differs significantly. This strongly suggests that type I IFN functions not only to reduce viral replication in cells but promote greater DC activation during influenza virus infections.
Type I IFN initiates an antiviral state in cells and inhibits viral replication [15],[29]. However, viruses differ in their sensitivity to the antiviral effects of IFN [30]. To examine the effects of type I IFN on the ability of human DCs to be infected by influenza virus, we performed a dose and time titration of IFN-β exposure in GM-CSF+IL-4 monocyte-derived DCs (hereafter referred to as ‘cDCs’). Figure 1 shows the impact of treatment with IFN-β on the replication of influenza virus as measured by qRT-PCR of influenza PR8 (PR8) viral product, NP protein. The results are expressed as the percent of the copy number for the NP gene relative to cells infected without IFN treatment. The cells were pretreated for 2, 3, 6, 12, or 24 hours with the indicated amount of IFN, after which the IFN was removed and the cells infected with virus. Virus replication was measured by qRT-PCR at 12 hours post infection (p.i.). Only pretreatment for 24 hours with the highest dose (5,000 units/ml) of IFN-β was able to completely prevent virus replication in DCs. Using the lower dose of IFN-β (50 units/ml) the impact on virus replication was relatively minor when the pre-incubation time was less than 6 hours for both cDC and pDC (Figure 1A and data not shown). Regardless of the length of pretreatment, the low dose of IFN was unable to completely inhibit virus replication. Figure 1B and 1C show the relative sensitivity of cDCs and pDCs to a three hour pretreatment with the indicated concentrations of IFN-β. The results demonstrate that IFN pretreatment reduces the ability of influenza virus to replicate but eliminates replication only with a high concentration and long incubation time.
After IFN-β treatment of human DCs we observed that genes coding for antiviral proteins such as MxA, viral sensors such as RIG-I, transcription factors like STAT1 and IRF7, and chemokines like IP-10 are upregulated (Figure 2). In these experiments cells were pretreated with the indicated concentration of IFN-β for 3 hours after which the cytokine was removed. MxA, STAT1 and IRF7 remain activated for a prolonged period after IFN is removed but mRNA for RIG-I and particularly IP-10 are quickly extinguished when the cytokine is withdrawn. In contrast, most of the other genes associated with DC maturation were not significantly upregulated by IFN treatment including IFN-α, IFN-β, IL-6, and MIP-1β (which was inhibited by IFN treatment). Gene activation was monitored at the indicated time points over a 24 hour period. Thus, IFN-β pretreatment does not result in global gene profile changes in DCs, but rather affects select genes with varying activation kinetics.
cDCs infected with PR8 virus demonstrate a minimal activation profile when compared to the profile observed after infection with viruses such as NDV or Sendai virus [17],[31]. However, cDCs that have been pretreated with a low dose of IFN-β for 3 hours prior to PR8 virus infection demonstrate a substantial increase in mRNA expression for numerous DC activation genes (Figure 3A). Viral RNA expression was moderately decreased in IFN pretreated samples, while all IFN-responsive genes tested demonstrated substantial increases above the level from IFN-β alone following infection with PR8 virus. Moreover, genes not activated by IFN showed enhanced activation following the three hour pretreatment of IFN-β and PR8 virus infection (Figure 3A). This priming effect was not limited to transcription since protein release was equivalently increased (Figure 3B).
In contrast to cDCs, pDCs are highly activated by PR8 virus infection (Figure 3C). Despite the increased basal level of pDC activation following exposure to PR8 virus, pDCs were further primed by IFN-β pretreatment to produce higher levels of mRNA and secrete more protein (Figure 3C and 3D). In conclusion, prior exposure to IFN-β promotes stronger DC activation in both cDCs and pDCs after infection by PR8 virus.
IFN-β pretreatment led to enhanced transcription and release of proteins from both subpopulations of DCs following virus infection. In order to determine whether exposure to IFN after virus infection would have a similar effect, cDCs were infected with PR8 virus at the 0 hour, and IFN-β (50 units/ml) was added at 0, 1.5, 3 and 6 hours post infection and left in the culture medium until the mRNA expression profile of the treated cDCs was analyzed 8 hours post infection. Specific viral RNA expression was inhibited by IFN-β as shown in Figure 4 with the highest inhibition of viral NP gene expression observed when IFN-β was added at the same time as the virus. Despite this reduction in viral replication, cDC priming for many genes was most enhanced at 0 hours post infection (Figure 4A) and decreased to basal levels from that point on. This priming effect was observed for both IFN-α and IFN-β genes, as well as genes IFN-responsive and IFN-independent (Figure 4A). Consistent with the mRNA expression patterns, similar results were observed at the protein level (Figure 4B).
When pDC were tested for priming by type I IFN after viral infection, we observed a similar trend but smaller magnitude to that seen with cDCs. Priming was minimally seen only at the early time points for IFN-α and IP-10 and the effect diminished when interferon was added at later time points (Figure 4C and 4D). These data argue that the enhancing effect of IFN-β on DC activation occurs also if it is given immediately after virus infection, but decreases as time after infection increases.
In order to determine the optimal time of IFN-β pretreatment needed to maximize DC activation, cDCs were pretreated with IFN-β (50 units/ml) for several intervals between 24 and 0.5 hours, prior to a 12 hour PR8 virus infection. Pretreatment with IFN-β for 1.5–6 hours led to optimal expression of DC activation genes and proteins in cDCs (Figure 5A and 5B). Priming occurred for IFN-α, IFN-β, IFN-responsive genes, and IFN-independent genes (Figure 5A). Surprisingly, cDCs incubated in IFN-β for 12 hours or longer became less responsive to the priming effect as compared with the shorter time points (Figure 5A). This may somewhat reflect the reduced replication of the virus after the prolonged pretreatment.
Due to cell number limitations and cell viability issues; the time course of pDC exposure to IFN-β was shortened (Figure 5C and 5D). Similar to the results seen in cDCs, virus replication was inhibited best in cells exposed to IFN for the longest interval. As a result of the shorter kinetics utilized with pDCs it is difficult to ascertain precisely the optimum pretreatment time, however, it is clear that pretreatment with IFN enhances the response of pDCs to influenza virus infection over a broad time range. Protein secretion from both cell types confirms the priming effects observed in RNA expression in cDC and pDCs (Figure 5B and 5D).
DCs do not get productively infected with influenza virus though the virus causes an abortive infection with viral message synthesis peaking at between 6–8 hours [32]. To determine the time points where the synergy between the viral and IFN triggering activity is maximal, a time course of infection was performed. cDCs, following a 3 hour pretreatment of IFN-β (50 units/ml), were infected with PR8 virus and samples were collected and analyzed for gene transcription and protein secretion at time points beginning at 0 hours and ending at 10 hours (Figure 6A). Viral mRNA levels show that the peak of viral replication occurs between 6 and 8 hours but were reduced in the IFN treated cells at all time points. Early stimulation of transcription can be observed for the IFN responsive genes but they are not enhanced by simultaneous infection at early time points. However beginning at 4–6 hours after infection the synergistic effect of IFN and infection is seen and correlates with maximal viral gene transcription (Figure 6A).
Due to pDC cell number limitations, the time course of infection for pDCs was shortened (Figure 6B and 6D). Similar to the RNA expression trends in cDCs, viral replication was reduced at all time points assayed in IFN+PR8 samples as compared to PR8 samples. In contrast to cDCs, the synergistic effect of virus and IFN treatment was observed earlier with pDC than with cDCs. This was true with both IFN-responsive and IFN-independent genes. The different kinetics observed with concomitant IFN treatment and infection most likely reflects different activation mechanisms used by the DC subtypes. pDC can be activated through a TLRs mechanism independent of virus replication, while cDCs signal by the virus replication dependent RLR pathway.
Both DCs subsets, show similar results at the level of protein secretion (Figure 6C and 6D). Regardless of the protein type, all cytokines and chemokines tested demonstrated an increase in secretion levels with time (Figure 6C and 6D).
To determine if the IFN-β priming was unique to live virus responses, the robustness of priming was compared between live influenza virus and several TLR ligands; in cDCs UV-inactivated virus, poly (I∶C) (TLR 3 ligand), CL-075 (TLR 7/8 ligand), and LPS (TLR 4 ligand) were used and in pDCs Gardiquimod (TLR 7 ligand), CpG, (TLR 9 ligand) and UV-inactivated PR8 virus were used. The DCs were pretreated for 3 hours with the low dose of IFN-β (50 units/ml) and treated with TLR ligands for between 0–12 hours. For each ligand, both dose and time courses were performed and the time point with the greatest priming (the largest differences between samples treated with IFN-β and TLR ligand compared to the other conditions) was determined. Table 1 represents the robustness of priming, as determined by the fold increase of mRNA expression of IFN+TLR ligand over the amount of expression from the IFN alone sample and TLR alone samples [IFN+TLR ligand sample / (IFN alone sample+TLR alone sample)] (Table 1, top part). Contrary to the significant IFN-β priming observed when cDCs were infected with live virus, only small differences were seen in cDCs mRNA expression or protein secretion levels (data not shown) regardless of exposure to IFN-β prior to TLR ligand addition (Table 1, top part). This demonstrates that IFN-β priming in cDCs may be unique to live virus and or activation by RLRs.
In contrast to the cDCs, pDCs demonstrated significant priming with the TLR 7 ligand, Gardiquimod (Gard) for most genes examined and to a lesser degree with UV-inactivated virus and CpG DNA (Table 1, bottom part, and Figure S1). The priming of pDCs with TLR 7 ligand is consistent with TLR 7 being the primary influenza viral sensor in the cell [7],[9]. This increase in mRNA expression was consistent with protein secretion levels (Figure S1 and data not shown).
Overall, these data suggest that there are differences between type I IFN priming in the DC subsets that follow with their pathways of viral activation. In cDCs, activation by TLR agonist is not significantly enhanced by IFN pretreatment, while in pDCs, substantial enhancement is seen with the appropriate TLR ligand.
IFN priming clearly enhances cDC activation suggesting that it may play an important role in the initiation of immunity. This function could likely be used to overcome viral immune inhibitors such as the NS1 protein from influenza virus that has been demonstrated to inhibit cytokine secretion and maturation in both mouse and human DCs. Therefore, we compared activation of cDCs by NS1 deficient PR8 virus (ΔNS1) to DCs infected with PR8 virus after a 3 hour pretreatment with 50–5,000 units/ml of IFN-β. Figure 7 demonstrates cDCs primed with IFN respond to wild type influenza virus at intensities comparable to an influenza virus lacking the IFN antagonist (ΔNS1). IFN priming can rescue the response to influenza for all cytokines and chemokines tested with the exception of IFN-β and TNF-α (Figure 7). These genes never reached the levels of ΔNS1 at any time point tested (0–12 hours) after 3-hour IFN-β pretreatment (data not shown). This may reflect differences in expression requirements for these proteins. Our data indicate that the priming effects of IFN-β counteract the inhibitory effects on DC activation genes induced by the influenza virus NS1 protein. Interestingly, IFN-β and TNF-α were still inhibited in the presence of the NS1 protein even when DCs were pretreated by IFN-β. While the transcription of both IFN-β and TNF-α is strongly dependent on NF-kβ activation other genes that are not so strongly dependent on this transcription factor can be induced by IFN-β treatment even in the presence of the influenza virus protein NS1 [19],[33].
Type I IFN has broad antiviral, immunological effects. It has been shown to impact NK and cytotoxic T cell elimination of virally infected cells, DC cross-presentation of viral antigens and B cell antibody production and isotype switching [34],[35]. Additionally, IFN-α/β has been found to alter pDC migration, development and maturation [36],[37],[38],[39]. However, the impact of IFN-β pretreatment on human DC activation by influenza virus infection had not been fully explored.
Osterlund and colleagues initially described an effect of IFN priming on DC responses to influenza virus [25]. In their studies they showed that pretreatment with IFN could enhance mRNA for type I and type III IFN and TNF [25]. These experiments were performed using high MOI of virus and did not show secreted protein data, leaving open the question of physiological relevance. In the current work, we have comprehensively examined the impact of type I IFN on the activation profiles of both subpopulations of DCs in the context of influenza virus infections and we demonstrate that IFN-β can potently enhance their response to virus induced activation in a dose and time dependent manner. Our data show that the priming effects of type I IFN on DCs impact both the levels of mRNA expression of IFN-responsive genes and the degree of viral replication. At all concentrations and time points explored, the low dose of IFN-β was able to impair viral replication but not able to completely eliminate this replication in DCs. This incomplete shut off may be necessary to allow DCs to be activated by the viral infection.
The novel question explored here was how DCs would respond to an influenza virus infection when they had been also exposed to type I IFN. In the context of a virus infection in vivo, it is very likely that epithelial cells may secrete type I IFN that can reach underlying DCs before the virus does. If the antiviral state had been initiated prior to or post infection, would DCs be activated by the viral infection or would the antiviral state block viral DC activation? Our results clearly demonstrate that both DC subsets are not only not impaired in their response to virus infection after exposure to type I IFN, but are primed by IFN-β, having increased activation following infection with an influenza virus.
The poor response of cDCs to wild type influenza virus infection in vitro is in contradiction to the immunological outcome of natural infection in vivo, since both humans and mice generate strong adaptive immunity and are able to clear influenza virus infection. Thus, DC activation must occur in vivo. Our data suggest that IFN priming may account for the ability of a host to respond to an infection that does not appear to elicit DCs activation in vitro. IFN priming could be a mechanism for the host to overcome the powerful ability of IFN antagonists such as the influenza NS1 protein to block IFN production, signaling and/or IFN-responsive genes actions. This has broad implications for the role of DC activation in the context of an antiviral immunological response. As shown by our data, very little viral replication is needed to elicit strong DC activation. This is in sharp contrast to cDC activation from viral infection in the absence of type I IFN, which is weak and viral dose dependent.
Our results suggest that pDCs also benefit from IFN signaling. Type I IFN has previously been shown to influence pDCs development [36], while in our studies we demonstrate that IFN has a substantial impact on the activation of pDCs following influenza virus infection. The importance of pDC activation, similar to the results of cDCs, is that despite very little viral input and replication, pDCs respond fully. This ability of pDCs to produce such large amounts of type I IFN with such small viral input, may be reflective of the role of pDCs during a natural infection. PDCs may be a host equivalent to IFN primed cDCs, in the sense that pDCs are not sensitive to the inhibitory effects of influenza virus IFN antagonist, the NS1 protein.
Despite the many similarities in IFN-β priming between the two subtypes of DCs, there were several important differences. cDCs demonstrated later priming kinetics with the majority of priming corresponding with viral replication. This delayed priming suggests several possible mechanisms. Priming may occur after 4 hours simply because input virus was not able to stimulate activation, and viral replication was necessary either to increase the amount of stimuli or to produce stimuli in a recognized structure. Another hypothesis for the late priming is that it occurs as a result of increased expression of IFN-responsive genes. One of the most likely proteins to account for cDC priming would be RIG-I, which is necessary for DC activation to influenza viruses [7]. The significance of crucial IFN-responsive genes acting as viral sensors, rather than other proteins involved in DC activation, like IRFs, is that IFN does not prime cDCs responses to TLR ligands (Table 1). This supports the notion that IFN priming in cDCs is augmenting the internal pathway of activation, most likely mediated by RIG-I. However, these two hypotheses of cDC priming are not mutually exclusive; and we propose that both may occur simultaneously. IFN priming in cDCs is dependent on viral replication being sensed by the RLR pathway and due to the increased expression of IFN-responsive genes like RIG-I, this internal pathway is able to stimulate a stronger cDC activation profile.
In contrast to the delayed cDC priming, pDCs demonstrate priming most substantially at 4 hours post infection and priming decreases with time. Again differing from cDCs, pDC activation did not follow the viral replication time course, suggesting a very different mechanism of priming than in cDCs. This finding is consistent with the profile of pDC activation by viruses being predominantly TLR dependent. In our experiments IFN-β priming in pDC was independent of viral replication and seen with both live virus and TLR ligand activation (Figures 3–6, and Table 1, bottom part). These results suggest that although IFN-β treatment did not enhance the TLR pathway in cDCs, IFN-β can enhance the overall activation within cells that utilize the TLR pathway as its primary viral sensor.
Lastly, the results from both pDCs and cDCs with IFN-β added post infection, demonstrate that while priming occurs over a broad time range, there is a point where the virus ‘wins’ and the enhancing effects of IFN-β treatment are not able to supplement the DC activation. It is possible that the viral sensors are made too late to be useful or they may not be made at all due to the inhibition of cellular machinery by the virus.
In summary, type I IFN priming overrides the inhibitory effects of viral antagonists on DC activation by eliciting strong responses in cDCs and even stronger responses in pDCs. The significance of this finding suggests the importance of evaluating DC responses in an environment similar to that in vivo. As DCs in vivo are responding to viruses in the context of setting that may contain multiple pro-inflammatory cytokines and chemokines, the effects of this environment cannot be disregarded. When evaluating a DC response, it is important to consider the actual stimuli the cells may have been exposed to prior to viral infection. Moreover, here we show that the establishment of antiviral state by type I IFN does not inhibit DC activation but rather, exerts priming effects, allowing for a more efficient detection and stronger response. Our data have important implications for the understanding of the initiation of immunity in the infected host, since differences in the micro-environment of the infected DC may account for different outcomes in adaptive immunity.
Influenza virus PR8 (H1N1) was grown in 9-day-old embryonated chicken eggs (SPAFAS; Charles River Laboratories). PR8 was titrated on MDCK cells by detection of hemagglutination (HA) activity in the supernatants after 48 h of infection, as previously described and by immunoflourescence, using a monoclonal antibody, PY102, specific for the HA protein (obtained from Jerome L. Schulman). All virus infections were performed in infection medium (Dulbecco's modified Eagle's medium, 0.35% bovine serum albumin, 0.12% NaHCO3, 100 µg/ml penicillin-streptomycin). For influenza virus titrations, 2.5 µg/ml trypsin was included in the infection medium.
MDCK and Vero cells were grown in tissue culture medium (Dulbecco's modified Eagle's medium [Invitrogen] with 10% fetal calf serum [HyClone], 1 mM sodium pyruvate [Invitrogen], 2 mM l-glutamine [Invitrogen], and 50 µg/ml gentamicin [Invitrogen]). All cells were grown at 37°C in 7% CO2.
Peripheral blood mononuclear cells were isolated by Ficoll density gradient centrifugation (Histopaque; Sigma Aldrich) from buffy coats of healthy human donors (Mount Sinai Blood Donor Center and New York Blood Center). CD14+ cells were immunomagnetically purified using anti-human CD14 antibody-labeled magnetic beads and BDCA4+ cells were immunomagnetically purified using anti-human BDCA4 (CD304)+ antibody- labeled magnetic beads and iron-based Midimacs LS columns (Miltenyi Biotec). After elution from the columns, CD14+ cells were plated (0.7×106 cells/ml) in DC medium (RPMI [Invitrogen], 10% fetal calf serum [HyClone] or 4% human serum [Cambrex], 100 units/ml of penicillin, and 100 µg/ml streptomycin [Invitrogen]) supplemented with 500 U/ml human granulocyte-macrophage colony-stimulating factor (GM-CSF; Peprotech) and 1,000 U/ml human interleukin-4 (IL-4; Peprotech) and incubated for 5 to 6 days at 37°C. Our cultured DCs were routinely 95–98% positive for CD11c as tested by flow cytometry, from over 40 independent isolations. BDCA4+ cells were treated immediately following elution. PDCs were tested for purity by flow cytometry. Briefly, BDCA4+ cells were stained with fluorescein isothiocyanate FITC)-linked CD123 and phycoerythrin (PE)-linked BDCA2 (CD303), according to the manufacturer's instructions (Miltenyi Biotec), and the expression of each marker was determined by flow cytometry using an FC500 flow cytometer from Beckman Coulter. Data were analyzed using Flowjo software. The average purity of BDCA4+ cells was 91.07±5.01% as defined as double positive for CD123 and BDCA2 (CD303) with n = 63. Each experiment used an independent donor with no overlap between pDC and cDC donors.
Immediately following isolation for BDCA4+ cells and after 5 to 6 days in culture for the CD14+ cells, DCs were either pre-treated with 5 to 5,000 U/ml IFN- β (PBL) and/or were treated with one of the following: live influenza PR8 virus at a multiplicity of infection (MOI) of 0.5, UV-inactivated influenza virus at a MOI = 5, 500 ng/ml LPS (Sigma-Aldrich), 6 ug/ml CpG (Coley Pharmaceutical Group), 2.5 ug/ml poly (I∶C) (InvivoGen), 0.5 ug/ml CL-075 (InvivoGen), 1 ug/ml Gardiquimod (InvivoGen). Cells were treated in medium (RPMI [Invitrogen], 4% human serum [Cambrex], 100 units/ml of penicillin, and 100 µg/ml streptomycin [Invitrogen]) at 1×106 cells/ml for different time periods, depending on the experiment. In experiments in which the IFN media was removed, fresh media was added prior to viral infection.
Capture enzyme-linked immunosorbent assays (ELISAs) for IFN-α, TNF-α, IL-6, IL-8, RANTES, IP-10 and MIP1-β (Upstate/Millipore) were performed as part of a multiplex assay following the manufacturer's protocol. Plates were read in a Luminex plate reader, and data were analyzed using software from Applied Cytometry Systems. All samples were assayed in duplicate or triplicate.
Samples of 0.15×106 to 0.5×106 DCs differentially treated according to the experimental protocol were pelleted, and RNA was isolated and treated with DNase by using an Absolutely RNA RT-PCR micro prep kit (Stratagene). RNA was quantified using a Nanodrop spectrophotometer (Nanodrop Technologies).
qRT-PCR of the extracted RNAs was performed by using a previously published SYBR green protocol with an ABI7900 HT thermal cycler by the Mount Sinai Quantitative PCR Shared Research Facility. Each transcript in each sample was assayed in triplicate, and the mean cycle threshold was used to calculate the x-fold change and control changes for each gene. Three housekeeping genes were used for global normalization in each experiment (actin, Rps11, and tubulin genes). Data validity by modeling of reaction efficiencies and analysis of measurement precision was determined as described previously [17].
Statistical analyses were performed using student's two-tailed t test. Unless otherwise indicated, means±standard deviation for each sample are shown.
|
10.1371/journal.pgen.1000274 | Mutations in AtPS1 (Arabidopsis thaliana Parallel Spindle 1) Lead to the Production of Diploid Pollen Grains | Polyploidy has had a considerable impact on the evolution of many eukaryotes, especially angiosperms. Indeed, most—if not all—angiosperms have experienced at least one round of polyploidy during the course of their evolution, and many important crop plants are current polyploids. The occurrence of 2n gametes (diplogametes) in diploid populations is widely recognised as the major source of polyploid formation. However, limited information is available on the genetic control of diplogamete production. Here, we describe the isolation and characterisation of the first gene, AtPS1 (Arabidopsis thaliana Parallel Spindle 1), implicated in the formation of a high frequency of diplogametes in plants. Atps1 mutants produce diploid male spores, diploid pollen grains, and spontaneous triploid plants in the next generation. Female meiosis is not affected in the mutant. We demonstrated that abnormal spindle orientation at male meiosis II leads to diplogamete formation. Most of the parent's heterozygosity is therefore conserved in the Atps1 diploid gametes, which is a key issue for plant breeding. The AtPS1 protein is conserved throughout the plant kingdom and carries domains suggestive of a regulatory function. The isolation of a gene involved in diplogamete production opens the way for new strategies in plant breeding programmes and progress in evolutionary studies.
| In the life cycle of sexual organisms, meiosis reduces the number of chromosomes from two sets (2n) to one set (n), while fertilization restores the original chromosome number. However, in case of failure of meiosis to reduce the chromosome number, the fecundation involving the obtained 2n gametes can lead to the formation of an organism with more than two sets of chromosomes (polyploid). Polyploidization occurred widely in the course of evolution of eukaryotes, especially of plants. Besides, many crops are current polyploids, and 2n gametes have been useful for their genetic improvement by allowing crosses between 2n and 4n species. 2n gametes formation is known to be under genetic control but none of the genes involved were identified. We have isolated and characterised a gene (AtPS1) involved in controlling diploid (2n) gamete formation in A. thaliana. In the Atps1 mutant, the second division of meiosis is disturbed, leading to the gathering of chromosomes that had been separated at the first division. Consequently, Atps1 mutants produce 2n male gametes and spontaneous triploid plants in the next generation. The isolation of a gene involved in diplogamete production opens the way for new strategies in plant breeding programmes and progress in evolutionary studies.
| Polyploidy, the condition of organisms having more than two sets of chromosomes, has had a considerable impact on the evolution of many fungi, invertebrate, and vertebrate lineages and is particularly prominent in plants [1],[2]. It is estimated that 95% of ferns are polyploids [3] and that almost all angiosperms have experienced at least one round of whole genome duplication during the course of their evolution [4]. Many important crop plants are currently polyploids or retain the vestiges of ancient polyploid events [5]–[7]. Even plants with small genomes, such as Arabidopsis thaliana, have been affected by polyploidy [8],[9]. However, the mechanisms involved in polyploid formation are still poorly understood. For a long time, polyploids were thought to originate from somatic chromosome doubling[10]. The realisation that gametes with somatic chromosome numbers (2n gametes or diplogametes) widely occur in diploid populations as a result of meiotic failure, led to a change of paradigm[11]; it is now believed that 2n gametes are the major route for polyploidy formation, in particular by leading to the formation of triploids, which then may serve as a bridge/step towards even ploidy levels [2], [12]–[15]. 2n gametes are also instrumental in the genetic improvement of several polyploid crops, where useful genes from diploid relatives are incorporated into cultivated genotypes [16],[17].
Given their importance in evolution and crop improvement, 2n gametes have been the focus of a considerable amount of research [12],[18]. The best documented and described meiotic abnormalities leading to 2n gamete formation include abnormal cytokinesis, the omission of the first or second division and abnormal spindle geometry. Co-orientation of 2nd division spindles (parallel spindles or fused spindles) is perhaps the most common mechanism resulting in 2n spore formation [12],[18], most notably in potato [14], and was first described more than eight decades ago [19],[20].
Environmental factors, notably temperature and chemical agents, were shown to affect the frequency of 2n gametes [12],[13],[21]. However, 2n gamete production is under strong genetic control [13]. The genetic determination of 2n pollen production was studied in several species [12] and usually fits the segregation pattern expected for a major locus in a background of polygenic variation. To date, however, none of the genes contributing to high frequency 2n gametes production were identified and characterised at the molecular level [22],[23]. This lack of information has slowed down our understanding of the origins of diplogametes, and limited the potential of diplogametes in crop breeding programmes.
In this paper we describe the isolation and characterisation of the first gene, called AtPS1 (Arabidopsis thaliana Parallel Spindle 1), implicated in the formation of a high frequency of diplogametes in plants. We show that meiosis in Atps1 mutants generates diploid male spores, giving rise to viable diploid pollen grains and spontaneous triploid plants in the next generation. Analysis of male meiosis showed that during meiosis II spindles are abnormally orientated, with frequent parallel or fused spindles, leading to the production of two sets of chromosomes instead of four at the end of anaphase II. Genetic analyses of the diploid gametes and epistasis experiments demonstrated that diplogamete formation in Atps1 results from these defects in spindle organisation.
AtPS1 was identified in a screen for genes potentially involved in meiosis using the Expression Angler tool [24], which selects co-regulated genes, in combination with the AtGenExpress tissue data set [25]. We first chose a subset of known meiotic genes (AtMER3 [26], AtDMC1 [27], SDS [28], AtMND1 [29],[30], AtHOP2 [31], AtMSH5 [32] and AtSPO11-1 [33]) for which the expression data appeared to be relevant: when one of these genes was used as the query in the Expression Angler tool (with default parameters [24]), other known meiotic genes appeared among the first hits. We thus selected a list of genes that appeared among the first 60 hits in at least one query, and in the first 100 hits in at least two independent queries, using one of these seven genes as bait. Following an additional manual selection, including elimination of genes with known function or essential character, we ended up with a list of 138 candidate genes. We examined the phenotype of one to three lines carrying an insertion in each of these genes (218 lines in total) [34]–[38]. Thirteen genes were not tested because corresponding mutant lines were not found in the databases. We visually screened ∼50 plants of each line obtained from stock centers [35],[39] for reduced fruit length, without genotyping. The meiotic products of plants with reduced fertility were then examined. In two independent lines carrying an insertion in the AT1G34355 gene, plants were found to have slightly reduced fertility and unbalanced meiotic products. Chromosome spreads revealed that these plants were polyploid, prompting us to analyze these lines further. Functional characterization of this gene led us to name it AtPS1 (see below). Two other genes with meiotic function were identified in the same screen (R. M., unpublished data).
We amplified the AtPS1 cDNA (EU839993) by RT-PCR on bud cDNA and sequencing confirmed that it is identical to that predicted in the databases (NM_103158). The AtPS1 gene contains 7 exons and 6 introns (Figure 1A) and encodes a protein of 1477 amino acids. BLASTp and PSI-Blast [40] analyses showed that the AtPS1 protein is conserved throughout the plant kingdom and contains two highly conserved regions. An FHA domain (forkhead associated domain) was predicted at the N-terminus (CD-search: 65–140 aa, E-value 2e-11) [41], while the C-terminal conserved region shows similarity to a PINc domain as identified using the SMART outlier homologue search (BLAST: PINc, 1237–1389 aa, E-value 1.00e-84) [42], the InterPro superfamily search (SSF88723: PIN domain-like, 1235–1412 aa, E-value 8.8e-09) [43] as well as borderline similarities in CD-search (smart00670: PINc, 1237–1305 aa, E-value 0.21) (Figure 1B and 1C). No close homologs of AtPS1 containing both the FHA and PINc domain were found outside of the plant kingdom. An FHA domain is a phosphopeptide recognition motif implicated in protein-protein interactions and is found in a diverse range of proteins involved in numerous processes including intracellular signal transduction, cell cycle control, transcription, DNA repair and protein degradation [44]. The PINc domain has been predicted to have RNA-binding properties often associated with RNAse activity [45], and this has now been experimentally confirmed [46]. Accordingly, several PINc domain containing proteins are involved in RNAi, RNA maturation, or RNA decay. The highest level of sequence similarity to the AtPS1 PINc domain in eukaryotes was found among others with S. cerevisiae Swt1[47], mammalian C1orf26, Drosophila CG7206 and SMG6 protein families [45],[48] (Figure 1D).
We investigated the role of the AtPS1 gene by isolating and characterizing a series of allelic mutants, identified in several public T-DNA insertion line collections [34],[35],[37]. The Atps1-1 (SALK_078818) and Atps1-2 (WiscDsLox342F09) insertions are in a Columbia (Col-0) background and are in the fourth exon and first intron, respectively. The Atps1-3 (FLAG_456A09) insertion is in a Wassilewskija (Ws-4) background and is located in the second exon (Figure 1A). RT-PCR was carried out using the pAtpsF/pAtpsR primers (Figure 1A) on RNA from the Atps1-3 and Atps1-1 mutants and no detectable levels of the AtPS1 transcript were amplified, indicating that these two alleles are null. When the same primers were used on RNA from the Atps1-2 mutant normal expression levels of this region of the AtPS1 transcript were observed (data not shown). Nevertheless, the phenotype analysis described below strongly suggests that this third allele is also null.
In A. thaliana, male meiosis produces a group of four spores, organised in a tetrahedron, called a tetrad. As expected, male meiotic products in wild type were almost exclusively tetrads (Figure 2). Rarely, (13/304) groups of three spores were also seen but these were most certainly the result of occasional spore superposition. In contrast, the meiotic products in the three independent Atps1 mutants were characterized by a high frequency of dyads and triads (Figure 2). Atps1 mutants did not show any other developmental defects. The Atps1-1 and Atps1-2 mutants produced a majority of dyads (∼65%). The Atps1-3 mutant phenotype appeared to be weaker and only 8% of its meiotic products were dyads. Complementation tests between Atps1-1 and Atps1-2 and Atps1-3 and Atps1-1 showed that these mutations are allelic, and thus demonstrated that the dyads observed in this series of mutants are due to disruption of the AtPS1 gene.
The Atps1-3 mutant exhibited a weaker phenotype than the two other alleles, whereas expression analysis suggested that this allele is also null. As this allele was in a different genetic background (Ws-4) to the two others (Col-0), we tested if this difference could be influencing the strength of the phenotype by introducing the Col-0 mutation into the Ws-4 background and vice versa. As expected for a background effect, the frequency of dyads increased with successive backcrosses when Atps1-3 was introduced into Col-0 (from 8% to 58% after four backcrosses) and decreased when Atps1-1 was introduced into the Ws-4 background (from 64% to 13% after four backcrosses). These results clearly indicate that the frequency of diploid gametes is influenced by multiple genes, with AtPS1 acting as a major gene.
Pollen grain viability was examined by Alexander staining [49] and showed that in the majority of cases the dyads and triads produced by the mutants result in viable pollen grains (more than 95% in the different Atps1 mutants : Col: 0 dead pollen grains out of 181 ; Atps1-1: 44 dead pollen grains out of 948 ; Atps1-2: 3 dead pollen grains out of 363). We did observe however that the pollen grains in mutant plants varied in size (data not shown). We then assessed the ploidy level of Atps1-1 and Atps1-2 pollen grains by quantifying spermatic nuclei DNA. Both mutants exhibited two different populations of pollen grains, one corresponding to viable haploid pollen grains (∼40% estimated by maximum likelihood) and another to viable diploid pollen grains (∼60% estimated) (data not shown). These proportions are compatible with the proportion of dyads, triads and tetrads observed in the mutants. In summary, the Atps1-1 and Atps1-2 mutants produce a high frequency of viable diploid pollen grains.
Next, we measured the ploidy level of the offspring of diploid Atps1 mutants by flow cytometry. Diploid and triploid plants (30%), but no tetraploid plants, were found among the progenies of Atps1-1 and Atps1-2 mutants (Atps1-1: 38 triploids out of 130 plants; Atps1-2: 30 triploids out of 103 plants). Flow cytometry results were confirmed by karyotyping a subset of 29 plants which were all confirmed to be triploid. This demonstrated that the diploid gametes produced in the Atps1 mutants are involved in fertilisation and produce viable triploid plants. The appearance of triploids, but not tetraploids, suggests that the Atps1 mutations only affect male meiosis. As expected for the absence of a female meiotic defect we never isolated triploid plants when ovules from plants with the Atps1 mutation were fertilised with wild type pollen grains (0 triploids out of 182 plants). When mutant pollen was used for the cross we again observed that 30% of the progeny were triploids (20 triploids out of 56 plants).
The observed frequency of triploid plants (30%) among Atps1-1 and Atps1-2 mutant progeny is lower than expected from the frequency of diploid pollen grains produced by these mutants (∼60%). In parallel, more than 50% of seeds obtained by selfing the Atps1-1 and Atps1-2 mutants were thinner than wild type, abnormally colored and shaped, and germinated at a rate of 57%, compared to 99.8% in wild type. We do not believe, however, that this seed mortality phenotype infers a possibly essential role for AtPS1 in embryo development, for the following two reasons: 1) 25% (56/210) of the progeny of selfed heterozygotes were mutant and no dead seed was obtained, showing that the Atps1 mutation does not impair embryo development. 2) The same seed defect (59% of germination) is observed when Atps1 is crossed as male with wild type as female, which shows that a seed with one functional AtPS1 allele may show developmental defects. Thus, a likely explanation for the discrepancy between the frequency of diploid pollen grains and triploids in the progeny is abnormal development of triploid seed, which is commonly observed during crosses between plant species with different ploidy levels. These problems are related to the paternal to maternal ratio, which is very important for normal endosperm development [50]. Using C24 and Ler accessions, Scott et al showed that triploid seeds obtained in diploid X tetraploid crosses germinated at a rate of 90%. We obtained stronger germination defects with Col0, suggesting a background effect on the susceptibility to the paternal/maternal ratio. Another, non-exclusive, explanation for the discrepancy could be that haploid pollen grains out-competed diploid pollen grains, which were shown in some cases to germinate more slowly [51],[52]. Nevertheless, approximately 25% of the triploid embryos appear to be able to overcome these constraints.
To unravel the mechanisms leading to dyad production in Atps1-1, we investigated chromosome behaviour during meiosis (Figure 3). Chromosome spreads showed that the meiosis in the Atps1-1 mutant progresses normally and is indistinguishable from the wild type until the end of the telophase I. Synapsis was complete, chiasmata formed (the cytological manifestation of crossovers) and bivalents were seen (compare Figure 3G–I with Figure 3 A–C, for example). At metaphase II, however, differences were seen compared to wild type with the 10 chromosomes aligned in a same plane, causing abnormal looking figures, rather than two well separated metaphase II plates containing five chromosomes each (Compare Figure 3J–K with Figure 3D). In rare cases, metaphase II in Atps1 did appear normal however (Figure 3L). At telophase II, we observed dyads (two sets of 10 chromosomes, Figure 3M), triads (2 sets of five chromosomes and one set of 10, Figure 3N) and normal tetrads (4 sets of 5 chromosomes, Figure 3O). These observations are consistent with the previous finding that Atps1 meiotic products are a mixture of dyads, triads and tetrads.
These results and specifically the alignment of the 10 chromosomes at metaphase II suggested that the meiotic spindles in Atps1 mutants are defective at this stage. We thus examined spindle organisation by immunolocalisation with an alpha-tubulin antibody (Figure 4). In wild type plants the majority of metaphase II spindles were roughly perpendicular to each other (Figure 4A), leading to four well separated poles at anaphase II (Figure 4B) and the formation of tetrads (Figure 4C). In the Atps1 mutant, while individual metaphase II / anaphase II spindles appeared regular in most cases their respective orientation was aberrant. The majority of cells had parallel spindles (Figure 4D to 4G), fused spindles (Figure 4H and 4I) or tripolar spindles (Figure 4J and 4K). This defect in spindle orientation explains the appearance of triads and dyads. These conformations cause chromatids, that had been separated at meiosis I, to gather at anaphase II. Occasionally, three to four sets of chromosomes encompassed by a spindle were dispersed in the cell at metaphase II (Figure 4L). This type of defect is probably the cause of the few unbalanced meiotic products observed in the Atps1 mutants.
The name AtPS1 for Arabidopsis thaliana Parallel Spindle 1 was chosen due to the high percentage of parallel spindles produced by the corresponding mutants.
Thus, parallel spindles at metaphase II in the Atps1 mutants appear to be leading to the formation of dyads. This proposed mechanism implies that unbalanced chromosome segregation at meiosis I would have no impact on the final distribution of chromosomes in the resulting dyad. To test this hypothesis we constructed a double Atspo11-1/Atps1 mutant. The Atspo11-1 mutant (N646172, Atspo11-1-3) [53] displays an absence of bivalents at meiosis [33] (Figure 5A) leading to frequent unbalanced first divisions (Figure 5B) that can be associated with lagging chromosomes (Figure 5C). At metaphase II, unbalanced plates are seen (Figure 5D), leading to unbalanced tetrads (Figure 5E). Lagging chromosomes at anaphase II, lead to multiple metaphase II plates and then polyads with more than four nuclei (Figure 5F).
In the Atspo11-1/Atps1 background the first division was identical to the single Atspo11-1 phenotype. We observed 10 univalents at metaphase I (Figure 5G), leading to missegregation at anaphase I, with two sets of unbalanced chromosomes (Figure 5H) or three sets because of lagging chromosomes (Figure 5I). At metaphase II, we regularly observed two unbalanced metaphase plates, which had a tendency to be parallel instead of perpendicular (Figure 5J). This led to the formation of dyads which were always balanced (Figure 5K to 5L, n = 44). We also observed triads with one set of 10 chromosomes caused by an unbalanced first division followed by the fusion of two of the four second division products (Figure 5N), which is highly consistent with our proposed mechanism. We also observed unbalanced tetrads (Figure 5P and 5Q), expected since the Atps1 mutation is not fully penetrant, and polyads due to lagging chromosomes at the first division (Figure 5R).
Another prediction of the proposed mechanism is that centromere distribution should resemble that seen during mitosis, e.g., any heterozygosity at the centromeres should be retained in the diploid gametes. Indeed, in Atps1, the first division is identical to wild type, with the co-segregation of sister chromatids and separation of homologous chromatids. Thus, in the case of a heterozygous genotype, A/a, at the centromere, following the first division the two A alleles will end up at one pole, and the two a alleles at the opposite pole. In wild type, the second division separates the two sisters leading to four spores with one chromatid. In Atps1, the second division would regroup the products of the first division, thus grouping the a and A allele in each cell, leading to systematic heterozygosis at the centromere. Because of recombination, loci unlinked to centromeres should segregate randomly. We tested this prediction by taking advantage of the two genetic backgrounds of the Atps1-1 (Col-0) and Atps1-3 mutants (Ws-4). F1 plants bearing the two mutations – thus mutant for AtPS1 and heterozygous for any Col-0/Ws-4 polymorphisms – were crossed as male to a third genetic background Landsberg erecta (Ler). Karyotyping and genotyping of the obtained plants for trimorphic molecular markers provided direct information regarding the genetic make up of the pollen grain produced by the mutant (Figure 6). All the diploid pollen grains tested had the predicted genetic characteristics. They were systematically heterozygous at centromeres and segregating–because of recombination–at other loci. These results confirm that the “parallel spindle” defect is indeed the cause of at least the vast majority of 2n pollen in Atps1.
In this study, we identified and described the AtPS1 gene and a corresponding set of mutants that produce pollen grains which are up to 65% diploid and give rise to numerous triploid plants in the next generation. Another Arabidopsis mutant that leads to severe meiotic defects and almost sterility [54],[55] was recently described and reported to produce diploid female gametes [56], but at a frequency of several orders of magnitude lower than the frequency of 2n gametes induced by the Atps1 mutation. By combining cytological and genetic analyses, we carried out a detailed investigation of the mechanism responsible for these 2n pollen grains in Atps1, and established that they result from abnormal orientation of spindles at meiosis II. Interestingly, defects in meiosis II spindles are the most common known mechanisms responsible for the formation of 2n spores, [12],[18] and are the main source of the 2n pollen which is extensively used in potato breeding programmes [14]. In potato, a major locus called ps was shown to be responsible for the parallel spindle phenotype more than 30 years ago [57], but the corresponding gene is still to be identified. As was observed in different ps potato lines, Atps1 mutations only affect male meiosis and the frequency of dyads formed depends on the genetic background. The AtPS1 gene is conserved in higher plants (Figure 1C) and is therefore a good candidate for the gene behind the major ps locus of potato [14].
The fact that Atps1 mutations only affect male meiosis points to a difference in regulation between male and female 2n gametes production. This phenomenon was previously described for mutations that had a specific impact on either male or female meiosis [12],[23],[55],[56],[58]. In the case of parallel spindles, it may stem from the 3-dimension organization of the spores (e.g. tetrahedron in male vs linear or multiplanar arrays in female [59]).
The AtPS1 protein has two domains, a FHA (ForkHead Associated) domain, a phosphopeptide recognition domain found in many regulatory proteins and a PINc domain, which is found in proteins involved in RNA processing [48]. In fungi/metazoa, the AtPS1 PINc domain shows highest similarity with the PINc domains of the Swt1/ C1orf26/ CG7206 and SMG6 protein families followed by SMG5, Dis3 and others. The mammalian C1orf26 and Drosophila CG7206 genes encode related proteins of unknown function, but Interestingly both are overexpressed in testis and ovaries, which is consistent with a putative meiotic role [60],[61]. SMG6 is an essential component of the Nonsense Mediated RNA Decay (NMD) machinery that degrades mRNAs containing premature translation termination codons. SMG6 also plays a role in RNAi [45],[48]. The SMG6 PINc domain has RNA degradation activity [46]. These features suggest that AtPS1 plays a regulatory function, perhaps via RNA decay, which may directly control the orientation of metaphase plates/spindles or be related to meiotic cell cycle control. There is growing evidence that NMD and its components have important functions in various cellular processes, including the cell-cycle [62]. A link between RNA decay and the control of meiosis progression was recently suggested because SMG7, which is a NMD essential component, is involved in progression through meiotic anaphase II in Arabidopsis [63].
Further studies involving AtPS1 should shed light on the poorly understood process of meiosis II. The isolation of a gene involved in 2n gamete production has important implications for deciphering meiosis mechanisms, as well as potentially fundamental applications in evolution studies and plant breeding programmes.
Arabidopsis plants were cultivated as described in [64]. For germination assays and cytometry experiments Arabidopsis were cultivated in vitro on Arabidopsis medium [65] at 21°C with a 16h day/8h night photoperiod and 70% hygrometry.
The Atps1-1 (SALK_078818) and Atps1-2 (WiscDsLox342F09) lines were obtained from the European Arabidopsis stock centre [39]. The Atps1-3 (FLAG_456A09) insertion is from the Versailles T-DNA collection[35]. Plants were genotyped by PCR (30 cycles of 30 s at 94°C, 30 s at 56°C and 1 min at 72°C) using two primer pairs. For each line the first pair designated is specific to the wild type allele and the second pair is specific to the T-DNA insertion.
Atps1-3: EQM96L (5′ACATCTCCCTTGTCGTAAC3′) and EQM96U (5′ATCTCTCAATCGTTCGTTC3′); EQM96L and tag3 (5′ CTGATACCAGACGTTGCCCGCATAA3′). Atps1-1: N578818U2 (5′TCGGAGTCACGAAGACTATG3′) and N578818L (5′CAGTCTCACTGATTATTCCTG3′); N578818U2 and LbSalk2 (5′GCTTTCTTCCCTTCCTTTCTC3′). Atps1-2: N851945U (5′AAGGCTGATATTCTGATTCAT3′) and N851945L (5′CTCTTGTTGGTCCGTATCTTA3′); N851945U and P745 (5′AACGTCCGCAATGTGTTATTAAGTTGTC3′). spo11-1-3: N646172U (5′AATCGGTGAGTCAGGTTTCAG3′) and N646172L (5′CCATGGATGAAAGCGATTTAG 3′); N646172L/ LbSalk2.
Genetic markers used to genotype Atps1-1/Atps1-3×Ler F1 triploid and diploid plants (40 cycles of 20 s at 94°C, 20 s at Tm and 30 s at 72°C): Microsatellite msat1.29450 (located on chromosome I at position 29450001) was amplified (Tm = 57°C) using 5′TCCTTTCATCTTAATATGC3′ and 5′TCTGTCCACGAATTATTTA3′ primers. Microsatellite Msat4.35 (Tm = 58°C) (located on chromosome 4 at position 7549125) was amplified using 5′CCCATGTCTCCGATGA3′ and 5′GGCGTTTAATTTGCATTCT3′ primers. Microsatellite NGA151 (Tm = 58°C) (located on chromosome 5 at position 4669932) was amplified using 5′GTTTTGGGAAGTTTTGCTGG3′ and 5′CAGTCTAAAAGCGAGAGTATGATG3′ primers. The 2 primer pairs specific for the Atps1-1 and Atps1-3 TDNA borders were used as a centromeric marker of the chromosome 1. CAPS markers Seqf16k23 (physical position: 14481813) and CAPSK4 51 (physical position: 5078201) were used as centromeric markers for chromosome 1 and 4, respectively. CAPS Seqf16k23 was amplified (Tm = 60°C) using 5′GAGGATACCTCTTGCTGATTC3′ and 5′CCTGGCCTTAGGAACTTACTC3′ primers and observed after TaqI digestion. CAPS CAPSK4 51 was amplified (Tm = 60°C) using 5′CAATTTGTTACCAGTTTTGCAG3′ and 5′TGAGTTTGGTTTTTTGTTATTAGC3′ primers and observed after MnlI digestion.
Final meiotic products were observed as describe in [28] and viewed with a conventional light microscope with a 40× dry objective. Chromosomes spreads and observations were carried out using the technique described in [33]. The DNA fluorescence of spermatic pollen nuclei was quantified using open LAB 4.0.4 software. For each nucleus the surrounding background was calculated and subtracted from the global fluorescence of the nucleus. Meiotic spindles were observed according to the protocol described in [55] except that the DNA was counter-stained with DAPI. Observations were made using an SP2 Leica confocal microscope. Images were acquired with a 63× water objective in xyz and 3D reconstructions were made using Leica software. Projections are shown. Cells were imaged at excitation 488 nm and 405 nm with AlexaFluor488 and DAPI respectively. Arabidopsis genome sizes were measured as described in [66] using tomato Lycopersicon esculentum cv “Montfavet” as the standard. (2C = 1.99 pg, %GC = 40.0%).
Arabidopsis total RNA was extracted using the QUIAGEN RNA kit.
Reverse transcription was done on 5 µg of total RNA using oligo(dT)18 as primer. The RevertAid M-MuLV Reverse Transcriptase enzyme (Fermentas) was used according to the manufacturer's instruction. RT-PCR was carried out on 1 µl of cDNA using the pAtpsF and pAtpsR primers and the following PCR conditions: 30 cycles of 30 s at 94°C, 30 s at 56°C and 1 min at 72°C.
|
10.1371/journal.pcbi.1004576 | Maximizing Sensory Dynamic Range by Tuning the Cortical State to Criticality | Modulation of interactions among neurons can manifest as dramatic changes in the state of population dynamics in cerebral cortex. How such transitions in cortical state impact the information processing performed by cortical circuits is not clear. Here we performed experiments and computational modeling to determine how somatosensory dynamic range depends on cortical state. We used microelectrode arrays to record ongoing and whisker stimulus-evoked population spiking activity in somatosensory cortex of urethane anesthetized rats. We observed a continuum of different cortical states; at one extreme population activity exhibited small scale variability and was weakly correlated, the other extreme had large scale fluctuations and strong correlations. In experiments, shifts along the continuum often occurred naturally, without direct manipulation. In addition, in both the experiment and the model we directly tuned the cortical state by manipulating inhibitory synaptic interactions. Our principal finding was that somatosensory dynamic range was maximized in a specific cortical state, called criticality, near the tipping point midway between the ends of the continuum. The optimal cortical state was uniquely characterized by scale-free ongoing population dynamics and moderate correlations, in line with theoretical predictions about criticality. However, to reproduce our experimental findings, we found that existing theory required modifications which account for activity-dependent depression. In conclusion, our experiments indicate that in vivo sensory dynamic range is maximized near criticality and our model revealed an unanticipated role for activity-dependent depression in this basic principle of cortical function.
| When many simple parts interact, the collective behavior of the whole can be astonishingly complex. A particularly striking example is our capacity for sensory perception, which results from the collective interactions of billions of relatively simple neurons. Another example is found in physical systems which undergo a phase transition–for example, liquid water turning to solid ice. When collective interactions among the water molecules are changed, the system transitions from a disordered state (liquid) to an ordered state (crystalline solid). At the tipping point of a critical phase transition, i.e. at criticality, physical systems exhibit very complex behavior. In this study, we show that phase transitions may occur in the cerebral cortex changing the neural activity from a disordered to an ordered state. Moreover, this neural phase transition may be intimately linked with sensory perception. We experimentally manipulate the interactions among neurons and show that sensory dynamic range is maximized when the cerebral cortex of a rat is closest to criticality.
| Cortical neuronal network dynamics shift among myriad states to cope with the changing needs of the organism [1–3]. Strikingly different cortical states are observed during different behaviors such as sleep [4], wakeful resting [5], active movement [6], or vigilant attention [7]. Externally-imposed manipulations of interactions among cortical neurons, like neuromodulators [7–9], anesthetics [10–12], and other drugs [13,14], also alter the cortical state. Which cortical states are optimal for gathering information about the world through sensory input? Answers to this question are only beginning to be understood. For example, previous studies have shown that changes in cortical state impact sensory adaptation [5], variability of cortical response [9,12,15,16], and the ability to track fast stimulus changes [12,17]. Here we focused on the ability of cortical neuronal networks to distinguish a wide range of stimulus intensities, i.e. sensory dynamic range. We sought to delineate how sensory dynamic range depends on cortical state. To meet this goal, we took advantage of changes in cortical state that occur naturally [15,18] without experimental control and we also imposed changes in cortical state by tuning cortical inhibitory interactions [19].
Our approach was motivated, in part, by theory [20–22] and in vitro experiments [19] which point to a potential general principle governing cortical dynamic range. They proposed that dynamic range is maximized by tuning the cortex to operate at criticality. Criticality is a boundary regime separating two distinct regimes of cortical state [23,24]. On one side of the critical boundary, the ‘subcritical’ cortical state is characterized by asynchronous population activity, low firing rates, and low sensitivity to stimuli. On the other side, the ‘supercritical’ cortical state is marked by large-scale, coordinated population activity and tends to be hyperexcitable in response to stimulation. Cortical dynamic range is thought to be low in the subcritical state due to insensitivity to weak stimuli. In contrast, existing theory suggests that dynamic range is low in the supercritical regime because the system tends to saturate with most neurons in the network firing at high rates, even without external input. Criticality is thought to be optimally balanced between these extremes, able to detect weak stimuli without saturating. However, this potentially fundamental relationship between criticality and sensory dynamic range has not been tested in an intact sensory system. Indeed, the theory may be irrelevant because in vivo cortical networks never reach the saturated firing regime that has been theoretically shown to be responsible for low dynamic range in the supercritical state. Synaptic depression and other mechanisms serve to prevent such saturated firing. Thus, it remains unclear if in vivo sensory dynamic range will indeed be highest when the cortex operates near criticality.
Here we directly measured the in vivo relationship between cortical state and somatosensory dynamic range in the rat whisker system. We found that dynamic range is highest in cortical states that exhibit signs of criticality. However, our experimental observations were not well-explained by existing theories, particularly in the experimentally observed supercritical regime. To account for our experimental results we used a model with strong activity-dependent depressive effects, thus avoiding the saturated response in the supercritical regime. Thus, we conclude that, for reasons not anticipated by previous theory, in vivo sensory dynamic range is maximized near criticality.
We investigated relationships among cortical state, inhibition, and sensory dynamic range in the whisker system of urethane anesthetized adult male rats (n = 13 rats, 94 recordings). We recorded multi-unit activity (MUA) in barrel cortex using 32-channel microelectrode arrays (Fig 1). Each rat was studied first in the normal anesthetized state, second, with pharmacologically altered synaptic inhibition, and, finally, under a wash condition. Changes in cortical state occurred for two reasons. First, we manipulated the cortical state by pharmacological modulation of synaptic inhibition, locally in the recorded region of cortex, by topical application of muscimol (GABA agonist) or bicuculline methiodide (GABA antagonist). Second, we found that changes in cortical state occurred naturally, without altered inhibition. Rather than ‘averaging out’ this natural variability of cortical state, we took advantage of it; we systematically parameterized the range of cortical states that we observed. This approach acknowledges that the cortical state is not static even under ‘normal’ conditions and state changes can result in significant experiment-to-experiment and animal-to-animal variability. During times with no whisker stimulus, we quantitatively assessed the cortical state using multiple features of ongoing MUA activity, including correlations, spatiotemporal variability, and the prevalence of different spatiotemporal scales. We analyzed the response to stimulation to quantify sensory dynamic range.
First, we parameterized the cortical state based on the prevalence of different spatiotemporal scales of population spiking activity. Our approach accounts for the relative importance of diverse scales, avoiding bias for any particular scale. Motivated by previous studies of spatiotemporal cascades of population activity called ‘neuronal avalanches’ [25–27], we began by making a population MUA spike count time series including spikes recorded on all electrodes. Then, ‘avalanches’ were defined as periods of time when the MUA spike count exceeded a threshold (Fig 2A). We note that our results were robust to variation (up or down by a factor of 2) in the choice of threshold and time bin duration (S1 and S2 Figs). The ‘size’ of each avalanche was defined as the total number of spikes occurring during the avalanche. To determine the prevalence of different spatiotemporal scales, for each recording, we examined the distribution of avalanche sizes (Fig 2B and 2D).
Examining avalanche size distributions over all of our experiments, we found that a continuum of different network states occurred (Fig 2D). At one end of the continuum, distributions were bimodal, indicating that large-scale avalanches played a prominent role in the cortex dynamics. This situation often occurred for pharmacologically reduced inhibition (Fig 2B and 2C). At the opposite end of the continuum, we observed cortical states in which small scales were dominant, often occurring when inhibition was enhanced (Fig 2B and 2C). The cortical state varied continuously between these extremes (Fig 2C). Near the middle of the continuum, we observed highly diverse avalanches with heavy-tailed distributions [23,28] of avalanche size, close to a power-law distribution with exponent -1.5 (Fig 2D). To quantitatively index the observed continuum of cortical states, we employed the parameter κ, which measured the deviation between the observed avalanche size distribution and a power law with exponent -1.5, as in previous work [11,19,27]. In brief, large κ entailed a cortical state with strongly coordinated population activity, commonly sweeping across the entire recording area (Fig 2E). For small κ, population activity was weakly coordinated, typically confined to small spatial extents (Fig 2E). Separating these extremes, the cortical state with κ = 1 exhibited more diverse population activity with power law distributed spatiotemporal scales. Power law avalanche size distributions have additional significance because they are predicted to occur in a specific cortical state called ‘criticality’, as discussed in the introduction section. The particular power law exponent -1.5 is associated with a particular type (universality class) of criticality, namely, directed-percolation [29] and has also been studied in other excitable networks [30].
The degree of correlations among cortical neurons plays an important role in population coding [31] and cortical state [1]. Our second approach for assessing the cortical state was based on pairwise correlations of MUA recorded at different electrodes. For this, we created MUA spike count time series for each electrode, excluding the times when whiskers were stimulated. Then, we computed the Pearson correlation coefficient between all pairs of electrodes. We found that correlations were closely related to our state index κ. The distribution of pairwise correlation coefficients was broadest for states near κ = 1 (Fig 3A and 3B), which reflects the diversity of avalanche sizes that occur in such states. States with either small or large κ had relatively narrow distributions of correlations with decreased or increased average correlations, respectively (Fig 3A and 3B). The average pairwise correlation of the population increased sigmoidally as κ is increased (Fig 3C). This demonstrates that the state with κ = 1 lies at the boundary separating two distinct dynamical regimes–one with low correlations, the other with high correlations.
Our ultimate goal was to determine how cortical somatosensory dynamic range depends on cortical state. For this, dynamic range was calculated based on average cortical neural response to a range of whisker stimulus intensities (Fig 4). We defined neural response to be the MUA spike rate during the 100 ms following stimulus onset (Figs 1 and 4A). We defined the stimulus to be the average whisker speed during the 100 ms following stimulus onset (Figs 1 and 4B). Repeated identical puff pressures generally resulted in different whisker motion. Therefore, we parameterized the stimulus based on measurements of the actual whisker speeds for each puff. Whisker speeds ranged from 0 to about 30 mm/s. For cortical states with small κ, the response curves tended to rise gradually with increasing whisker speed (Fig 4C). For states with large κ, the response curve tended to rise sharply and saturate for a relatively small whisker speed (Fig 4C). Dynamic range was defined based on the range of whisker speeds over which the response increased from 10% to 90% of the response range (Fig 4D, inset). The main result of our work is that dynamic range was low in experiments with either low κ or high κ; dynamic range was maximized for cortical states with κ≈1 (Fig 4D). We remind the reader that κ is based solely on ongoing activity; periods of stimulus-evoked activity are excluded when computing κ. Comparing dynamic range (Fig 4D) to correlations (Fig 3C), our results establish that cortical dynamic range is maximized for cortical states with an intermediate degree of correlations.
In the model, we obtained stimulus-response curves and dynamic range trends similar to those we observed in our experiments (Fig 4E). To obtain this match, we had to limit the range of stimulus intensities to be below 10−3 stimulus driven spikes per model time step. As discussed further below, further increases in stimulus intensity resulted in a further rise in the response curve and disagreement between the model and experiment.
Our findings offer a specific answer to a long standing debate concerning what degree of correlations among neurons is optimal for sensory information processing [31–33]. If correlations are too strong—many neurons firing synchronously—then coding is redundant and metabolically inefficient. At the other extreme, sufficiently weak correlations may compromise the robustness of information transfer among cortical circuits. Thus, functionally effective correlations must lie between these extremes, but pinpointing the optimal level of correlations has been an elusive goal. In the context of sensory dynamic range, our results demonstrate that the specific intermediate level of correlations that coincides with power law distributed avalanches (κ = 1) is optimal.
Our study was motivated by predictions of maximized dynamic range at criticality based on pioneering analytical and computational studies [20–22,34]. At first glance, these previous predictions appear to agree with our main findings here. However, taking a closer look, we found that these previous models and theories did not explain our experimental observations. The discrepancies were in the putative supercritical regime (κ>1), experimentally observed when inhibition was suppressed. In this case, we observed large bursts of synchronous spiking activity occurring at irregular intervals, emerging from a mostly inactive baseline activity (red, Fig 2A). In contrast, in the supercritical regime of most previously studied models, ongoing activity manifests as persistent activity with nearly all neurons active at all times, with no quiet periods and no synchronous bursts (Fig 5A). We found that a simple way to modify previous models to produce more realistic, bursty dynamics was to introduce activity dependent depression–spiking probability was reduced in proportion to how may spikes occurred in the recent past (see Materials and Methods). This naturally resulted in large bursts of population activity separated by times of relative silence (Fig 2E and Fig 5A), as seen experimentally when inhibition was reduced. Without such depressive adaptation, our model produced sustained, saturated activity in the supercritical regime and dynamic range was maximized near criticality like in previous models (Fig 5A and 5B). Including depressive adaptation dramatically altered the shape of the stimulus response curve compared to that of a model without activity dependent depression (Fig 5B and 5C). In fact, if the full range of stimuli studied in previous models was considered, including those large enough to activate a large fraction of the network, then dynamic range was no longer maximized at criticality, as shown in Fig 5C. However, such large stimuli are not relevant in real sensory systems; even the most intense whisker stimulation does not result in a neural response that approaches the system size, i.e. all neurons firing. Thus, the most plausible comparison to our experiments was to exclude the large-stimulus section of the response curve. This limitation leads to response curves (Fig 4E) which match well with our experimental observations (Fig 4C) and, most importantly, recovers the result that dynamic range is maximized near criticality. In conclusion, our results indicate that a different mechanism than previously predicted is responsible for the experimental observation of peak somatosensory dynamic near criticality in vivo.
A natural question arises due to the fact that activity dependent depression dramatically changes the nature of network dynamics in the supercritical regime. Does activity-dependent depression change the nature of the phase transition; do we expect a continuous phase transition or some other type of phase transition? We leave this question to be answered by future theoretical work, but we speculate, based on the following reasoning, that the phase transition remains continuous. The activity-dependent depression has no effect on the model dynamics if spike rates are sufficiently low (< 1 spike per 80 time steps). Indeed, in the subcritical regimes where spike rates are relatively low, the stimulus response curves in Fig 5C (with depression) are not significantly different than those in Fig 5B (without depression). Since the tipping point of the phase transition is close to this low activity regime, it is likely that a continuous phase transition remains continuous when the model includes depression. This speculation is partially supported by the fact that our model produces power-law distributed avalanches (Fig 2G) when inhibitory modulation is near 1. Such power-laws are expected for continuous phase transitions.
Finally, our findings highlight a promising hypothesis for future research; changes in cortical state due to changes in behavioral context [1] may tune sensory dynamic range to suit the needs of the organism. For example, a highly focused task may benefit from a state with lower dynamic range, away from criticality. In contrast, a critical cortical state with high dynamic range may be optimal when vigilance or readiness for unknown input is important. Confirmation of this hypothesis would establish a general principle of sensory information processing: sensory dynamic range can be optimized by tuning the cortical state and maximized specifically in the critical cortical state.
All procedures were carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and approved by University of Arkansas Institutional Animal Care and Use Committee (protocol #12025). We studied adult male rats (n = 13, 328±54 g; Rattus Norvegicus, Sprague-Dawley outbred, Harlan Laboratories, TX, USA). Anesthesia was induced with isoflurane inhalation and maintained with urethane (1.5 g/kg body weight (bw) dissolved in saline, intraperitoneal injection (ip)). Dexamethasone (2 mg/kg bw, ip) and atropine sulphate (0.4 mg/kg bw, ip) were administered before performing a 2 mm x 2 mm craniotomy over barrel cortex (1 to 3 mm posterior from bregma, 5 to 7 mm lateral from midline).
Extracellular voltage was recorded using 32-channel microelectrode arrays. The electrode arrays were comprised of 8 shanks with 4 electrodes per shank, 200 μm inter-electrode distance, 400 μm inter-shank distance (A468-5 mm-200–400-177-A32, NeuroNexus, MI, USA). Each shank was made of silicon and each electrode contact was made of iridium. Shanks were 50 μm x 15 μm in cross section. Electrode impedances were approximately 1 MΩ at 1 kHz. Insertion depth was 650 μm, centered 2 mm posterior from bregma and 6 mm lateral from midline. Voltages were measured with respect to an AgCl ground pellet placed in the saline-soaked gel foams, which protect the exposed tissue surrounding the insertion site. Voltages were digitized with 30 kHz sample rate (Cereplex + Cerebus, Blackrock Microsystems, UT, USA). Recordings were filtered between 300 and 3000 Hz and thresholded at -3 SD to detect multi-unit activity (MUA).
All whiskers were trimmed except 2–4 whiskers from rows A-C and arcs 1–4. A computer-automated, pressure-controlled air puff (1 s duration) was used to deliver 10 different puff intensities, each repeated 20 times in pseudorandom order at 7 s intervals. As shown in Fig 1 and previously described [35], two-dimensional (rostrocaudal and mediolateral) multi-whisker motion was measured with millisecond, micron precision using two line cameras (LC100, Thorlabs Inc, NJ, USA). Response curves were based on the speed of the whisker which evoked the largest MUA neural response, which we call the dominant whisker.
Up to nine 20 min recordings were conducted with each rat. First, three recordings were performed with no direct manipulation of inhibition (n = 32, indirect effects may be imposed by anesthetics [36] and atropine sulfate). Then, three recordings were performed with a drug topically applied via gel foam pieces soaked in saline mixed with drug. Finally, three wash experiments were performed with drug-free gel foams. Three drug conditions were studied (one condition per rat): 1) 20 μM muscimol (6 rats, 15 recordings), 2) 20 μM bicuculline methiodide (4 rats, 10 experiments), 3) 40 μM bicuculline methiodide (3 rats, 8 experiments). The wash condition for bicuculline typically returned to activity similar to that measured in pre-drug experiments, but this was not the case for muscimol.
The model was comprised of N = 1000 binary, probabilistic, integrate-and-fire neurons. At each time t, the state si(t) of neuron i was 0 (quiescent) or 1 (firing). At each time there was a probability pext that each neuron would fire due to external input and a probability pi(t) that it would fire due to input from other neurons Ii(t),
pi(t)={1forIi(t)>1Ii(t)for0≤Ii(t)≤10forIi(t)<0,
Ii(t)=(∑j≠iWijsj(t−1))1hi(t),
where hi(t) depends on recent firing history
hi(t)=∑τ=t−Ttsi(τ).
In cases where this sum was zero, we set h to 1. Note that T plays an important role determining the character of the ongoing network dynamics as well as the shape of the stimulus response curves presented in Results (see Fig 5). For the results shown in Figs 2, 3, 4 and 5C, we set T = 80 ms (assuming that one model time step takes 1 ms) to obtain good qualitative fit with observed experimental results. For the results shown in Fig 5B, we set T = 0, i.e. there was no history-dependent depression.
The default synaptic weight matrix W is constructed as follows. First, all entries are drawn from a uniform distribution [0, 1]. Second, 20% of neurons are designated as inhibitory and the corresponding columns of W are multiplied by -1. Third, all entries are multiplied by a constant to enforce that the largest eigenvalue of W is 1. This third step ensures that the network operates at criticality [21]. Thus, the network topology is all-to-all coupling, but with non-uniform coupling strength. To model pharmacological manipulation of inhibition we multiply all the negative entries of W by a constant ranging from 0 (model of strong GABA antagonist) to 3 (model of strong GABA agonist). This is the quantity labeled ‘inhibitory modulation’ in Figs 2, 3, 4 and 5. These manipulations change the largest eigenvalue of W, thus pushing the system away from criticality.
To simulate the onset of sensory stimulation, pext undergoes a step increase. The pre-stimulus low rate (pext = 5 × 10−6) resulted in 5 externally-driven spikes per second for the entire network, assuming that one model time step was 1 ms. The during-stimulus high rate was fixed during a single trial, but varied across trials to model different intensities of sensory stimulation. High rates ranged from pext = 5 × 10−6 to pext = 1 × 10−3, i.e. generating 5 to 1000 externally-driven spikes per second for the entire network. Each level of pext we repeated 20 times, just as each experimental stimulus intensity was repeated 20 times. We note that increasing pext to values approaching 1 does result in a saturation of network activity, which changes the shape of the stimulus-response curves and changes dynamic range. However, such highly saturated response-curves were not observed in our experiments. The chosen range of pext, better fits our experimental observations.
MUA spike count time series were based on time bins of duration DT = 7.5±3 ms (mean±SD), depending on the number of good electrodes, signal/noise, and baseline spike rates for each animal. The threshold for avalanche detection was TH = 11±5 spikes per time bin. For experiments with overall higher MUA spike rates, presumably due to differences in experimental details like electrode quality, we chose smaller DT and smaller TH. However, this experiment-specific tuning was not necessary to support our conclusions. Indeed, we found that changes in the choice of time bin durations (in the range 5 ms > DT > 20 ms) and avalanche thresholds (in the range 5 > TH > 20) can cause small changes in the shape of the avalanche distribution (S1 Fig) and, consequently, small changes in κ. However, we emphasize that our primary conclusion–peak dynamic range near κ = 1 –was robust to such parameter variations (S2 Fig). Deviation from the reference power-law (-1.5 exponent) was quantified with κ, which is a previously developed non-parametric measure with similarities to a Kolmogorov-Smirnov statistic [19]; κ equals 1 plus the sum of 10 differences (logarithmically spaced) between the observed avalanche size distribution (recast as a cumulative distribution) and a perfect power-law (in cumulative form). In the summary plots of κ versus correlation and κ versus dynamic range, experiments were grouped according to their κ values into 13 equally spaced κ bins.
For dynamic range calculations, each point in the response curve was the average MUA response for a range of whisker speeds. The binning of whisker speeds was based on 10 equally spaced values spanning the range of observed speeds. Finally, the response curve was fit with a sigmoid function,
f(x)=Rmax1+e−b(x−c)+Rmin
where Rmin was defined as the ongoing spike count per time bin and the constants Rmax, b, and c were fitting parameters. This fit function was used to compute dynamic range, as defined in previous studies [19,20]. First, two stimulus levels, S10 and S90 are defined as those stimuli which give rise to response R10 and R90 as illustrated in Fig 4D (inset). R10 is defined as the response level Rmin+0.1(Rmax- Rmin) and R90 is defined as the response level Rmin+0.9(Rmax- Rmin). Finally, we define dynamic range as
Δ=10log10S90S10
Analysis of model data paralleled the experimental data analysis with a few exceptions. MUA spike count time series were based on time bins of duration 1 time step. The threshold for avalanche detection was 10 spikes per time bin. For computing stimulus-response curves, the response was defined as the total number of spikes from the entire network during the first 200 time steps following onset of stimulation (increase in pext). For dynamic range calculations based on model data, we did not fit the response curves with a sigmoid because they were less noisy than the experimental response curves. The correlation coefficients in Fig 3 were computed in a way to mimic the experiments, as follows. First, the 1000 model neurons were broken up into 32 groups, like the 32 electrodes in experiments. Then, a spike count timeseries was created for each group. Finally, all pairwise zero-lag Pearson correlation coefficients were computed and averaged together.
|
10.1371/journal.pbio.1000337 | Enhanced Memory for Scenes Presented at Behaviorally Relevant Points in Time | The ability to remember a briefly presented scene depends on a number of factors, such as its saliency, novelty, degree of threat, or behavioral relevance to a task. Here, however, we show that the encoding of a scene into memory may depend not only on what the scene contains but also when it occurs. Participants performed an attentionally demanding target detection task at fixation while also viewing a rapid sequence of full-field photographs of urban and natural scenes. Participants were then tested on whether they recognized a specific scene from the previous sequence. We found that scenes were recognized reliably only when presented concurrently with a target at fixation. This is evidence of a mechanism where traces of a visual scene are automatically encoded into memory at behaviorally relevant points in time regardless of the spatial focus of attention.
| What determines whether a visual scene is remembered or forgotten? The ability to remember a briefly presented scene depends on a number of factors, such as its saliency, novelty, degree of threat, or relevance to a behavioral outcome. Generally, attention is thought to be key, in that you can only remember the part of a visual scene you were paying attention to at any given moment. Here, we show that memory for visual scenes may not depend on your attention or what a scene contains, but when the scene is presented. In this study, attention to one task enhances recognition performance for scenes in a second task only in situations when the first task has behavioral relevance. Our results suggest a mechanism where traces of a visual scene are automatically encoded into memory, even though the scene is not the spatial focus of attention.
| Photographs of urban and natural scenes can be perceived and understood very quickly. However, to form a memory of a scene requires substantially more processing time. The dissociation between scene perception and memory has been shown by using rapid serial visual presentation (RSVP) to present a series of images and measuring detection and recognition performance while manipulating exposure duration [1]. These studies have shown that scene understanding requires 100 ms or less while memory formation requires at least an additional 300 ms of processing [1]–[5]. The amount of time required for memory formation is dependent on a number of factors; for example, it may take longer to form a memory if images from the set being remembered are highly confusable and similar [4].
There are a number of factors that can increase the memorability of a scene. For example, any feature that increases its “distinctiveness” or novelty—from low-level image features (e.g., a low contrast foggy scene among high contrast daylight scenes) to high-level semantic information—can lead to enhanced scene memory [6],[7]. Novelty is often believed to transiently increase attention, which leads to enhanced memory—a contention supported by experiments suggesting that spatial attention is necessary for a visual item to be encoded into memory [8]–[11]. In addition, the processing of novel events is known to rely on unique neural processing [12]–[17].
Although particularly salient or distinctive information in a scene enhances scene memory, we hypothesized that scene memory would also be enhanced at specific moments in time. A clear example is “flashbulb memory,” where details of the context in which people experience shocking news are stored into long-term memory such as where they were, what they were doing, and with whom they were [18]. This suggests the hypothesis that there may be a mechanism in which unattended (but not necessarily physically salient, novel, or threatening) information is implicitly encoded at behaviorally relevant points in time. We explored this hypothesis by testing participants' ability to recognize a particular scene as a member of a sequence of rapidly presented scenes while performing a demanding detection task at fixation. We found that recognition memory was enhanced for test scenes presented concurrently with an unrelated target at fixation. This is evidence of a mechanism where traces of a visual scene are automatically encoded into memory at behaviorally relevant points in time—operationally defined as a point of time that is important for the future execution or completion of an auditory or visual task—regardless of the spatial focus of attention.
We adapted a standard RSVP task [19] into an RSVP recognition memory task similar to other paradigms used to measure recognition memory for scenes [1],[20]–[22]. In Experiment 1, after being familiarized with a large set of photographs of natural and urban scenes, participants viewed a sequence of 16 scenes presented in an RSVP. Each sequence was then followed by a single test scene in which participants were asked whether they recognized the test scene from the previous RSVP sequence. A typical display sequence is shown in Figure 1. Baseline or chance performance on this task was 50%. Results for the scene recognition task are shown in the grey bar in Figure 2. A t test showed that participants performed no better than chance, 51.32%±4.03%, t(11) = 0.3079, p = 0.7639. Here, participants were unable to recognize whether or not a specific test scene had just appeared in the prior sequence, suggesting a previously unknown difficulty in recognizing a familiar and meaningful scene from short-term memory.
In Experiment 2, the same set of scenes was presented, but attention was directed to a demanding task at fixation where the goal was to identify a white target letter among a stream of black distractor letters. As in Experiment 1, one scene was presented immediately after each sequence for the recognition test. Mean performance on the letter identification task in Experiment 2 was 95.22%±1.09%, suggesting that participants were complying with instructions to focus their attentional resources on the fixation task. Results for the scene recognition task in Experiment 2 are shown in the white and black bars in Figure 2. The black bar shows recognition performance for scenes presented during distractor frames (black letters). For scenes presented behind black, non-target letters, performance remained at chance—52.49%±1.66%, t = 0.5951, p = 0.5638.
Surprisingly, scene recognition was significantly greater than chance for test scenes presented concurrently with white target letters (white bar in Figure 2, 67.21%±3.82%). A paired-samples t test reveals a significant difference between recognition task accuracy for test scenes that had previously been presented with black distractor letters versus white target letters, suggesting that scenes presented concurrently with white target letters were remembered better, t(10) = 2.746, p = 0.021. An additional remarkable feature of Experiment 2 was that participants claimed to have no awareness of their enhanced performance. In debriefing after Experiment 2, all participants claimed that they could not perform the scene recognition task despite performing near 70% on target-present test scenes.
We next explored whether this improved performance for scene recognition at the time of target detection was specific to detecting visual targets. Participants performed an auditory target detection task while viewing sequences of scenes as in Experiments 1 and 2. Displays and timing parameters were identical to Experiment 2 except that the alphabetical letters were removed from the scenes and replaced with a fixation marker. With every scene, a baseline auditory tone was presented and a unique tone was designated as the target. Mean performance on the auditory detection task was 90.15%±8.19%, which suggests that participants were complying with instructions to focus their attentional resources on the auditory task.
Scene recognition accuracy for Experiment 3 is presented in Figure 3. Similar to Experiment 2, participants performed near chance levels for scenes presented concurrently with distractor tones, 53.59%±1.65%, t(10) = 0.7290, p = 0.4827. However, performance for scenes presented concurrently with target tones were more accurately encoded into memory, 64.78%±3.69%, t(10) = 3.573, p = 0.005. This shows that enhanced scene encoding occurs for targets detected across modalities, suggesting that the concept of “behavioral relevance in time” is independent of modality.
In both Experiments 2 and 3, the attended targets were perceptually novel compared to distractor stimuli. Thus, enhanced encoding of scenes during target presentation may be simply due to the physical novelty of the stimuli and not due to performing the detection task. To test this, we used stimuli identical to Experiment 2 including the letter stream at fixation, but participants were instructed to ignore the letters and only perform the scene recognition memory task.
Given that the white letter serves as a perceptually novel event, one might expect enhanced performance for scenes presented concurrently with the novel event. However, recognition performance (shown in Figure 4) was at chance for both test scenes presented concurrently with black distractor letters and with novel white letters, t(14) = 0.6798, p = 0.5077, and t(14) = 0.8373, p = 0.4165, respectively. A paired-samples t test revealed no significant differences for test scenes presented concurrently with black letters (52.89%±1.33%) and novel white letters (53.13%±3.96%), t(14) = 0.1494, p = 0.8834, suggesting that the enhanced performance in prior experiments was not simply due to the perceptual novelty of the physical stimulus.
Together, these four experiments demonstrate that at behaviorally relevant points in time—operationally defined as a point of time that is important for the future execution or completion of an auditory or visual task—a memory trace of the visual field is automatically encoded into memory, enhancing later recognition of information even at unattended regions of visual space. This “screen capture” mechanism is likely to play an important role in the retrospective analysis of important events.
A defining characteristic of the human visual system is its ability to rapidly extract details of a scene, but it takes substantially longer to encode a scene into memory [1],[4]. However, recognition memory for scenes is remarkably good when given sufficient encoding time [23]. Traditionally, the encoding of pictures into memory has been studied using single-task, undivided attention paradigms exploring the effects of stimulus duration and visual and conceptual masking on effective encoding and later memory. Consequently, less is known about memory encoding under conditions of reduced attention.
What determines whether an item is remembered or forgotten? It has been shown that observers are very poor in discriminating or recognizing obvious and significant changes in scenes unless they happen to be attending to the item that was changed [24],[25]. As an extension of this, it is generally believed that focused spatial attention is necessary for a visual item to be encoded into memory [8]–[11]. In the present studies, however, even focal attention on the scenes in Experiment 1 was not sufficient to maintain familiar scenes in short-term memory.
In sharp contrast, in Experiment 2, when spatial attention was directed towards fixation on an attentionally demanding task, the presentation of a target item resulted in enhanced recognition memory for the scene presented concurrently with the target in the background. This result suggests a new mechanism that may play a role in determining what and when information about a scene is encoded into memory. A counterintuitive feature of this enhanced recognition memory effect is that it occurs in spite of the known effects of focusing of spatial attention around a target item [26],[27]. Our results indicate that target detection, engagement, or processing has a strong, non-stimulus-specific influence on memory formation—the enhanced encoding into memory of all items that are temporally coincident with a behaviorally relevant target event. The data suggest that behaviorally relevant points in time trigger a “temporal novelty” effect on memory encoding that appears to be a sufficient prerequisite for the successful encoding of visual stimuli into memory under conditions of reduced attention [28]–[30].
It is unlikely that this non-stimulus-specific influence on memory formation was due to the attentional blink [19],[31] suppressing scenes presented after focal targets were identified; indeed, recognition memory for scenes presented immediately before or after the temporal positions of the targets was still at chance. Moreover, the rate of presentation (two pictures/s) is considerably slower than rates that produce an attentional blink. In addition, recognition memory for the scenes presented before or after the temporal positions of the targets being at chance also suggests that the effects were not due to a general arousal [32] triggered by the onset of a perceptually novel stimulus and thereby increasing recognition memory for all subsequent scenes presented after the targets.
Perceptual learning for task-irrelevant peripheral stimuli can occur when attention is focused away from the peripheral stimuli and towards fixation and these learning effects are greatest for peripheral stimuli presented at the time of foveal target detection [33]–[35]. These results were surprising because it had generally been assumed that perceptual learning requires attention be focused on the target stimulus being learned. However, even in the absence of attention, it must be necessary for the target stimulus being learned to be encoded into memory for learning to occur. Here, we show that short-term memory for a peripheral scene is enhanced when it is presented at a behaviorally relevant point in time. It seems likely that a version of this “task-related screen capture” is one of the mechanisms that could support the phenomenon of perceptual learning in the absence of attention.
Recently, researchers have shown that repeated presentation of movie clips produces detectable “memory traces” in subsequent resting state activity in cat visual cortex [36]. It is plausible that given a behaviorally relevant point in time, a strong reverberation or memory trace was triggered and the residual of this imprint was being tapped into when performing the scene recognition task.
Finally, one might assume these results suggest that the processes associated with enhanced vividness, memory, and attention for novel events act globally throughout the visual field; however, Experiment 4 suggests that at first glance, perceptual novelty is not the source of these effects. When passively viewing the same displays as Experiment 2 and asked to perform the recognition memory task while ignoring the black distractor letters and novel white target letters, no significant differences were found in recognition performance. Overall, our results suggest a mechanism where traces of a visual scene are automatically encoded into memory at behaviorally relevant points in time regardless of the spatial focus of attention.
All participants reported normal or corrected-to-normal visual acuity and gave informed consent to participate in this experiment, which was approved by the University of Washington Human Subjects Institutional Review Board. In every experiment prior to testing, participants performed a practice block of 24 trials. Each participant was then tested for a total of 240 trials, in 10 blocks of 24 trials. Blocks were separated by brief breaks.
Different participants participated in each of the five experiments. All received financial compensation in one 1 h session. Experiment 1 consisted of 12 participants (10 females, 2 males). Experiment 2 consisted of 11 participants (7 females, 4 males). Experiment 3 consisted of 11 participants (6 females, 5 males). Experiment 4 consisted of 15 participants (11 females, 4 males).
Displays were presented on a 45 cm ViewSonic Graphics Series G90fB monitor at 1024×768 resolution, refreshed at 60 Hz. Participants sat with their eyes approximately 50 cm from the screen. The backgrounds of all displays were gray (15 cd/m2). Display items consisted of 192 700×700 pixel (28.07 degrees of visual angle) photographs depicting natural or urban scenes from eight distinct categories (i.e., mountains, cityscapes, etc). Scenes were obtained from the LabelMe Natural and Urban Scenes database [37] at 250×250 pixels of resolution, then up-sampled to 700×700 pixels of resolution.
Display items during the experiment were sampled from the 192 scenes with replacement. In each sequence, observers were shown 16 of these scenes at 133 ms per scene, followed by a blank ISI of 367 ms for a SOA of 500 ms.
All experiments (1, 2, 3, and 4) used the scene recognition task. Following each rapid sequence of 16 full-field scenes, observers were presented with a test scene and asked to recall whether the test scene appeared in the previous RSVP sequence of scenes. The test scene was presented for 3,000 ms or until participants responded to whether they recognized the test scene from the RSVP stream with a “Y” or “N” on the keyboard. In 50% of the trials, the test scene was randomly drawn from the scenes presented in serial positions 9 to 16 of the RSVP; in the other 50% of trials, the test scene was drawn from the set of scenes not shown in the current RSVP stream. When test scenes were drawn from serial positions 9 to 16, there was a random 1/8 chance that the test scene matched the scene presented behind the white target letter in the RSVP stream, meaning that the white target letter task was irrelevant to the secondary recognition memory task and did not predict the test scene participants would be tested on. All scenes were sampled from our database with replacement. Distractors and target letters were embedded in randomly selected scenes over the entire session.
It is important to note that although our scene recognition task is similar to earlier studies that tested picture memory for novel scenes [1],[20], our task requires the participant to remember whether an already-familiar test picture appeared in the most recent sequence. Previous studies have used unfamiliar pictures on each trial. We presume that observers would have no difficulty detecting the presence or absence of a familiar scene in a sequence if they knew beforehand what scene to detect [3],[4]. In addition to the main result, the last scene in the RSVP sequence was often recognized with higher accuracy, in line with well-known recency effects of memory [38]–[43] and the fact that the last scene was not conceptually masked by a subsequent item. In Experiment 1, we only tested the second half of scenes presented in the RSVP to maintain consistency with subsequent experiments and therefore do not have data on potential performance differences for the first scene presented in the RSVP sequence. This new recognition memory task that measured participants' ability to encode a familiar set of scenes into short-term memory using RSVP sequences served as a starting point for examining potential temporally related enhancements to the encoding of briefly presented scenes into memory.
For the letter detection task (Experiment 2), a gray aperture (1 degree of visual angle) was embedded in the center of each scene and a random alphabetical letter (20 font size) was centered within the aperture. New random letters were embedded into the gray apertures of every scene, with the only requirement being that no duplicate letters could be presented within the same trial. Alphabetical letters were either black (indicating its identity as a distractor) or white (indicating its identity as a target; see Figure 1). In every trial, random black alphabetical letters representing distractors were embedded at central fixation in 15 of the scenes and a random white alphabetical letter representing the target was embedded in 1 scene. White target letters could only appear concurrently with scenes presented in serial positions 9 to 16 to avoid having white target letters presented at the onset of a RSVP stream. Participants were instructed to fixate on a point in the center of the screen and search for and identify a white target letter while memorizing the series of 16 scenes presented in RSVP.
In Experiment 2, immediately following the RSVP, participants were instructed to type the letter key corresponding to the identity of the white target letter for the current trial. Following the response to the letter detection task, participants performed the scene detection task. Participants were instructed to ignore the letter stream in Experiment 4.
The auditory target detection task in Experiment 3 was similar to the letter detection task in Experiment 2 except the alphabetical letters were removed from the apertures centered in the scenes. Instead, an auditory tone was presented with each scene. Tones were sampled at 44,000 Hz, with durations of 50 ms. Baseline tones were presented at 261.50 Hz, while target tones were either 130.75 Hz or 523.0 Hz. Immediately following the RSVP stream, participants were instructed to discriminate the pitch of the unique tone as either lower or higher via key press, then were again presented with a test scene and asked to recall whether they recognized the scene from the RSVP stream.
|
10.1371/journal.ppat.1002518 | Concerted Actions of a Thermo-labile Regulator and a Unique Intergenic RNA Thermosensor Control Yersinia Virulence | Expression of all Yersinia pathogenicity factors encoded on the virulence plasmid, including the yop effector and the ysc type III secretion genes, is controlled by the transcriptional activator LcrF in response to temperature. Here, we show that a protein- and RNA-dependent hierarchy of thermosensors induce LcrF synthesis at body temperature. Thermally regulated transcription of lcrF is modest and mediated by the thermo-sensitive modulator YmoA, which represses transcription from a single promoter located far upstream of the yscW-lcrF operon at moderate temperatures. The transcriptional response is complemented by a second layer of temperature-control induced by a unique cis-acting RNA element located within the intergenic region of the yscW-lcrF transcript. Structure probing demonstrated that this region forms a secondary structure composed of two stemloops at 25°C. The second hairpin sequesters the lcrF ribosomal binding site by a stretch of four uracils. Opening of this structure was favored at 37°C and permitted ribosome binding at host body temperature. Our study further provides experimental evidence for the biological relevance of an RNA thermometer in an animal model. Following oral infections in mice, we found that two different Y. pseudotuberculosis patient isolates expressing a stabilized thermometer variant were strongly reduced in their ability to disseminate into the Peyer's patches, liver and spleen and have fully lost their lethality. Intriguingly, Yersinia strains with a destabilized version of the thermosensor were attenuated or exhibited a similar, but not a higher mortality. This illustrates that the RNA thermometer is the decisive control element providing just the appropriate amounts of LcrF protein for optimal infection efficiency.
| Many important virulence genes remain silent at moderate temperatures in external environments and are rapidly and strongly induced by a sudden temperature upshift sensed upon host entry. Thermal activation of virulence gene transcription is frequently described, but post-transcriptional control mechanisms implicated in temperature-sensing and induction of virulence factor synthesis are less evident. Here, we present a novel two-layer regulatory system implicating a protein- and an RNA-dependent thermosensor controlling synthesis of the most crucial virulence activator LcrF (VirF) of pathogenic yersiniae. In this case, moderate function of a thermosensitive gene silencer is coupled with the more dominant action of a unique intergenic two-stemloop RNA thermometer. Thermally-induced conformational changes in this RNA element control the transition between a ‘closed’ and an ‘open’ structure which allows ribosome access and translation of the lcrF/virF transcript. This mechanism guarantees optimal virulence factor production during the course of an infection, ideal for survival and multiplication of yersiniae within their warm-blooded hosts. The hierarchical concept combining two temperature-sensing modules constitutes a new example of how bacterial pathogens use complementing strategies to allow rapid, energetically cheap and fine-tuned adaptation of their virulence traits.
| Pathogenic yersiniae, including Y. pestis, the causative agent of the bubonic plague, and the two enteric species Y. enterocolitica and Y. pseudotuberculosis which cause gut-associated diseases (yersiniosis) such as enteritis, diarrhea and mesenterial lymphadenitis express different sets of virulence factors important for different stages of the infection process [1]–[2]. It is well known that most of the Yersinia virulence genes are tightly controlled in response to temperature [3].
Some of the early stage virulence factors, including the primary internalization factor invasin of both enteric Yersinia species, are mostly produced at moderate temperatures to allow efficient trespassing of the intestinal epithelial barrier shortly after infection [4]–[6]. These virulence genes are controlled by RovA, an intrinsic protein thermometer, which undergoes a conformation change upon a temperature shift from 25°C to 37°C, that reduces its DNA-binding capacity and renders it more susceptible to proteolysis [7]–[9].
Most other known Yersinia virulence genes remain silent outside the mammalian hosts and are only induced after host entry in response to the sudden increase in temperature. One important set of thermo-induced virulence factors is encoded on the 70 kb Yersinia virulence plasmid pYV (pCD1 in Y. pestis) [10]. These pathogenicity factors are crucial to avoid phagocytosis or other attacks by the innate immune defense system and comprise a type III secretion system (T3SS), the secreted Yersinia outer proteins (Yops) and regulatory components of the secretion system [11]–[13]. The Yop secretion genes (ysc) are organized in two operons yscB-L (virC operon) and yscN-U, or encoded elsewhere (e.g. yscW, yscX, yscY and yscV) on pYV [10] and are required for the formation of the T3S apparatus (injectisome). Body temperature (but not 20–25°C) and host cell contact trigger expression and translocation of Yop proteins by the T3S machinery into the cytoplasm of targeted host cells [14]–[17]. The Yop proteins can be divided into the group of translocators implicated in the formation of the translocation pore and the Yop effector proteins which manipulate numerous signal transduction pathways to prevent phagocytosis and the production of proinflammatory cytokines [18]–[21].
Expression of the majority of pYV-encoded virulence genes (yadA, yop, lcr and ysc genes for T3S and regulation) is induced by temperatures above 30°C in all pathogenic Yersinia species. Temperature-dependent induction of these genes requires the AraC-type DNA-binding protein LcrF (VirF in Y. enterocolitica) [22]–[24]. The LcrF protein contains a poorly conserved N-terminal oligomerization domain which is connected to a flexible highly conserved C-terminus with two helix-turn-helix DNA-binding motifs [25]. It exhibits high homology to the main regulator of T3S in Pseudomonas aeruginosa, ExsA and has been shown to bind specifically to TTTaGYcTtTat DNA motifs in the promoter regions of yopE, lcrG, virC and yopH [26]. The transcriptional activator LcrF is mainly produced at 37°C. Hoe and Goguen [27] showed that the lcrF mRNA produced in E. coli or Y. pestis could not be translated at 26°C, but was readily translated at 37°C. Based on predicted mRNA structure, these authors proposed that translation was dependent on melting of a stem-loop which sequestered the lcrF ribosomal binding site. Calculated thermal stability agreed well with observed translation, but no experimental work testing this hypothesis by manipulating stability of the structure was performed. In contrast, for Y. enterocolitica it has been reported that transcription of the lcrF homologous gene virF is increased at higher temperatures. This activation was shown to depend on topological changes and thermo-induced melting of intrinsically bent DNA identified upstream of the lcrF/virF gene [28]–[29]. Also chromosomally encoded factors that contribute to the temperature-dependent regulation of yadA and yop transcription have been identified in Y. enterocolitica. Below 30°C, induction of these virulence genes was only observed in the absence of the Yersinia modulator A (YmoA) [30]–[31]. YmoA belongs to the superfamily of nucleoid-associated proteins and shares 82% sequence identity with the regulator of “high hemolysin activity” (Hha) in E. coli and Salmonella [32]. The E. coli Hha protein represses the transcription of the hlyCABD operon encoding the pore-forming toxin hemolysin at moderate temperatures [33]–[34]. YmoA was shown to influence DNA supercoiling and forms heterodimers with the nucleoid-associated protein H-NS [33]–[36]. However, YmoA or H-NS binding to pYV promoter sequences has never been reported. Hence, the molecular mechanisms by which YmoA controls yadA and yop gene expression is still unclear. A recent analysis of type III secretion in Y. pestis indicated that regulated proteolysis of YmoA by the ATP-dependent Clp and Lon proteases plays an important role in the temperature-dependent expression of the type III secretion operons [37]. It was shown that YmoA is rapidly degraded at 37°C, but remains stable at environmental temperatures [37]. Whether the thermo-control mechanisms of LcrF vary between Y. pestis and Y. enterocolitica or whether they are connected, and if so, how they contribute to LcrF production in the different species remained elusive.
In this study, we investigated the molecular mechanism underlying thermoregulated production of the LcrF virulence regulator of Y. pseudotuberculosis. We found that concerted actions of the thermo-labile YmoA regulator protein and an unusual intergenic RNA thermosensor assured best possible production of LcrF for the highest infection efficiency. YmoA repressed lcrF transcription through sequences located within the 5′-UTR of yscW located upstream of the lcrF gene, and contributed moderately to the thermo-dependent production of LcrF. This activity is supplemented by a two-hairpin RNA thermometer composed of four uracil residues (fourU) that pair with the ribosome-binding site (AGGA) within the intergenic region of the yscW-lcrF mRNA. Using a mouse model system we provide evidence that this RNA thermosensor is mainly responsible for thermo-induced LcrF production, and show that its function is relevant for a high pathogenic potential, optimal survival and multiplication of Yersinia during infection.
The AraC-type transcriptional activator protein LcrF induces the expression of crucial Yersinia pathogenicity factors (e.g. YadA, T3SS and Yop effectors) in response to temperature. Initial efforts in this study to unravel the molecular control mechanisms of lcrF expression in Y. pseudotuberculosis demonstrated that the lcrF gene is organized in an operon with yscW (formerly named virG) located 124 bp upstream of the lcrF coding region on the Yersinia virulence plasmid pYV (Figure 1A). As shown in Figure 1B, a yscW-lcrF-lacZ (pSF4) and a yscW-lacZ (pKB10) translational fusion harboring the yscW regulatory region up to position −572 relative to the yscW start codon were expressed, whereas a construct carrying yscW sequences to position −7 (pSF3) was not. The yscW-lcrF-lacZ fusion was thermo-regulated in dependence of the YmoA protein. Expression was about 2-fold increased in the ymoA mutant strain and showed a significantly higher expression at 37°C than at 25°C (Figure 1B). To confirm this result, western blot analysis was performed to detect the LcrF protein in cell extracts from the Y. pseudotuberculosis wildtype strain YPIII and the isogenic ymoA mutant YP50 grown at 25°C and 37°C. As shown in Figure 1C, the LcrF protein could only be detected in extracts of the ymoA mutant but not in the wildtype strain when the bacteria were grown at 25°C. In contrast, LcrF production was significantly increased and detectable in both strains at 37°C, whereby the overall level of LcrF was significantly higher in the ymoA-deficient strain. This indicated that lcrF expression occurs from a temperature- and YmoA-dependent promoter located upstream of the yscW gene.
In order to investigate yscW-lcrF transcription in more detail, total RNA of Y. pseudotuberculosis was prepared for Northern blot analysis using an lcrF specific probe (Figure 2). The yscW-lcrF transcript was found to be highly unstable and was rapidly degraded into lower molecular weight transcripts in the wildtype (Figure 2). In contrast, higher concentrations and higher molecular weight transcripts were detectable in the ymoA mutant strain consistent with the conclusion that yscW and lcrF originate from the same promoter. Moreover, as judged from the length of the yscW-lcrF transcript, the transcriptional start site appeared to be about 300 bp upstream of the yscW start codon, leading to the formation of a long 5′-untranslated region (5′-UTR).
To identify the yscW-lcrF promoter we performed primer extension analysis. We found that the transcription of the yscW-lcrF operon starts at a G found 264 nt upstream of the start codon GTG of yscW with a −35 and a −10 region of a typical σ70-dependent promoter (Figure 3) leading to a 264 nt 5′-UTR. Several shorter reverse transcripts were consistent with the Northern results suggesting rapid processing of the yscW-lcrF transcript.
Increased expression of the yscW-lacZ and yscW-lcrF-lacZ fusions in the ymoA deficient Yersinia strain suggested that YmoA influences expression on the transcriptional level (Figure 1B, S1). Continuous deletions of the promoter region showed that elimination of the identified promoter region by a 5′-upstream deletion up to position −2 abrogated transcription of the fusion construct and confirmed presence of a single promoter driving yscW-lcrF expression (Figure S1A, 3C). Further analysis demonstrated that YmoA-dependency was maintained when sequences upstream of the yscW promoter were deleted, but it was lost when the 5′-UTR of yscW was removed (Figure S1A,B). This indicated that YmoA acts through sequences located downstream of the yscW promoter.
Next, we tested whether YmoA influence on yscW-lcrF was direct or involves (an)other regulatory factor(s). Experimental evidence support the hypothesis that members of the YmoA(Hha) protein family modulate gene expression by interacting with the nucleoid-structuring DNA-binding protein H-NS or its paralogs [36], [38]. However, other studies reported that the Hha/YmoA protein binds specifically to regulatory sequences of virulence genes. Unfortunately copurification of H-NS was not ruled out in these studies [39]–[41]. In order to test YmoA binding, YmoA was overexpressed and purified from E. coli strain KB4 (Δhns, ΔstpA, Δhha) deficient of all E. coli full-length and partial H-NS family proteins and used for band shift analysis with an yscW promoter fragment harboring the entire 5′-UTR. However, even at very high protein concentrations YmoA was not able to interact specifically with the yscW regulatory region (Figure S2A). In addition, we purified YmoA overexpressed in E. coli strain KB4 also expressing the Y. pseudotuberculosis hns gene. This YmoA protein sample included copurified H-NSY.pstb (data not shown) and was able to interact specifically with the 5′-UTR sequences of the yscW gene (Figure S2C). We further expressed and purified H-NSY.pstb in the absence of YmoA and found that also H-NSY.pstb alone is capable to interact with the yscW regulatory sequences (Figure S2B). This indicated that YmoA influences yscW-lcrF expression directly and this involves heterocomplex formation with H-NS.
To confirm these data we also analyzed whether YmoA influence on thermal regulation of LcrF is lost, when the 5′-UTR important for H-NS/YmoA binding is absent. To do so, we compared expression of the yscW-lcrF-lacZ construct and a derived deletion variant of this fusion (yscW(Δ13–241)-lcrF-lacZ) at 25°C and 37°C. We found that the expression level is still thermoregulated, but lcrF transcription became independent of YmoA (Figure S2D). This clearly demonstrated that YmoA influences expression of lcrF via the 5′-UTR region of the yscW gene.
The YmoA protein of Y. pestis was shown to be subject to proteolysis by the Lon- and ClpP proteases at 37°C but not at 25°C [37], and this post-translational control was also observed for YmoA in Y. pseudotuberculosis YPIII (K. Böhme, unpublished results). However, expression of the yscW-lcrF-lacZ fusion and LcrF synthesis was still thermoregulated in the ymoA-deficient strain (Figure 1B,C), suggesting that contribution of YmoA to lcrF thermoregulation is rather small and predominantly mediated by an additional YmoA-independent control mechanism.
To localize the region responsible for this type of control, we exchanged the yscW promoter (PyscW) against the PBAD promoter and analyzed yscW-lcrF-lacZ expression after induction with 0.05% arabinose at 25°C and 37°C. Thermoregulation was maintained when yscW-lcrF was transcribed by PBAD independent whether the fusion was expressed in E. coli or in Y. pseudotuberculosis (Figure 4, S3). In contrast, expression of lacZ fused to the 5′-UTR of the Y. pseudotuberculosis 6-phosphogluconate dehydrogenase gene (gnd) in the identical vector system was not affected by the growth temperature. These experiments strongly suggested that the temperature control of lcrF expression is mediated by a post-transcriptional mechanism as previously demonstrated in Y. pestis [24] and is independent of Yersinia-specific factors. Deletions removing different portions of the yscW locus or the entire yscW gene further demonstrated that presence of the yscW gene is dispensable and that the intergenic region of the yscW-lcrF operon is sufficient for temperature control of lcrF translation (Figure 4, S4).
Systemic inspection of the 124 nt yscW-lcrF intergenic region, comparison with related bacteria (Y. pestis, Y. enterocolitica) and secondary structure predictions by Mfold [42] revealed a potential RNA structure composed of two stemloops (hairpin I and II) with a free energy of −19.67 kcal mol−1 (Figure 5A). The first hairpin (57 nt) consists of three base-pairing stretches interrupted by two internal loops and is separated from the second hairpin (hairpin II, 46 nt) by 11 nt. In hairpin II, the ribosomal binding site (RBS) of lcrF pairs with a stretch of four uracil residues (fourU) located 26 to 29 nt upstream of the translation initiation site of lcrF. This structure resembles a fourU thermometer identified in the 5′-untranslated region of the Salmonella agsA gene [43]. Presence of two small loops (C-5/A-6/A-38 and A-12/A-31/A-32) and three imperfect base-pairs in the RBS region (G-15/U-28; G-16/U-27; U-19/G-24) in hairpin II suggested a temperature-labile structure prone to melting at increasing temperatures.
To investigate whether the intergenic region of yscW-lcrF forms a functional RNA thermometer we deleted hairpin I (Dhairpin I: −111/−57) or parts of hairpin II (Dhairpin II: −44/−25) and introduced stabilizing (AG-46/-45CC; UU-28/-27CC) and destabilizing (AUA-36/-34CCC; GUU-30/-28AAA) point mutations into in the PBAD::lcrF-lacZ fusion construct and in the yscW-lcrF intergenic region of the virulence plasmid pYV (Figure 5A). Absence of hairpin I (Δhairpin I) resulted in a significant reduction of lcrF thermo-induction from 5- to 2-fold (Figure 5B,C). Expression was already high at 25°C and induction was lost when sequences implicated in the formation of hairpin II (Δhairpin II) were deleted. Similarly, thermo-induced expression of lcrF was strongly decreased in both mutations designed to destabilize hairpin II, whereas expression of variants with stabilizing mutations remained repressed upon a temperature upshift and only very small amounts of the LcrF protein were produced at both 25°C and 37°C (Figure 5B,C). Increase of LcrF levels in the destabilized mutant from 25°C and 37°C demonstrates a two-layer regulation by the thermo-labile YmoA protein and the RNA thermometer. In the following experiments the stabilizing mutation UU-28/-27CC and the destabilizing mutation GUU-30/-28AAA are also referred to as ‘closed’ and ‘open’, respectively. In summary, our data demonstrate that the intergenic region of the yscW-lcrF operon contains a thermo-responsive RNA element composed of two hairpins that mediate post-transcriptional control in an RNA thermometer-like manner.
In order to examine the architecture of the predicted RNA structure experimentally, we determined the structure and the nature of thermo-induced conformational changes of the intergenic yscW-lcrF mRNA by enzymatic probing at 25°C and 37°C using RNAse T1 (cleaves 3′ of unpaired guanines) and double-strand specific RNase V1. Due to the large size of the full-length yscW-lcrF transcript, the structure of a shorter RNA fragment including the entire yscW-lcrF intergenic region (5′-UTR of lcrF) was probed (Figure 6A, B). The cleavage pattern at 25°C was in full agreement with the predicted two hairpin structure (Figure 5A). RNase T1 digestion at positions −81 to −85 and positions −101 to −102 as well as the sensitivity of adjacent regions to RNase V1 cleavage (positions −71 to −68; −78 to −75; −87 to −90; −94 to −99; −103 to −108) confirmed the predicted secondary structure of hairpin I containing three loop segments. Also hairpin II seems to form the predicted structure (protection to RNase V1 at positions −24 and −20). Consistent with its function as a temperature sensor, this stemloop is more dynamic and adapts a thermo-sensitive conformation that seems to open after a shift to 37°C. As shown in Figure 6, the stem region including the imperfect UUUU/AGGA base pairs with the RBS and flanking regions is more resistant to RNases T1 at 25°C than at 37°C. Temperature-induced melting of the stemloop II at 37°C is also supported by digest with the RNase V1, which is less active at 25°C. To confirm that RNase T1 cleavage at the RBS is the result of structural changes and not the result of an induced activity of RNase T1 at higher temperatures, we quantified the intensities of the T1 cleavage sites at 25°C and 37°C using the AlphaEaseFC program (Cell Biosciences, USA). The cleavage intensity at 37°C relative to 25°C was 4-fold at G15/G16 within the RBS, compared to 1.3-fold at G82 in hairpin I, and 1.7- to 2.3-fold at adjacent T1 cleavage sites (G11, G24, G45) in hairpin II. These results indicated that the RBS within hairpin II represents the primary temperature-sensing site within the lcrF RNA thermometer. We also performed enzymatic probing with the yscW-lcrF mRNA derivatives including the stabilizing and destabilizing nucleotide exchanges in hairpin II. Analysis of the UU-28/-27CC mutation indicated the generation of a thermostable stemloop II structure as neither the RBS nor the anti-RBS fourU sequence was accessible to RNases T1 at 25°C and 37°C (Figure 6C). Complete protection of the ribosomal binding site is in full agreement with reduced expression of lcrF (Figure 5B). In contrast, introduction of the derepressing GUU-30/-28AAA exchanges resulted in an altered, less stable structure in which the RBS sequence is more accessible at 25°C and 37°C (Figure 6D).
To demonstrate temperature-dependent interaction of the 30S ribosome with the RBS in the intergenic region of the yscW-lcrF mRNA, we performed toeprinting analysis. Ribosomal subunits and the initiator tRNAfMet were added after annealing of the lcrF specific reverse primer to the yscW-lcrF template and incubated at 25°C or 37°C. The primer extension reaction was not inhibited at 25°C and/or in the absence of the 30S ribosome. However, at 37°C a toeprint (prematurely terminated product) was detected at position +14/+18 relative to the translational start site, demonstrating the formation of a ternary translation initiation complex composed of the yscW-lcrF mRNA, the 30S ribosome and tRNAfMet (Figure 7). More prominent toeprint signals were observed when the destabilizing GUU-30/-28AAA exchanges were introduced, whereas significantly read-through up to the full length transcript and less toeprint signals were found with the stabilizing UU-28/-27CC variant (Figure 7). Taken together, this experiment showed that a thermo-induced interaction of the ribosome with the lcrF translation initiation site is facilitated at body temperature and occurs in the absence of any other bacterial factors.
To analyze whether this mechanism of post-transcriptional thermoregulation has an important impact on virulence, we first tested whether introduction of the ‘open’ (GUU-30/-28AAA) and ‘closed’ (UU-28/-27CC) mutations into the yscW-lcrF intergenic region resulted in mis-regulation of the LcrF-dependent virulence genes yadA, and yopE encoded on the Yersinia virulence plasmid pYV (Figure 8). Consistent with previous results, yadA and yopE transcription as well as YadA synthesis was temperature-induced in the wildtype. Expression was abolished in mutants with stabilizing nucleotide exchanges in the 5′-UTR of lcrF. In contrast, destabilizing substitutions led to increased yadA and yopE expression already at 25°C. The LcrF-dependent yadA-lacZ expression and LcrF synthesis increased at 37°C in the presence of the ‘open’ mutation (Figure 5, 8) which can be explained by the loss of YmoA-dependent control of lcrF expression.
In vitro, Yop secretion is generally blocked in the presence of millimolar amounts of extracellular Ca2+ but it can be induced upon Ca2+-complexation with sodium oxalate (Na2C2O4) [44]–[46]. As expected, concentration of secreted Yop proteins by the wildtype (YPIII) and the ‘open’ strain YP95 (GUU-30/-28AAA) was high at 37°C in the absence of Ca2+, but no Yops could be detected in the supernatants of the ‘closed’ strain YP90 (UU-28/-27CC) under the same growth conditions. Strikingly, although LcrF synthesis and the LcrF-dependent yopE gene expression are already induced in the derepressed strain YP95 (GUU-30/-28AAA) at 25°C, no Yop secretion was detectable after Ca2+ depletion, indicating that a temperature-dependent mechanism blocks YopE production and/or secretion at low temperatures.
In order to define the influence of the lcrF RNA thermometer on bacterial pathogenesis, we compared survival and dissemination of the Y. pseudotuberculosis wildtype YPIII and the repressed and derepressed mutant strains YP90 (UU-28/-27CC) and YP95 (GUU-30/-28AAA) in the mouse model. Presence of each strain was examined three days after intragastrically infection of a group of BALB/c mice (n = 12) with 5·108 bacteria by quantifying the number of bacteria that reached and survived in the Peyer's patches (PP), the mesenterial lymph nodes (MLN), liver and spleen. Significantly reduced numbers of the repressed YP90 (UU-28/-27CC) mutant strain were recovered from the Peyer's patches and organs (Figure 9), very similar to the lcrF mutant strain YP66 (Figure S5). We also introduced the ‘closed’ and ‘open’ mutation into the more virulent Y. pseudotuberculosis strain IP32953. Oral infections with IP32953 and the isogenic ‘open’ variant (YPIP02) led to a higher organ burden three days post infection. However, the number of bacteria was similarly reduced in the host tissues with the ‘closed’ mutant (Figure S6). This demonstrated that a repression of the fourU RNA thermometer reduced virulence and showed that the structural rearrangements of the 5′-UTR of the lcrF mRNA affects pathogenesis of both Yersinia strains. Our results also showed that introduction of derepressing nucleotide exchanges (GUU-30/-28AAA) had no or only a minor effect on the colonisation of host tissue (Figure 9, S6).
To complement the infection experiments, the potential of the different lcrF RNA thermometer mutant strains to cause a lethal infection was determined. Groups of BALB/c mice (n = 10) were infected intragastrically with 2·109 bacteria of each mutant, YP90 or YPIP01 (UU-28/-27CC) and YP95 or YPIP02 (GUU-30/-28AAA) and the wildtype strains YPIII or IP32953. Survival of the mice was followed over 14 days and date of death was recorded (Figure 10). All mice infected with the wildtype strain showed visible signs of infection by day three post infection (e.g. lethargy, rough fur) and succumbed to infection between day three to six post challenge. Strikingly, none of the mice infected with the repressed mutant strain YP90 or YPIP01 (UU-28/-27CC) developed disease symptoms and all were still alive 14 days after infection, similar to the ΔlcrF mutant strain YP66 (Figure 10). This indicated that stabilization of hairpin II renders the bacteria avirulent. In contrast, the destabilizing mutations in the lcrF RNA thermometer had no apparent effect on the initial rate of death, and did not cause a higher mortality than the wildtype over a 14-day period (Figure 10). The average day to death of mice challenged with the “open” mutant variants YP95 or YPIP02 (GUU-30/-28AAA) was similar or increased from four to seven days. Taken together, this illustrates that the RNA thermometer is crucial for virulence, as it plays an important role adjusting the appropriate amounts of LcrF for maximal pathogenicity.
Many environmental signals are sensed by enteric pathogens such as Y. pseudotuberculosis in order to induce and adjust expression of virulence factors upon host entry and during ongoing infections. Temperature is among the most important decisive parameters for an intestinal pathogen, indicating that it successfully invaded a warm-blooded host. A prerequisite for an appropriate response to temperature changes is precise thermosensing, and different principles governing the temperature-sensing mechanism have been uncovered for a variety of macromolecules [47]–[48]. Thermo-induced structural changes in supercoiled or intrinsically curved DNA have long been known to manipulate gene expression by altering the accessibility of promoter elements [49]–[50]. Recently, also regulatory proteins were shown to act as intrinsic thermosensors to adjust their DNA-binding properties [8], [51]–[53], and experimental evidence accumulated that also RNA plays a fundamental role in temperature sensing [54]–[56].
Although control of translation initiation by limiting the access to the ribosome-binding site has been reported earlier, the full dimension to which structured mRNAs contribute to thermosensing has only recently been recognized. Several distinct and structurally unrelated RNA sensors have been identified in bacteria, but almost all control the synthesis of heat shock proteins. To our knowledge only one RNA thermometer located upstream of the virulence regulator gene prfA of Listeria monocytogenes has been described to regulate virulence gene expression and invasion into cultured cells [57]. However, its impact for pathogenesis, e.g. initiation or progression of the infection has not been investigated. In this study, we report the existence of an unusual intergenic, two-stemloop RNA thermometer and provide first experimental evidence that its function is crucial for Y. pseudotuberculosis virulence in a mouse model. Two temperature-sensing modules, the thermo-sensitive virulence modulator protein YmoA and the RNA thermosensor, act in concert to optimize temperature perception and fine-tune virulence gene expression during infection (Figure 11).
A comprehensive expression analysis revealed, that the lcrF gene is transcribed from a single σ70-specific promoter of the yscW gene (formerly named virG) which is located 124 bp upstream of lcrF. Cotranscription is consistent with the observation that the yscW-lcrF locus is similar to the last two genes of the exsC-exsB-exsA operon of Pseudomonas aeruginosa required for the ExoS effector synthesis [58]. It also reconciles previous contradictory models for temperature control of lcrF (virF) expression in Y. pestis and Y. enterocolitica. Cornelis et al. showed that virF of Y. enterocolitica itself is thermoregulated at the transcriptional level [44]. In that study, virF::cat fusions and virF Northern blots demonstrated transcription activation at elevated temperature. Based on the present analysis, thermo-dependent virF expression can be explained by YmoA-dependent repression of the yscW promoter that is eliminated at higher temperatures due to increased degradation of YmoA by the Lon and Clp proteases [37].
In contrast, lcrF-lacZ reporter fusions in Y. pestis, including only 206 bp upstream of the lcrF start codon, were found to be insensitive to temperature changes, although much higher levels of the LcrF protein were produced in Y. pestis with raising temperature [24], [27]. This implied that a different post-transcriptional mechanism modulates LcrF levels in response to temperature in this organism. A simple model for lcrF thermal regulation has been suggested in which a predicted thermo-labile stem-loop (identical to the upper part of hairpin II) sequesters the lcrF ribosomal binding site [27], [43], but its function has never been proven.
Reporter gene assays and a detailed structure-function analysis of the isolated intergenic region in this study provide evidence for a functional RNA thermometer in which temperature regulation of lcrF is mediated in the absence of the natural promoter. Structural probing experiments demonstrated the formation of two hairpins of which hairpin II includes a consecutive stretch of four uridine nucleotides (fourU motif), which base pair with the RBS, and two internal unpaired bulges. Mutational analyses and toeprinting experiments further showed that this RNA structure is sufficiently stable to resist melting at moderate temperature (25°C), but it allows partial unfolding at body temperature (37°C) which permits access of ribosomes and initiation of lcrF translation. Hairpin I was not essential for thermosensing, but it seems to support proper folding and/or the stability of the ‘closed’ RNA thermometer structure, as generally higher amounts of the LcrF protein were detectable in Δloop1 mutant variants. Importance of this RNA structure is further supported by the fact that the RNA thermosensor sequence is 100% identical in all human pathogenic Yersinia species, although the homology between Y. pseudotuberculosis and Y. enterocolitica is less than 70% and several nucleotide substitutions are detectable in the adjacent yscW and lcrF genes (Figure S7). In fact, a PBAD::lcrF-lacZ reporter in Y. enterocolitica 8081 exhibited a similar thermo-dependent expression pattern, indicating that the RNA thermometer is also functional in this Yersinia species (R. Steinmann, K. Böhme, unpublished results).
The intergenic position of the Yersinia RNA thermosensor is unique. All previously known RNA thermometers are positioned at the 5′-end of heat shock or virulence transcripts [54]. Also sequence and structure of the Yersinia thermometer deviate significantly from the thermosensor controlling virulence genes of L. monocytogenes. The listerial RNA thermometer is positioned within the untranslated region (5′-UTR) of the prfA mRNA and forms one extended stemloop structure (130 nt) in which the ribosome binding site and the start codon locate in two small and unpaired bulges within the long hairpin structure. This overall structure prevents translation at moderate temperature but is destabilized at 37°C through additional melting of the loops facilitating the access of ribosomes [57].
The hairpin II of the lcrF 5′-UTR bears highest resemblance with fourU elements (UUUU pairing with AGGA) predicted in the 5′-UTR of the heat shock genes groES and dnaJ of Staphylococcus aureus and Brucella melitensis, and agsA of Salmonella enterica serovar Typhimurium [43]. Only the agsA leader sequence has been studied in detail. It is short (58 nt), simply structured and folds into two hairpins. Hairpin II with the fourU region and an unpaired internal G-A loop was temperature-responsive and melted at heat shock temperature while hairpin I remained stable [43]. A stabilizing G-C pair in close vicinity to the fourU motif and Mg2+ ions are required to set the melting temperature to heat shock conditions [59], [60]. Hairpin II of the Yersinia thermometer is devoid of the stabilizing G-C pair, and contains a large number of weak A–U and G–U base pairs interrupted by two asymmetric internal loops (Figure 5A). These features might contribute to setting the melting temperature to a more moderate temperature provided by the mammalian host.
The Listeria prfA and the fourU elements are clearly distinct from the complex structured thermometer embedded in the coding region of the rpoH gene of E. coli [61] and the widespread class of ROSE-type thermometers, a conserved regulatory element found in the 5′-UTR of heat shock genes in many α- and γ-proteobacteria [54]–[55]. ROSE elements range from 60 to 110 nt and form a complex secondary structure of 2–4 hairpins, of which the 5′-located stemloop(s) remain folded whereas the 3′-proximal hairpin including the ribosome binding site is thermo-labile and melts upon heat shock. The ROSE class of thermometers contains a UUGCU/AGGA motif in which the highly conserved 5′-G residue pairs in a syn-anti conformation with the second G in the AGGA stretch of the ribosome binding site followed by non-canonical interactions including a triple UC-U and a U-U pair [62].
The structure and thermo-induced conformational changes have been studied with several prototypic RNA thermometers. However, the physiological relevance, e.g. for heat resistance and for recovery in the post-stress situation, has only been proven for the Syncheocystis hsp17 thermometer [63]. Here we provide first experimental evidence that a functional lcrF RNA thermometer is crucial for Yersinia pathogenesis. A repressing ‘closed’ mutation resulted in a strong reduction of the bacterial burden in the PP, MLN, liver and spleen, and afforded a dramatic survival advantage, similar to lcrF-deficient strains. Presence of the Yersinia LcrF protein is very critical for virulence as it controls production of the well-characterized virulence determinants, the antiphagocytic Yop effectors and their type III secretion machinery. Evidence of the importance of this protective immune defense strategy derives from pYV-cured strains, which rendered the bacteria completely avirulent, and from studies with Yersinia lcrF knock-out strains which were severely attenuated in mouse models of septic, oral and nasal infection [10], [64]–[65].
Another interesting aspect is that Yersinia does not profit from elevated LcrF levels provided by a destabilized RNA structure during infection. Strains carrying the ‘open’ lcrF variant are attenuated or exhibit the virulence potential of the wildtype strain. Likewise, overproduction of Hsp17 by an ‘open’ thermometer in Syncheocystis provided a burden to photosynthetic performance and bacterial fitness [63]. It is very likely that additional control mechanisms prevent Yop production and/or secretion when not needed to maintain maximal bacterial fitness. Although significantly higher levels of the LcrF protein were produced in the ‘open’ UU-28/-27CC mutation at 25°C Yop proteins were not detectable in the supernatants. This is consistent with a previous study demonstrating that LcrF(VirF) overexpression in Y. enterocolitica under the control of the tac promoter did not result in Yop secretion at 25°C [31]. This strongly indicates that low temperature not only prevents the opening of the lcrF RNA thermometer, but also impedes Yop secretion which might be explained by the fact that some T3S genes are only activated by temperature and not by LcrF [31].
Similar to other RNA thermometers, the yscW-lcrF intergenic region accounts only in part for the drastic induction of LcrF after temperature upshift. Apparently, very rapid and efficient activation of LcrF production is achieved by the combination of separate regulatory modules. Thermal induction of LcrF translation mediated by the RNA thermometer is combined with temperature-regulated proteolysis of YmoA repressing yscW-lcrF expression. Regulatory cascades composed of an alternative sigma factor and an RNA thermometer have been reported [55], [63], [66]. They might be able to integrate several independent signals (e.g. heat and unfolded proteins), whereas the two-layered control by a thermo-sensitive regulator and an RNA thermometer discovered in Yersinia presents a novel strategy to strictly adjust virulence gene expression to the presence in the warm-blooded host.
Both thermosensing mechanisms appear to be complemented by yet another control level. The yscW-lcrF RNA is highly unstable in Y. pseudotuberculosis, which is consistent with a previous study reporting that lcrF mRNA degradation was so fast in Y. pestis that the transcript could not reliably be detected [24]. Recently, it has been shown that the translocator pore protein YopD recognizes the 5′-ends of transcripts of all type III secretion genes and facilitates their degradation [67]. A negative YopD effect was also observed for LcrF synthesis (R. Steinmann, unpublished results) which could be explained by YopD-mediated stabilization of hairpin II or YopD-induced degradation of the yscW-lcrF transcript. Translocation of the YopD protein upon host cell contact would then result in stabilization and increased translation of the LcrF virulence activator, adjusting injectisome and Yop production according to effector translocation.
Presence of highly homologous 5′-UTRs and structurally conserved RNA thermometers in the yscW-lcrF intergenic region of all pathogenic Yersinia species suggests that they are highly important to adjust biological fitness and virulence. RNA-based regulation of LcrF synthesis is very rapid, highly efficient and less energy-consumptive as it allows fast induction upon host entry and tissue contact, and permits immediate shut-off when the initiating signals are removed. This is particularly important as failure to perform type III secretion results in avirulence due to rapid clearance by the host. On the other hand, uncontrolled secretion of Yops was found to be highly detrimental and cause a severe growth defect of the bacteria, which would also be disadvantageous for their survival and host persistence [68]–[70]. As a consequence a complex feedback mechanism must be responsible for perfect adaptation. The molecular mechanism coupling transcription and translation with Yop export is currently under investigation.
Animal work was performed in strict accordance with the German regulations of the Society for Laboratory Animal Science (GV-SOLAS) and the European Health Law of the Federation of Laboratory Animal Science Associations (FELASA). The protocol was approved by the Niedersächsisches Landesamt für Verbraucherschutz und Lebensmittelsicherheit: animal licensing committee permission no. 33.9.42502-04-055/09.
E. coli and Yersinia strains were routinely grown under aerobic conditions at 25°C or 37°C in LB (Luria Bertani) broth on solid or in liquid media if not indicated otherwise. The antibiotics used for bacterial selection were as follows: ampicillin 100 µg ml−1, chloramphenicol 30 µg ml−1, tetracyclin 10 µg ml−1, and kanamycin 50 µg ml−1.
All DNA manipulations, polymerase chain reactions, restriction digestions, ligations and transformations were performed using standard genetic and molecular techniques [71]–[72]. Plasmid DNA was purified using the Qiagen Plasmid Mini or Midi Kits. Restriction and DNA-modifying enzymes were obtained from Roche, Fermentas, Promega or New England Biolabs. The oligonucleotides used for amplification by PCR, sequencing and primer extension were purchased from Metabion. PCR reactions were performed routinely in a 100 µl mix for 25 cycles using Taq polymerase or Phusion High-Fidelity DNA polymerase (New England Biolabs) according to the manufacturer's instructions. PCR products were purified with the QIAquick PCR purification kit (Qiagen) before and after digestion of the amplification product. Site-directed mutagenesis to delete or substitute nucleotides in the yscW-lcrF intergenic region of pBO1817 and pBO1818 was performed as described in the instruction manual of the QuikChange mutagenesis kit (Stratagene, LaJolla, USA) with plasmids harboring the yscW-lcrF wildtype sequence as template and the mutagenic primers listed in Table S1. Sequencing reactions were performed by GATC (Konstanz, Germany) or by the in-house facility.
Strains and plasmids used in this study are listed in Table 1 and primers for plasmid generation are listed in Table S1. The ymoA+ fragment of Y. pseudotuberculosis of pAKH71 was generated by PCR using primers 1 and 2, digested with BamHI and SalI, and inserted into pACYC184. To construct pAKH77 a DNA fragment carrying the ymoA gene was amplified by PCR using the primer pair 79/80. The fragment was digested with EcoRI and XhoI and inserted into the corresponding sites of pASK-IBA5plus. A fragment carrying the yscW-lcrF intergenic region and the first 10 nt of the lcrF gene under control of the T7 promoter was amplified with primer pair 3/4. The resulting fragment was cloned into the SmaI site of pUC18 to generate plasmid pBO1817. Plasmid pBO1823 and pBO1855 were derived from pBO1817 by site-specific mutagenesis using the mutagenesis primer pairs 5/6 and 7/8, respectively. For toeprinting analysis pBO1818 was constructed by insertion of a PCR fragment amplified with primer pair 9/10 into the SmaI site of pUC18. Plasmids pBO1824 and pBO1855 were derived from pBO1818 by site-specific mutagenesis using the mutagenesis primer pairs 11/12 and 13/14, respectively. Primer 15 and 16 were used for amplification of the 5′-untranslated region of the gnd gene of E. coli from chromosomal DNA of MC4100 and the resulting fragment was cloned into the NheI/EcoRI sites of pBAD18-lacZ(481) to generate pED05. For construction of plasmids pED06–pED08 and pED12–pED13 harboring different mutations within the yscW-lcrF intergenic region upstream of the lcrF-lacZ fusion on pKB14, a two-step PCR was performed. The first PCR reaction was always performed with primer 17 and mutagenesis primer I, and the second PCR with primer 18 and mutagenesis primer II. The following listed mutagenesis primer I/II were used for plasmid: pED06 (19/20), pED07 (21/22), pED08 (23/24), pED12 (25/26), and pED13 (27/28). The two generated PCR fragments for each plasmid were used as templates, amplified with primer pair 17 and 18, and cloned into the NheI/EcoRI site of pBAD18-lacZ(481). Plasmids pED10 and pED11 contain lcrF-lacZ fusions with different portions of the yscW locus located upstream of the lcrF gene, starting from position +5 and +246 relative to the yscW start codon. For generation of the different fusion fragments, primer combination 25/8, and 26/8 and template plasmid pSF4 was used for PCR, and the resulting fragments were cloned into the NheI/EcoRI sites of pBAD18-lacZ(481). To analyze expression of the yscW gene, a PCR-derived fragment harboring the yscW regulatory region from position −310 to +281 relative to the transcriptional start site of yscW was amplified from chromosomal DNA of Y. pseudotuberculosis strain YPIII with primers 32 and 33 and cloned into the PstI site of pGP20 to generate pKB10. To engineer an in frame deletion of yscW from nucleotide position +378 to +564 in the yscW-lcrF-lacZ fusion, first a two-step PCR was performed. In the first step, two fragments containing the region upstream and downstream of the yscW deletion was amplified from chromosomal DNA of YPIII using primer pair 32/73 and 74/37. Subsequently a third PCR was performed with primers 32 and 37 using the upstream and the downstream fragments as templates. The PCR product was digested with PstI and ligated into the vector pGP20 generating pKB12. To construct pKB14, the yscW-lcrF intergenic region was amplified with primer pair 17 and 18 and cloned into the NheI/EcoRI sites in plasmid pBAD18-lacZ(481). To obtain equivalent plasmids (pKB13 and pKB18) in which hairpin II (Δ-44/-25) or hairpin I (Δ-111/-57) was deleted from the yscW-lcrF intergenic region, primer pairs 17/34 and 18/35 or 17/42 and 18/43 were used to synthesize overlapping fragments which were used for a third amplification reaction with primer pair 17 and 18. The resulting PCR fragments were also inserted into the NheI/EcoRI sites of pBAD18-lacZ(481). Plasmid pKB34 carrying a yscW-lcrF-lacZ fusion starting from position −310 was constructed by insertion of a PCR fragment amplified with primer pair 36/37 into the PstI site of pTS02. Continuous deletions of the 5′-regulatory region were obtained by amplification using different forward primers 38–41 and reverse primer 37. The resulting fragments were ligated into the PstI site of pTS02 to generate the yscW-lcrF-lacZ fusion plasmids pKB39–42. Plasmids pKB84 and pKB85 were constructed by amplification of the yscW upstream region either without (pKB84) or with the yscW 5′-UTR (pKB85) using primer pairs 65/66 or 65/67, respectively. The resulting PCR fragments were cloned into the KpnI site of pFU68. To construct plasmid pKB90 harboring a deletion of the 5′-untranslated region (5′-UTR) of yscW from +13 to +241 relative to the transcriptional start site, a two-step PCR reaction was performed. Two fragments containing the region upstream and downstream of the yscW 5′-UTR were amplified from chromosomal DNA of YPIII using primer pairs 87/89 and 88/90. Subsequently a third PCR was performed with primers 87 and 88 using the upstream and the downstream fragments as templates. The product was digested with PstI introduced into pTS02. To engineer a deletion of yscW from nucleotide position +271 to +651 two fragments containing the region upstream and downstream of the yscW deletion were amplified from chromosomal DNA of YPIII using primer pairs 75/77 and 76/78. Subsequently, a third PCR was performed with primers 77 and 78 using the upstream and the downstream fragments as templates. The product was digested with SpeI and SphI and introduced into pDM4 resulting in pRS29. For construction of plasmids pRS41–46 harboring different mutations within the yscW-lcrF intergenic region upstream of the lcrF gene, a two-step PCR was performed and the resulting fragments were cloned into the SphI/SpeI sites in pDM4. For the construction of pRS41, first two PCR reactions were performed with primers 46/50 and 47/51, and the two PCR fragments were used as templates to amplify the yscW-lcrF (AG-46/-45CC) mutant version with primer pair 44/45. Plasmids pRS42–44 were constructed by insertion of SpeI/SphI fragments amplified with primers 44/45 from different PCR fragments used as templates. For the production of the template fragments the two primer pairs 44/52 and 45/53, 44/54 and 45/55 as well as 44/56 and 45/57 were used for the synthesis of the GUU-30/-28AAA, UU-28/27CC and AUA-36/-34CCC fragments. To engineer plasmid pRS45, (Δloop2 −44/−25) both template fragments were obtained by amplification with primers 48/49 and 64/49. After annealing of the template fragments, the yscW-lcrF fragment harboring the Δloop2 −44/−25 mutation was amplified with primer pair 63/64 and cloned into the SpeI/SphI sites of pDM4. For the construction of pRS46, first two PCR reactions were performed with primers 48/43 and 49/42, and the two PCR fragments were used as templates to amplify the yscW-lcrF (AG-46/-45CC) mutant version with primer pair 48/49. The yadA-lacZtranslational fusion encoded by pSF1 was constructed by insertion of a yadA promoter fragment amplified with the primer pair 91/92 into the PstI site of pGP20. The lcrF-lacZ and yscW-lcrF-lacZ fusion plasmids pSF3 and pSF4 were constructed by insertion of PCR fragments amplified from Y. pseudotuberculosis YPIII genomic DNA with primer pairs 37/58 and 37/32 into the PstI site of pGP20.
Y. pseudotuberculosis strain YP50 was constructed by insertion of a kanamycin cassette into the locus of wildtype strain YPIII using the RED recombinase system as described [73]. First, the kanamycin resistance gene was amplified using the kanymoA primers (Table S2) and plasmid pACYC177 as template. Next, the Yersinia genomic DNA was used as a template to amplify 500-bp regions flanking the target gene. The upstream fragment was amplified with a primer pair of which the reverse primer contained additional 20 nt at the 5′-end which were homologous to the start of the kanamycin resistance gene. The downstream fragment was amplified with a primer pair of which the forward primer contained additional 20 nt at the 3′-end which were homologous to the end of the kanamycin resistance gene (for primer see Table S2). In the next step, a PCR reaction was performed with the forward primer and the reverse primer using the upstream and downstream PCR products of the target gene and the kan gene fragment as templates. The PCR fragment was transformed into Y. pseudotuberculosis YPIII pKOBEG-sacB and chromosomal integration of the fragments was selected by plating on LB supplemented with kanamycin. Mutants cured of pKOBEG-sacB were proven by PCR and DNA sequencing.
For the construction of the lcrF knock-out mutant strain YP66, a lcrF::AmpR PCR fragment was generated using an AmpR-resistance plasmid as a template with primers composed of 55 nucleotides which are homologous to the up- or downstream region of the lcrF gene followed by 20 nucleotides homologous to the 5′- or 3′-end of the ampicillin resistance gene (for primer see Table S2). The resulting PCR fragment was integrated into the lcrF locus of Y. pseudotuberculosis YPIII on pYV by the RED recombinase system (Derbise et al., 2003). Selection of the mutant strain was performed as described [74]. One strain, YP66, harboring the lcrF::AmpR mutation in the lcrF locus, as proven by PCR and DNA sequencing, was used for further studies.
All mutant strains with deletions or nucleotide substitutions in the yscW-lcrF intergenic region (YP82–86, YP90, YP95, YP96, YPIP01 and YPIP02) were constructed by homologous recombination using suicide plasmids pRS41–46 and pRS29. Plasmids were mated from E. coli S17-1 λpir (tra+) into Y. pseudotuberculosis YPIII or IP32953 and transconjugants were selected on Yersinia selective agar (Oxoid) supplemented with chloramphenicol. The recombination of the plasmid into the Yersinia virulence plasmid pYV yielded a merodiploid strain, including a wildtype and the mutant copy of yscW-lcrF. Subsequently, the resulting strain was plated on 10% sucrose and fast growing, large colonies were selected. Because sucrose induces the expression of the sacB gene on the integrated plasmids and leads to the production of a toxic substance that prevents growth, a spontaneous second recombination process resulting in the excision of the integrated plasmid is advantageous. 50 selected fast-growing strains were screened for chloramphenicol sensitivity to prove the loss of the integrated plasmid. One strain (Table 1), harboring the desired yscW-lcrF mutation, as proven by PCR and DNA sequencing, was taken for further analysis.
The E. coli mutant strains were constructed with the RED recombinase system as described previously [75]. E. coli strain KB1 was constructed by introducing a stpA deletion into strain BL21lDE3 and used to generate KB3 (BL21λDE3 stpA− hns−). Subsequently, KB3 was used to construct KB4 (BL21λDE3 stpA− hns− hha−). First, a kanamycin cassette was amplified by PCR with primers homologous to the resistance gene encoded on pKD4 followed by homologous sequences of adjacent regions of the target gene (for primer see Table S2). The PCR fragment was transformed into E. coli BL21 pKD46. Chromosomal integration of the fragment was selected by plating on LB supplemented with kanamycin. Subsequently, mutant derivatives were cured of the temperature-sensitive plasmid pKD46 by cultivation at 37°C. To remove the resistance gene at its FLP recognition sites the mutants were transformed with the helper plasmid pCP20 encoding the FLP recombinase. For thermal induction of FLP synthesis and subsequent removal of the temperature-sensitive plasmid pCP20, mutants were incubated at 37°C.
Overnight cultures were diluted 1/50 in fresh medium and grown to stationary phase (OD600 of 3). 2.5 ml culture were withdrawn, mixed with 0.2 volume of stop solution (5% water-saturated phenol, 95% ethanol) and snap-frozen in liquid nitrogen. After thawing on ice, bacteria were pelleted by centrifugation (2 min, 14.000 rpm, 4°C), and RNA was isolated using the SV total RNA purification kit (Promega) as described by the manufacturer. RNA concentration and quality were determined by measurement of A260 and A280.
Total cellular RNA (20 µg) was separated on MOPS agarose gels (1.2%), transferred by Vacuum Blotting for 1.5 h onto positively charged membranes (Whatman) in 10× SSC using a semi-dry blotting system and UV cross-linked. Prehybridization, hybridization to DIG-labelled probes and membrane washing were conducted using the DIG luminescent Detection kit (Roche) according to the manufacturers instructions. The yscW-lcrF transcript was detected with a DIG-labelled PCR fragment (DIG-PCR nucleotide mix, Roche) with primer pair 59 and 60 (Table S1).
KB4 transformed with pAKH77 or pAKH11 was grown at 37°C in LB broth to an A600 of 0.6. Anhydrotetracycline was added (0.2 µg/ml) to induce the expression of YmoA-Strep-Tag or 2 mM IPTG was used to induce H-NS-His6 expression. For purification of the YmoA-H-NS heterodimer KB4 transformed with pAKH74 and pAKH77 was used for overexpression of the YmoA in the presence of the Yersinia H-NS protein. The cells were grown for an additional 3 h before being harvested. The purification procedure for the Strep-tagged YmoA protein was performed according to the manufacturers instructions (IBA GmbH, Germany). H-NS purification was performed as described [76]. The purity of the YmoA and the H-NS protein was estimated to be >95%.
For DNA-binding studies the purified YmoA and H-NS proteins were dialysed against the DNA-binding buffer (10 mM Tris-HCl pH 7.5; 3 mM DTT; 7,5% glycerol; 100 mM KCl; 100 mM MgCl2). The yscW fragment (−2 to +272) for DNA band shift analysis was obtained by PCR using primer pair 68/69 with chromosomal DNA of Y. pseudotuberculosis YPIII and the csiD1 and csiD2 control fragments were amplified by PCR from chromosomal DNA of E. coli strain MC4100 with primer pairs 71/72 and 70/71 (see Table S1). The DNA fragments and increasing concentrations of purified YmoA or H-NS were incubated for 30 min in DNA-binding buffer at room temperature and immediately loaded on 4% polyacrylamide gels.
The lcrF RNA for structural probing was obtained by run-off transcription with T7 RNA polymerase from plasmids pBO1817, pBO1823 and pBO1855 linearized with MlsI. The RNAs were radioactively 5′-end labelled according to Brantl & Wagner [77]. Partial digestion of the RNAs was performed using the ribonucleases T1 and V (Ambion, USA) as described previously [43]. RNA corresponding to about 30.000 cpm was mixed with 1 µl of 5× TMN buffer (100 mM tris acetate pH7.5, 10 mM MgCl2, 500 mM NaCl), 0.5 µg tRNA (Invitrogen, Germany), and destilled water to a total volume of 4 µl. The samples were incubated at 25°C or 37°C for 5 min before 1 µl of RNAse T1 (0.001 U/µl), RNase V (0.0002 U/µl) or RNAse-free water were added. After 5 min of digestion at the appropriate temperature, the reaction was stopped by addition of 5 µl formamide stop solution. The samples were denatured at 95°C, and separated on a denaturing 8% poyacrylamide/urea gel. The alkaline ladders were generated with 60.000 counts of lcrF mRNA as described previously [77].
RNAs for primer extension inhibition experiments (toeprinting analysis) were synthesized in vitro by runoff transcription with T7 RNA polymerase from linearized plasmids pBO1818, pBO1824, and pBO1833. Toeprinting experiments were performed with 30S ribosomal subunits, lcrF mRNA and tRNAfMet mainly as described previously [43]. The 5′-[P32]-labelled lcrF-specific primer 61 (2 pmol) was used for reverse transcriptase reaction. About 1 pmol of the lcrF mRNA was annealed to the radioactive primer and incubated for 20–30 min at 25°C and 37°C in a 20 µl reaction mix with 16 pmol of uncharged tRNAfMet (Sigma-Aldrich, USA) in the presence or absence of 6 pmol of the 30S ribosomal subunits isolated as described [78]. To initiate the primer extension reaction, 2 µl VD-Mg2+ buffer (0.05 M Tris-HCl pH 7.4, 0.3 M NH4Cl, 30 mM β-mercaptoethanol, 0.05 M MgO acetate) containing 80 U of the MMLV reverse transcriptase (USB, USA), dNTPs, and BSA was added and incubated for 10 min at 25°C. cDNA synthesis was stopped by the addition of 20 µl formamide stop solution. In parallel, sequencing reactions using the same lcrF specific primer was performed with the Thermo Sequenase cycle sequencing Kit (USB, USA) according to the manufacturer's instructions. All samples were denatured at 95°C for 5 min, and separated on a denaturing 8% polyacrylamide/urea gel.
Primer extension analysis was performed to determine the transcriptional start site of the yscW-lcrF mRNA from strain YPIII. At an OD600 of 2.0 (early stationary phase), total RNA was extracted of the samples using the SV total RNA purification kit (Promega) as described by the manufacturer. Annealing was performed with 20 µg extracted RNA and the 5′-Dig-labelled oligonucleotide 62 for yscW in 20 µl of 1× First Strand Buffer (Invitrogen) by slow cooling of the sample (0.01°C/sec). 8 mM dNTPs and 5× FS Buffer (Invitrogen) with 200 U Superscript II reverse transcriptase (Invitrogen) was added to the reaction mix and incubated for 1 h at 42°C. The size of the Dig-labelled reaction products was determined on a denaturing 6% DNA sequencing gel by a detection procedure as described [79].
The activity of the β-galactosidase activity of the lacZ and the phoA fusion constructs was measured in permeabilized cells as described previously [71], [80]. The activities were calculated as follows: β-galactosidase activity OD420 · 6,75 · OD600−1 ·Δt (min)−1 · Vol (ml)−1; alkaline phosphatase activity OD420 · 6,46 · OD578−1 · Δt (min)−1 · Vol (ml)−1. Reporter fusions emitting bioluminescence were measured in non-permeabilized cells with a Varioskan Flash (Thermo Scientific) using the SkanIt software (Thermo Scientific) for 1 s per time point. The data are given as relative light units (RLU/OD600) from three independent cultures performed in duplicate. The level of statistical significance for differences in reporter gene expression was determined by the Student's t test.
For immunological detection of the LcrF and YadA proteins, Y. pseudotuberculosis cultures were grown under specific environmental conditions as described. Cell extracts of equal amounts of the bacteria were prepared and separated on a 15% (LcrF) or 10% (YadA) SDS-PAGE [72]. Subsequently the samples were transferred onto an Immobilon-P membrane (Millipore) and probed with polyclonal antibodies directed against YadA or LcrF (generous gift of Greg Plano) as described [74]. The cell extracts used for Western blotting were also separated by SDS-PAGE and stained with Coomassie blue to ensure that the protein concentrations in the different cell extracts are comparable; about 10 mg protein was applied of each sample.
The Yop secretion assay was performed as described previously [69]. Bacteria were grown overnight at 25°C in LB medium, diluted 1∶50 in fresh LB medium and grown at 25°C until the culture reached an OD600 of about 0.4–0.5. Subsequently, the cultures were shifted to 37°C for 3–4 h in the absence or presence of 20 mM Mg2+ and 20 mM Naoxalate, a Ca2+ chelator. Proteins in the medium supernatant were harvested, filtered and precipitated with TCA. Precipitated proteins were resuspended in equal amounts of sample buffer, separated on 15% SDS polyacrylamide gels and visualized by Coomassie brilliant blue staining.
In order to assess the impact of the RNA thermometer on Y. pseudotuberculosis virulence, groups of 7-week-old female BALB/c mice were orally infected with 5·108 bacteria using a ball-tipped feeding needle. To prepare the inocula, the Y. pseudotuberculosis wildtype strain YPIII and isogenic mutant strains were cultured overnight in LB at 25°C. The bacteria were harvested by centrifugation, washed and resuspended to the appropriate concentration in PBS. Three days after infection of the mice the colony forming units (CFU) per gram tissue were determined in the Peyer's patches, mesenterial lymph nodes, liver and spleen. Isolated Peyer's patches were rinsed with sterile PBS and incubated with 100 µg/ml gentamicin in PBS in order to kill the bacteria on the luminal surface. After 30 min, gentamicin was removed by extensive washing with PBS for three times. Subsequently, all organs were weighted, homogenized in PBS, and plated in three independent serial dilutions on Yersinia selective agar (Oxoid, Germany). For the survival assays, groups (n = 10) of 7-week-old BALB/c mice were orally infected with a lethal dose of 2·109 bacteria (Y. pseudotuberculosis strains YPIII and mutant strains cultured overnight in LB at 25°C). The infected mice were monitored for 14 days every day to determine the survival rate.
|
10.1371/journal.pgen.1002683 | Genetic Inhibition of Solute-Linked Carrier 39 Family Transporter 1 Ameliorates Aβ Pathology in a Drosophila Model of Alzheimer's Disease | The aggregation or oligomerization of amyloid-β (Aβ) peptide is thought to be the primary causative event in the pathogenesis of Alzheimer's disease (AD). Considerable in vitro evidence indicates that the aggregation/oligomerization of Aβ is promoted in the presence of Zn; however, the functional role of Zn in AD pathogenesis is still not well clarified in vivo. Zn is imported into the brain mainly through the solute-linked carrier (Slc) 39 family transporters. Using a genetically tractable Drosophila model, we found that the expression of dZip1, the orthologue of human Slc39 family transporter hZip1 in Drosophila, was altered in the brains of Aβ42-expressing flies, and Zn homeostasis could be modulated by forcible dZip1 expression changes. An array of phenotypes associated with Aβ expression could be modified by altering dZip1 expression. Importantly, Aβ42 fibril deposits as well as its SDS-soluble form were dramatically reduced upon dZip1 inhibition, resulting in less neurodegeneration, significantly improved cognitive performance, and prolonged lifespan of the Aβ42-transgenic flies. These findings suggest that zinc contributes significantly to the Aβ pathology, and manipulation of zinc transporters in AD brains may provide a novel therapeutic strategy.
| Alzheimer's disease (AD) is characterized by extracellular amyloid plaques and altered metal ion (including Zn, Cu, Fe) concentrations in the brain. Amyloid plaques are the result of increased aggregation of Aβ, while the in vivo role of metal ions such as Zn remains poorly understood. We found that the expression of a zinc transporter (dZip1) is altered in the brains of AD flies. Genetic manipulation of dZip1 to modulate its expression was accompanied by altered Aβ accumulation, resulting in changes in the neurodegeneration development, cognitive performance, and lifespan of the AD flies. These genetic findings support the zinc role in AD pathology and implicate a new therapeutic target for treating AD.
| Alzheimer's disease (AD) is a major neurodegenerative disease affecting the elderly. The accumulation of amyloid-β (Aβ) peptides, which either form the major component of senile plaques (SP) or the oligomer state in patient brains, is hypothesized to be the primary causative event in AD pathogenesis [1]–[3]. However, what drives the Aβ accumulation and how this accumulation links to progression of the disease is not well understood.
Increasing evidence indicates that the disruption of metal homeostasis, particularly in Zn and Cu concentrations, is strongly correlated with the pathophysiological process of AD [4]–[6]. Although copper and, to a lesser extent, iron can induce partial Aβ aggregation, they need a mildly acidic condition (pH 6.6) [7]. Zn2+ is the only metal ion available to aggregate Aβ at pH7.4-the normal physiological pH [8]–[10]. Elevated Zn was found and co-purified with Aβ from AD brain tissues, associated with markedly high Zn level in cerebral spinal fluid (SP) [11]. Measurements from well characterized late stage AD (LAD) also showed a significant increase of Zn in brain sections of hippocampus, multiple neocortical areas and amygdala compared to age-matched normal control subjects [4], [5], [12]–[14]. Although several reports indicate that Zn induces Aβ aggregation at low physiological concentrations [8], [15], [16], later studies showed that higher Zn concentrations are required for significant fibril formation [17], [18]. These pieces of evidence were obtained mostly from in vitro experiments. Therefore how Zn status influences Aβ pathology in vivo throughout life course remains unclear. Dietary intervention of zinc intake with zinc chelators in animals show some encouraging results [19], [20], [21]. However, genetic evidence is still lacking. More importantly, zinc chelators are usually not zinc specific, and may associate with other nonspecific phenotypes [22], precluding accurate mechanism analysis.
Transport of Zn into cells is mediated by a set of zinc transporters called Zrt-Irt like proteins (Zips). Zips are characterized as influx transporters that mediate Zn2+ uptake into cytoplasm from extracellular or vesicular sources [23], [24], and are encoded by the solute-linked carrier (Slc) gene family, Slc39 [23]. At least 14 Zips have been identified in the human genome [23] and 8 in Drosophila [25]. Most Zips are predicted to have eight transmembrane domains (TM) with a histidine-rich loop between TM3 and TM4, and to be located at the plasma membrane [26]. Although the uptake of Zn from the brain's extracellular environment to intracellular compartments in neurons and glia cells is not completely understood, the Zips are thought to be involved in this process [23], [24]. To our knowledge, the relationship between Zips and AD has not been explored to date.
Previously we and others have shown that expression of human Aβ42 in Drosophila brains recapitulates the main symptoms of AD including age-dependent memory loss, formation of amyloid deposits and neurodegeneration [27], [28]. In the current study, we found that the time course expression change of dZip1, an orthologue of human Slc39 family transporter hZIP1 in Drosophila, was reversed in brains of Aβ42 flies as compared with normal control flies. We hypothesize that modulating dZip1 expression level might affect Zn accumulation in the brain and modify the AD pathological process. By creating dZIP overexpression and RNAi transgenic flies, we demonstrated that dZip1 is critically involved in Aβ-induced AD pathological process, and by lowering dZip1 expression Aβ toxicity can be markedly ameliorated.
Using the amino acid sequence of hZip1, BLAST searches revealed 8 putative Zips in Drosophila, among which CG9428-encoded putative protein shared the highest similarity with human Zip1 (29% identity). We designate it as dZip1. Topology analysis of dZip1 revealed the presence of eight putative transmembrane domains, a histidine–rich loop between domains 3 and 4 which was predicted to occur within the cytosol, and extracellular N- and C-terminals (Figure 1A). All these features are typical of Slc39 family members [23], [24].
We measured dZip1 transcript levels in organs of w1118 flies by semi-quantitative RT-PCR (sqRT-PCR). dZip1 is expressed in the gut and other organs including the brain (Figure 1B left). To further explore the physiological role of dZip1 in flies, we created dZip1 over-expression (OE) and RNAi transgenic flies. We confirmed elevated and decreased dZip1 transcript levels accordingly by dZip1 over-expression or RNAi in whole body of transgenic flies driven by actin-Gal4 (Figure 1B Right). We then tested the zinc sensitivity of these flies. Figure 1C shows that dZip1 OE flies were more sensitive to zinc overdose (t-test, p<0.001), while dZip1-RNAi flies were more tolerant to zinc overdose (t-test, p<0.001) in comparison with controls. These results suggest that dZip1 is indeed involved in zinc uptake.
To test whether brain Zn levels could be changed accordingly by specifically modulating brain dZip1 level, the pan-neuronal elav-Gal4 driver was used to drive dZip1 OE and RNAi in fly brains. Figure 1D showed qRT-PCR results of dZip1 transcript level in different transgenic fly brains. Two dZip1-RNAi transgenic lines were used, in which the knockdown effect of dZip1-RNAi #2 transgenic line (∼1/10 of the dZip1 transcript level in control elav-Gal4 flies) was much stronger than dZip1-RNAi #1 transgenic line (∼3/5 of the dZip1 transcript level in control elav-Gal4 flies) (Figure 1D). dZip1 OE transgenic line showed ∼6–7 fold increase of dZip1 transcript level compared with control brains (Figure 1D). Inductively coupled plasma optical emission spectrometry (ICP-OES) result showed that over-expression of dZip1 in the fly brain markedly increased brain Zn accumulation (t-test, P<0.01), while knocking down dZip1 via RNAi decreased brain Zn level compared with the control elav-Gal4 flies (t-test, p<0.05) (Figure 1E). Therefore, specific manipulation of dZip1 expression level in fly brains could affect the brain Zn status.
Aβ42 expression in fly brains could induce an age-dependent formation of amyloid deposits and neurodegeneration which may correlate with disturbed metal homeostasis, especially for zinc and copper. We therefore checked native dZip1 mRNA levels in brains of elav-Gal4>UAS-Aβ42 (Aβ42) flies and normal elav-Gal4 flies at different ages. We found that the dZip1 mRNA level was developmentally altered in Aβ42 fly brains compared to elav-Gal4 fly brains (Figure 2A). The dZip1 mRNA level in brains of w1118 flies showed similar results as elav-Gal4 flies (data not shown), indicating that this is not due to the effect of the introduced Gal4 gene. These results imply that Aβ expression may indeed lead to a zinc dyshomeostasis in the fly brains. Of note is that the endogenous dZip1 expression was lower in young adult Aβ flies as compared to the control, although the brain zinc level of the Aβ flies at the young stage was not significantly different (Figure 2B), suggesting a complex regulation of zinc metabolism (involving participants such as other zinc importers and exporters besides dZip1) is involved in the brain zinc control.
Using ICP-OES we directly measured the Zn content in brains of 7- and 30-day old flies. By 7 days of age, differences of Zn content among different groups were still not apparent (Figure 2B). With ageing, the Zn content in all the brains significantly increased; however, in comparison with 7-day old elav-Gal4 flies without Aβ expression, 30-day old Aβ42 flies increased ∼170% of their Zn level, significantly higher than that of 30-day old elav-Gal4 flies (average ∼65% increase of the control Zn). Over-expression of dZip1 further increased Zn accumulation in Aβ42 fly brains, while dZip1 knockdown slowed brain Zn accumulation during aging and significantly reduced Zn accumulation compared to 30-day old Aβ42 flies (t-test, p<0.01). By testing Zn content of these flies at 20-day old (Figure S1), although it's not obvious as that of 30-day, the trend is already apparent and statistically significant. These results indicate an intimate connection among Aβ42 expression, aging and brain zinc accumulation, and the latter can be strongly affected by dZip1 expression interference.
Next, we used Hematoxylin and Eosin (H&E) staining to examine whether the extent of brain neurodegeneration in aged Aβ42 flies (visualized as vacuolization in the brain region, arrowhead in Figure 3) could be changed by modulation of dZip1 expression. Compared to the control, Aβ42-expressing brains with dZip1 knockdown were to a large extent normal (Figure 3C), but when dZip1 was overexpressed degenerative changes were significantly more apparent in both the cortex and the neuropil region, where there were more and bigger bubbles (Figure 3B). Counting the number of vacuoles in the cortex and neuropil revealed that over-expression of dZip1 increased brain vacuolization more than 2-fold (Figure 3D, t-test, p<0.001), whereas dZip1 knockdown dramatically decreased brain vacuolization in Aβ42 flies (t-test, p<0.001). Meanwhile, no significant different of neurodegenerative bubbles were found between dZip1 OE alone and age-matched control elav-Gal4 flies in 20-day old fly brains (Figure 3E and 3F). To exclude possible off-target effect of the RNAi action, we confirmed the results with a differently constructed RNAi line, V3986, from the VDRC stock center. V3986 exhibited a similar level of dZip1 reduction as our own dZip1-RNAi line #2 (Figure S2), and could roughly to the same extent rescue Aβ-associated brain vacuolization (Figure S2C and S2D). These results indicated that modulating dZip1 expression could change the brain neurodegenerative process.
It has been shown that Aβ42 flies start to display locomotor dysfunction after three weeks of age and their lifespan is significantly reduced [27], [28]. We therefore tried to examine whether dZip1 expression levels could affect Aβ42 flies' locomotion and lifespan. Two dZip1-RNAi transgenic lines were used, in which the knocking down effect of the dZip1-RNAi #2 transgenic line was much more obvious than the dZip1-RNAi #1 transgenic line (Figure 1D). Assay of climbing ability demonstrated that Aβ42 flies with dZip1 overexpression started to have a locomotor defect at 15-day old age comparable with that of the control elav-Gal4>UAS-Aβ42 flies at between 20–25 days (Figure 4A). In contrast, Aβ42 flies with decreased dZip1 levels through RNAi (#1) had a delayed climbing deficit (Figure 4A). This rescuing effect on climbing ability was even more pronounced when the stronger line dZip1-RNAi #2 was used. As a further control, we tested the climbing ability of the transgenic flies without Aβ42 expression (Figure 4B). Only flies with over-expression of dZip1 manifested mild locomotor defect 30 days after eclosion as compared with the age-matched control elav-Gal4 flies. Again we confirmed the climbing rescuing effect with a different RNAi line V3986 and found it compared similarly with our dZip1-RNAi flies (Figure S2E). We conclude the locomotion defect as a result of Aβ42 toxicity can be modulated through the change of dZip1 expression level.
Consistent with the result obtained in the climbing assay, the lifespan of Aβ42 flies was shortened by over-expression of dZip1 and prolonged by RNAi-based knockdown of dZip1 expression (Figure 5A and 5B). The dZip1-RNAi #2 transgenic line exhibited the strongest rescue, with 33.3% and 88.8% increase respectively in the median lifespan of Aβ42 flies reared at 25°C and 29°C (Figure 5D). Similar to that in the climbing assay, the lifespans of flies without Aβ42 expression were largely indistinguishable except for that of the dZip1 OE (elav-Gal4>UAS-dZip1) flies, which displayed a noticeable 7.2% reduction (Figure 5C and 5D). Our results indicate that a reduction of dZip1 expression in Aβ42 flies leads to improved locomotor ability and longer lifespan. Consistently, zinc chelation with clioquinol extended Aβ42 survival (Figure S3). Interestingly, clioquinol appeared rescuing male Aβ42 more effectively than females.
A cardinal defect in Alzheimer's disease is memory loss. With extensively characterized Pavlovian olfactory aversive conditioning [29], a memory defect in adult Aβ42 flies started to appear as early as 5-day-old. dZip1 knockdown significantly rescued memory loss at this stage (Figure 6A). Paradoxically, overexpression of dZip1 also resulted in obvious memory recovery. As a control, we examined how alteration of dZip1 expression alone (in the absence of Aβ42) might impact memory scores. Overexpressing or knocking down dZip1 did not significantly influence memory of 5-day-old normal flies (Figure 6B and 6C), although overexpression of dZip1 might have a marginal beneficial effect. These results suggest that modulating dZip1 could alleviate the Aβ42 toxicity on memory ability at early stages.
The aforementioned experiments demonstrated dZip1 expression modulation can markedly alter the course of Aβ-associated neurodegeneration. Towards further analysis of the mechanism underlining dZip1 effect on Aβ toxicity, we first tried to determine where Zn was concentrated in these Aβ fly brains. We raised 10-day old Aβ42 flies on normal food supplemented with ZnCl2 and then used Zinquin staining to detect Zn distribution in vivo. The purpose of applying extra Zn is to enhance the fluorescent signal. Without Zinquin treatment, little signal was detected (Figure 7B). Over-expression of dZip1 (Figure 7E and 7F) appeared to produce stronger signals in the neocortex and neuropile region of the fly brain compared with control Aβ42 (Figure 7C and 7D) flies. Quantitative analysis of these signal intensities showed a significant difference (Figure 7K, t-test, p<0.01). Only a faint signal was detected in dZip1-RNAi fly brains (Figure 7G, 7H and 7K, t-test, p<0.05 at 40 h and t-test, p<0.001 at 72 h). Staining of the brains of elav-Gal4 flies without Aβ expression showed faint signals similar to dZip1-RNAi flies (Figure 7I and 7J). These results demonstrate a positive correlation between dZip1 level and Zn accumulation in the neocortex and neuropile region of the fly brain where Aβ deposits were revealed (Figure 8).
To determine if Zinc is causally related to Aβ42 deposition, we overexpressed or knocked down dZip1 under the control of elav-Gal4 in Aβ42 flies and subjected the brains to histochemical analysis. Aβ42 peptides could form diffused amyloid deposits in fly brains [27], [28]. Thioflavin-S (TS) staining of the whole brain was used to specifically visualize the Aβ42 fibril deposits (Figure 8) [23]. Aβ42 deposits were observed in the Kenyon cell body region of Aβ42-expressing brains. The number of TS-positive deposits in Aβ42 flies was significantly increased with aging (Figure 8A and 8A1, 8B and 8B1). Over-expression of dZip1 markedly increased the number of Aβ42 deposits compared with age-matched Aβ42 flies (Figure 8A and A1, ∼212% at 25-day old relative to Aβ42 flies at 25-day old, t-test, p<0.001). Conversely, RNAi-based knockdown of dZip1 in Aβ42 flies significantly decreased the Aβ42 deposits compared with age-matched Aβ42 flies (Figure 8B and B1, ∼48% at 30-day old relative to Aβ42 flies at 25-day old, t-test, p<0.001). Taken together, our results demonstrate that dZip1 over-expression can increase Aβ42 accumulation whereas inhibiting dZip1 can decrease Aβ42 deposition.
The above TS staining was used to specifically detect the Aβ42 fibril deposits. Recently, an alternative model for the Aβ toxicity is put forth hypothesizing that amyloid oligomers rather than plaques are responsible for the disease [30]. The oligomer form, together with the monomeric form of Aβ, is soluble in SDS whereas the fibril aggregate is not. We next investigated how SDS-soluble Aβ was affected by modulating dZip1 expression. Fly brain lysates were used for a Western blotting analysis. The result showed that the SDS-soluble Aβ42 (low level aggregation forms) was dramatically decreased in accordance to reduced dZip1 level and increased when dZip1 was over-expressed (Figure 8C and 8C1). In a separate experiment, SDS-insoluble but formic acid-soluble Aβ42 (high level aggregation forms) was also examined; much decreased formic acid-soluble Aβ42 level was observed when dZip1 expression was inhibited (Figure S4), consistent with the TS-staining result.
Whole-mount immunohistochemical staining with Aβ42 antibody also revealed a reduction of Aβ42 level in the dZip1-RNAi fly brains (Figure 9). Abundant amyloid deposits were observed in the Kenyon cell body region of 20-day old Aβ42-expressing brains (Figure 9A, arrow). Such deposits were markedly decreased when dZip1 was knocked down (Figure 9C) and greatly increased when dZip1 was over-expressed (Figure 9B). Similar results were found in 30-day old fly brains (Figure 9E and 9F) compared with age-matched Aβ42-expressing control brains (Figure 9D). In the case of co-overexpression of dZip1 and Aβ42, vacuolization became much more pronounced with aging (Figure 9E). As a control for the Aβ42 antibody, we used it against elav-Gal4 fly brains and found no meaningful signal, indicating the antibody reacts specifically with Aβ (Figure S5).
Because dZip1 RNAi led to a general reduction of Aβ42 level, we tried to ask whether this was due to Aβ42 expression inhibition or an increase of Aβ42 degradation. Aβ42 gene was directly under the control of elav-Gal4, and indeed no changes of RNA expression under the various genetic manipulations were observed (Figure 10A and 10B, Figure S2B). We then suspected that zinc might reduce the rate of Aβ42 clearance. Several proteases (NEP1-3, IDE) were proposed to act in the Aβ degradation [31], [32] and we thus explored whether they were affected by dZip1 expression in the Aβ flies. Aβ expression brought some changes to the expression of these genes, but introduction of dZip1-RNAi transgene in the Aβ flies resulted no significant expression increase of these genes (Figure 10C-10F). We did however, observed some decrease of NEP2 expression in dZip1 OE/Aβ flies. Therefore we saw little evidence that the observed rescuing effect of dZip1 RNAi on Aβ42 flies is mediated by an increase of these degrading proteases.
Together, we conclude dZip1 reduction decreases levels of Aβ42, in both the high (fibril aggregates) and low aggregation forms.
Previous studies suggest that heavy metals, especially Zn, have a close relationship with the development of Alzheimer's disease [4]–[6], [8]–[10]. However, little genetic evidence exists that demonstrates a functional link between Aβ and proteins involved in zinc assimilation. In this study, we showed that manipulating the Slc39 family protein dZip1 greatly altered the Aβ toxicity. In particular, knocking down dZip1 in brains of Aβ42 flies markedly decreased both Aβ42 deposits and zinc accumulation. dZip1 knockdown ameliorated early memory loss, decreased the number of neurodegenerative vacuoles, significantly enhanced locomotor ability and prolonged life-span in Aβ42-expressing flies. Taken together, our results provide strong evidence to support our hypothesis that knocking down protein dZip1 may mitigate Aβ pathology and Aβ-dependent behavioral defects in a Drosophila model of Alzheimer's disease.
The accumulation and aggregation of Aβ42 peptide in the neocortex has been suggested to be caused by its abnormal interactions with neocortical metal ions especially Zn, which is constitutively found at high levels in the neocortical regions where they play important roles in normal physiology [11], [33], [34]. Our qRT-PCR results showed that the dZip1 transcript level was higher in brains of young control elav-Gal4 and w1118 flies and decreased with age, but lower in brains of young Aβ42 flies and increased with age. Paradoxically, dZip1 transcript level was lower in young adult Aβ42 flies than age matched normal flies. Since at this stage Zn level is not any lower (Figure 2B) in Aβ42 flies, we suspect other Zn homeostasis genes might also be affected. Indeed, besides dZip1 quite a few other Zip or ZnT (Slc30 family transporter) genes are also expressed in the brain (data not shown and the Flyatalas: http://flyatlas.org), likely contributing to Zn uptake or export. Supporting this notion is a recent report showing that the expression level of the Slc30 family protein ZnT3 decreased with age in AD brains [35], [36]. Thus the general Zn status is the result from the combination effect of all these Zn homeostasis genes. It is possible that Aβ42 expression alters the Zn homeostasis starting from an early stage, although we are not totally clear why dZip1 is reduced at early stages but increased at late stages. One thing worthy of consideration is that total zinc level does not reflect well available cellular zinc. In other words, two brains with the same level of total zinc may have very different levels of zinc for use. dZip1 expression regulation may reflect the cell's native response to its own physiological states-to bring more or less zinc into cells. Because Aβ42 can likely bind to zinc, and monomer and oligomers may have different binding characteristics, we speculate that Aβ42 can affect zinc homeostasis even in the absence of noticeable total zinc level alteration. This dyshomeostasis could result or be reflected by native dZip1 expression changes.
Not all pathogenic effects of Aβ in Drosophila correlate directly with its aggregation. An artificial mutation (L17P) with decreased Aβ42 aggregation tendency is associated with lower toxicities, in term of locomotor ability and lifespan, but induces even earlier onset of memory defects than its normal counterpart [37]. Furthermore, although both Aβ40 and Aβ42 affect learning, only Aβ42 causes degeneration [27]; inhibition of PI3K activity ameliorated the Aβ42-induced early memory loss, but did not rescue neurodegeneration [38]. These results lead to the speculation that neuronal dysfunction and neurodegeneration may be mediated by different mechanisms. In our study, while knocking down dZip1 and overexpressing dZip1 were associated with opposite effects on all other aspects of Aβ42-induced toxicity, it is interesting that both knocking down dZip1 or overexpressing dZip1 lessened Aβ42-induced early memory loss. While inhibiting Zn accumulation in fly brains could promote memory in Aβ42-expressing flies by reducing aggregation of Aβ42, the amelioration of memory loss with dZip1-overexpression is likely due to a different mechanism. At the moment, we are not certain how this memory gain with dZip1 over-expression was achieved.
One might ask whether the dramatic effect seen with dZip1 modulation in Aβ42 flies could be reproduced with dietary zinc supplement or chelation. High levels of zinc supplement (such as 2.5 mM) do significantly worsen the viability of Aβ42 flies, however at this level normal flies are also affected (Figure S3C and S3D). Using a zinc chelator clioquiniol, dietary feeding can rescue to some extent Abeta flies, interestingly mostly in male flies (Figure S3A and S3B). Thus genetic intervention is a much more effective method. We interpret this as systemic Zn overloading may cause damages to other tissues before enabling a dramatic Zn increase in the target organ. Likewise, zinc depletion at the organismal level is generally harmful: zinc is an essential nutrient vital to many biological processes. Indeed, high levels of clioquinol greatly impact even the survival of normal flies. Thus, we believe low zinc level can affect fly development and survival so that overall beneficial effect of zinc reduction by dietary measures is significantly less effective than targeted neuronal zinc reduction through genetic interventions.
dZip1 repression results in less Aβ42 level. Although other possibility cannot be excluded, we favor the model that zinc induces oligomerization of Aβ42, as supported by numerous in vitro evidences. More oligomers result more fibril deposits. Perhaps the oligomer and the aggregated Aβ42 are more stable than the monomer, so that the Aβ42 level is dramatically reduced in dZip1 RNAi flies, where a larger fraction of Aβ42 adopt the monomeric form more susceptible to clearance (Figure 11).
In summary, we have demonstrated the modulating effect of dZip1 on Aβ42 toxicity in a Drosophila model of Alzheimer's disease. We observed Aβ42 expression could cause a change of dZip1 expression pattern during ageing. Through genetic manipulation of dZip1 expression, we can modify the pathological process of Aβ42. These results raise the possibility that Zip1, or more broadly Zn transporter genes expressed in the brain, could be a new kind of promising therapeutic target in AD pathology.
Flies were raised and maintained at 25°C or otherwise indicated temperatures. All general stocks were obtained from the Bloomington Drosophila Stock Center, which includes Actin-Gal4, elav-Gal4. UAS-Aβ42 transgenic strain was reported previously [37]. To make UAS-dZip1 transgenic fly, corresponding genomic DNA including a 76 bp intron was cloned into the pUAST vector. The primers used for PCR amplification were: UAS-dZip1-F: 5′-CCGAATTCAAGATGAGCGCTACCGC-3′ and UAS-dZip1-R: 5′- GGAAGATCTCTA GGAACAGGTTAGGCTG-3′. The UAS-dZip11-RNAi constructs were generated according to Lee and Carthew [39]. The primers used were: WIZ-dZip1-RNAi-F: 5′-GGGTCTAGAATGAGCGCTACCGC-3′ and WIZ-dZip1-RNAi-R: 5′-GGTCTAGACC ACACAGTGCTCACAG-3′. All transgenic flies were generated in w1118 background following standard protocols.
Total RNA was extracted from the brain, gut and carcase (whole body minus brain and gut) of 10 adults for each sample using TRIzol Reagent (Invitrogen) according to the manufacturers' instructions and subjected to DNA digestion using DNAse I (Ambion) immediately. The concentration and quality of DNAse-treated total RNA were then tested, and 800 ng total RNA from each sample was used to synthesize cDNA by using Superscript™ II Reverse Transcriptase kit (Invitrogen) with oligo(dT) primers. Semi-quantitative RT-PCR (sqRT-PCR) was performed using primers for rp49 (forward: 5′- TACAGGCCCAAGATCGTGAA-3′; reverse: 5′- TCTCCTTGCGCTTCTTGGA-3′) and dZip1 (forward: 5′-ATTATCCTCGCCCTTTCGC-3′; reverse: 5′-TCACCCTCCGCT TCGTCAG-3′). rp49 was used as the loading control. For quantitative RT-PCR (qRT-PCR), 20 fly brains were used for each sample, RNA extraction and cDNA synthesis were the same as described for sqRT-PCR. Primers for amplifying dZip1, NEP1, NEP2, NEP3 and Aβ42 were listed as Table S1. Real-time PCR reactions were monitored on an iCycler (Bio-Rad) by means of SYBR Green (Bio-Rad) dye. mRNA expression levels were determined relative to rp49 expression by relative quantification. Statistical analysis was performed using the Student's t-tests.
For immunostaining analysis on paraffin sections, antigen retrieval was achieved by boiling the samples in 10 mM sodium citrate (pH 6.6) for 15 min. Immunostaining was performed using an avidin-biotin-peroxidase complex (ABC) kit (Vector Laboratories). For Aβ staining, the primary antibody used was anti-Aβ42 (Promega; 1∶500). Appropriate secondary antibodies were diluted 1∶200, and histochemical detection was done with DAB (Sigma-Aldrich) color development.
Adult fly heads were fixed in Carnoy solution (ethanol∶chloroform∶acetic acid = 6∶3∶1) overnight at 4 C°, and then embedded in the paraffin and sectioned at 6 m thickness. H&E staining was performed following standard protocols. Neurodegeneration was assessed by quantification of vacuoles with diameter greater than 3 µm in the fly brains. At least five fly brains were analyzed for each genotype.
Thioflavin-S (TS, Sigma) staining was performed to detect fibril Aβ42 deposits. Fly brains were fixed in 4% paraformaldehyde and permeabilized by 2% triton. Brains were then transferred to 0.25% TS in 50% ethanol overnight. After 1× wash for 10 min in 50% ethanol and 3× wash with PBS, they were mounted with focusclear (Pacgen Biopharmaceuticals Inc.) and covered by cover slips. Slides were inspected with a Zeiss LSM 510 confocal microscope, aided by LSM 510 analysis software. TS-positive deposits located in the mushroom body somatic region were counted for comparison analysis.
The training and testing procedures were as previously described [40], [41]. During one training session, a group of 100 flies was sequentially exposed for 60 s to two odors, octanol (OCT) or methylcyclohexanol (MCH), with 45 s of fresh air in between. Flies were subjected to foot-shock (1.5 s pulses with 3.5 s intervals, 60 V) during exposure to the first odor (CS+) but not to the second (CS−). To measure “immediate memory (also referred to as “learning”)”, flies were transferred immediately after training to the choice point of a T-maze and forced to choose between the two odors for 2 min, at which time they were trapped in their respective T-maze arms, anesthetized, and counted. A performance index (PI) was calculated from the distribution of flies in the T-maze. A reciprocal group of flies was trained and tested by using OCT as the CS+ and MCH as the CS+, respectively. PIs from these two groups finally were averaged for an n = 1 and multiplied by 100. A PI of 0 represented a 50∶50 distribution, whereas a PI of 100 represented 100% avoidance of the shock-paired odor.
SDS-soluble and SDS-insoluble but formic acid-soluble Aβ42 were prepared as previously reported [27]. Lysates from equal number of fly heads were diluted in SDS sample buffer and separated by 10–20% Tris-Tricine gels (Invitrogen), and transferred to nitrocellulose membranes (Invitrogen). Membranes were boiled in PBS for 3 min. Membranes were blocked with 3% BSA and blotted with primary antibody. Primary antibodies used in this study were mouse anti-Aβ42 (6E10, Covance Research Products) and rabbit anti-Actin (Sigma). After washing in TBST for 3 times, membranes were incubated with secondary antibodies for 1 hr at RT. After 3 times wash in TBST, membranes were incubated with ECL working solution (GE healthcare) and developed with films (Kodak). Data were analyzed with ImageJ sofeware (NIH).
10-day old adult flies reared on normal condition were transferred to vials with normal food supplied with 4 mM ZnCl2. Fly brains were then dissected at 40 and 72 h after transferring, respectively. The dissected brains were then incubated with 25 µM Zinquin (Sigma) for 30 min at 37°C, and washed 3 times with 1×PBS buffer for 5 min each time. After that, brains were examined using conventional epifluorescence microscope (Nikon, Diaphot 300) equipped with a Nikon 100×, 1.4 NA Plan Apo oil-immersion objective. Zinquin signals in the neocortex and neuropile region of fly brains were quantitated by using ImagJ software.
For the metal stress assay, Drosophila was fed on normal medium containing 2.5 mM ZnSO4. Control flies were fed on normal medium in the absence of drug. Mortality was recorded every 24 h or a longer intervals. Each vial contained 20–25 flies, and the experiments were repeated at least three times.
For the metal content analysis, flies were reared on normal food and fly heads were collected at day 7, 20 and 30 after eclosion. Fly heads were dissolved in 1 ml 65% HNO3, boiled in 100°C water bath for 10 min and diluted to 10 ml for metal content analysis with inductively coupled plasma optical emission spectrometry (ICP-OES, IRIS Intrepid II XSP, Thermo Electron Corporation, USA).
The climbing assay was referenced to Iijima et al. (2004) [27]. Briefly, twenty flies were placed in a plastic vial and gently tapped to the bottom. The number of flies at the top of the vial was counted after 18 s of climbing under red light (Kodak, GBX-2, Safelight Filter). The data shown represent results from a cohort of flies with four repeats tested serially for 5–50 days. The experiment was repeated more than three times.
Flies of two days after eclosion were used for the experiment. Twenty to 23 flies were placed in a food vial. Each vial was kept at 25 or 29°C, 70% humidity, under a 12-h light–dark cycle. Food vials were changed every 2–3 days, and dead flies were counted at that time. At least 150 flies were prepared for each genotype, and the experiments were carried out more than three times. Percent increases in life span are based on comparing the median survivals. Prism (GraphPad) was used for statistical analysis of lifespan data. Mantel-Cox log-rank statistical analysis was used for testing statistical significance of the differences between the survivorship curves.
All data were analyzed by Student's t-test. Statistical results were presented as means ± SEM. Asterisks indicate critical levels of significance (*P<0.05, **P<0.01 and ***P<0.001).
|
10.1371/journal.pbio.1001937 | In Vitro Generation of Neuromesodermal Progenitors Reveals Distinct Roles for Wnt Signalling in the Specification of Spinal Cord and Paraxial Mesoderm Identity | Cells of the spinal cord and somites arise from shared, dual-fated precursors, located towards the posterior of the elongating embryo. Here we show that these neuromesodermal progenitors (NMPs) can readily be generated in vitro from mouse and human pluripotent stem cells by activating Wnt and Fgf signalling, timed to emulate in vivo development. Similar to NMPs in vivo, these cells co-express the neural factor Sox2 and the mesodermal factor Brachyury and differentiate into neural and paraxial mesoderm in vitro and in vivo. The neural cells produced by NMPs have spinal cord but not anterior neural identity and can differentiate into spinal cord motor neurons. This is consistent with the shared origin of spinal cord and somites and the distinct ontogeny of the anterior and posterior nervous system. Systematic analysis of the transcriptome during differentiation identifies the molecular correlates of each of the cell identities and the routes by which they are obtained. Moreover, we take advantage of the system to provide evidence that Brachyury represses neural differentiation and that signals from mesoderm are not necessary to induce the posterior identity of spinal cord cells. This indicates that the mesoderm inducing and posteriorising functions of Wnt signalling represent two molecularly separate activities. Together the data illustrate how reverse engineering normal developmental mechanisms allows the differentiation of specific cell types in vitro and the analysis of previous difficult to access aspects of embryo development.
| Stem cells are providing insight into embryo development and offering new approaches to clinical and therapeutic research. In part this progress arises from “directed differentiation” – artificially controlling the types of cells produced from stem cells. Here we describe the directed differentiation of mouse and human pluripotent stem cells into cells of the spinal cord and paraxial mesoderm (the tissue that generates muscle and bone that is normally found adjacent to the spinal cord). During embryo development, spinal cord and paraxial mesoderm arise from a shared group of precursors known as neuromesodermal progenitors (NMPs). We show that signals to which NMPs are exposed in embryos can be used to generate NMPs from pluripotent stem cells in a dish. We define conditions for the conversion of these NMPs into either spinal cord or mesoderm cells. Using these conditions, we provide evidence that the decision between spinal cord and mesoderm involves a gene, Brachyury, that promotes mesoderm production by inhibiting spinal cord generation. Together the data illustrate how mimicking normal embryonic development allows the generation of specific cell types from stem cells and that this can be used to analyse cells that are otherwise difficult to study.
| The differentiation of embryonic stem cells (ESCs) to specific cell types offers insight into developmental mechanisms and has potential therapeutic applications. For example the differentiation of neural progenitors (NPCs) from monolayers of ESCs seeded in serum free conditions is a model of neural induction and regional patterning [1]. In the absence of additional signals, NPCs differentiated from ESCs adopt an anterior-dorsal neural (telencephalon) identity [1],[2]. The addition of Sonic Hedgehog (Shh) ventralises these neural progenitors, mimicking the in vivo role of Shh [3],[4]. Exposing NPCs to retinoic acid (RA) results in the repression of anterior identity and the induction of genes that typify hindbrain and anterior spinal cord (cervical) identity [5]. This has been taken as support for the idea that newly generated NPCs are by default anterior and are then posteriorised by exposure to specific extrinsic signals [6],[7]. It is notable, however, that RA is actively excluded in the progenitors of the posterior spinal cord after gastrulation [8] and that commonly used ESC differentiation protocols do not efficiently generate neural cells of the more posterior spinal cord such as thoracic and lumbar spinal cord cells marked by posterior Hox gene expression, including Hoxc8–10 expression [9].
The anterior and posterior nervous system has distinct origins [10]–[12]. Anterior epiblast expresses Otx2 and contributes cells to the anterior nervous system [2],[13] whereas spinal cord progenitors are located posteriorly [14]–[16]. Clonal analysis indicates that the spinal cord shares a common lineage, at least in part, with the trunk paraxial mesoderm that forms the somites [15]. The dual-fated neuromesodermal precursors (NMPs) of these tissues are located in the node-streak border (NSB), caudal lateral epiblast (CLE) cell layer adjacent to the regressing node and the chordoneural hinge of the tail bud [13],[14],[17],[18]. Cells in these regions coexpress the neural marker Sox2 and nascent mesoderm marker Brachyury [8],[19],[20]. Genetic lineage tracing experiments confirm that many spinal cord cells previously expressed Brachyury [21] indicating that as cells from regions harbouring NMPs move into the neural tube they downregulate Brachyury but maintain Sox2 expression and consolidate neural identity. By contrast, NMPs that enter the primitive streak delaminate basally, downregulate Sox2 and acquire expression of the paraxial mesoderm marker Tbx6 [22] en route to somite formation. Strikingly, in embryos lacking Tbx6, paraxial mesoderm cells express Sox2 and transdifferentiate into neural cells, providing additional support for the inter-relationship between spinal cord and somitic mesoderm [22]–[24]. As yet, however, the existence of NMPs has only been revealed in vivo and the inaccessibility of this population makes them difficult to study.
The region occupied by NMPs is exposed to Wnt and Fgf ligands [16]. These signals are required for body axis elongation [16] and both Wnt and Fgf signalling have been implicated in mesoderm and neural induction [22],[25]–[31]. In vivo and in vitro evidence has suggested that Wnt signalling is responsible for posteriorising tissue by inducing posterior Hox genes [29],[32],[33]. Together, the data suggest that the generation of posterior neural tissue and paraxial mesoderm proceeds by Wnt and Fgf signalling inducing a neuromesodermal bipotential intermediate. To test this idea, we developed an efficient in vitro differentiation method for spinal cord and paraxial mesoderm from mouse and human pluripotent stem cells. We show that carefully timed and calibrated pulses of Wnt and Fgf signalling generate a population of cells that transiently coexpress Sox2 and Brachyury in which the expression of posterior Hox genes are induced. Transcriptome analysis is consistent with the equivalence of these cells to the NMPs found in vivo. In vivo grafting and directed in vitro differentiation confirm the ability of NMPs to assume spinal cord or paraxial mesoderm cell fates. We further show that Brachyury is not required for the production of posterior neural cells or for the induction of posterior Hox genes, hence separating the posteriorising and mesoderm inducing functions of Wnt signalling. Taken together the data define a means to generate posterior neural and paraxial mesodermal tissues in vitro and illustrate how the directed differentiation of stem cells provides novel insight into developmental mechanism.
To identify conditions for the generation of posterior neural cells from monolayers of mouse ES cells (mESCs), we cultured mESCs in serum free media containing bFgf for 3 days (D1–D3) and then transferred these to media lacking bFgf for an additional 2 days [1] (Figure 1A). This resulted in the induction of a post-implantation epiblast-like intermediate by D2, indicated by the downregulation of the “naïve” pluripotency marker Zfp42 (Rex1) and the upregulation of the epiblast marker Fgf5 (Figure 1F) [34]. At this stage, Pou5f1, which is expressed in both mESCs and epiblast-like cells, is maintained (Figure 1F) [34]. In all experiments a Shh agonist, SAG, was added at D3 in order to generate a predictable ventralised identity for subsequent comparisons. The transcriptome of cells was then analysed at D5 by mRNA-seq. Consistent with previous studies [35]–[39], cells in these conditions had acquired an anterior neural identity (NA), exemplified by the expression of Otx1 and Otx2 [2]. The presence of SAG induced the expression of ventral neural markers (Figure S1A). Addition of retinoic acid (RA) and SAG to differentiating mESCs at D3 downregulated anterior neural markers (e.g. Otx2, Six3, Lhx5) and instead genes typical of hindbrain identity, including Hoxa2, Hoxb2, Mafb, Epha4 and Ephb2 were expressed (Figure 1B) [40]. However markers of spinal regions of the neural tube, such as the 5′ Hox genes Hoxc6, Hoxc8 and Hoxc9 were not detected (Figure 1B) [5],[9]. Changing the timing or concentration of RA used in these experiments did not result in the efficient induction of more posterior spinal cord identity [29].
To recapitulate the sequence of signalling events that generate the spinal cord, we seeded mESCs into serum free media containing bFgf. At D2, Wnt signalling was induced by the addition of the Wnt agonist CHIR99021 (CHIR). bFgf and Wnt agonist were removed at D3 and cells exposed to media containing RA and SAG until D5. Examination of gene expression profiles indicated that cells subjected to the FGF/CHIR/RA regime expressed genes characteristic of the spinal cord including high levels of 5′ Hox genes Hoxb6, Hoxb8, Hoxc6, Hoxc8, Hoxc9 and low levels of the anterior neural and brainstem markers Otx2 and Mafb (Figure 1B). Together, the data suggested that a brief pulse of Wnt signalling between D2–D3 was sufficient to posteriorise differentiating mESCs. We termed the neural cells generated in this regime NP cells and cells that display anterior and brainstem identity NA and NH, respectively (Figure 1A).
We confirmed the posteriorisation and neural identity of Np cells using qRT-PCR and immunostaining (Figure S1C–D). Analysis of the time course of Hox gene expression in NH and NP cells indicated that their temporal sequence of induction matched the in vivo time course [40]: Hoxb1 was induced within 12 h of exposure to Wnt signalling, whereas more 5′ Hox genes were induced later (Figure 1C). Notably the more posterior Hox genes, e.g, Hoxc6 and Hoxc9 were not induced in NH cells. In Np cells Hoxc6, Hoxc8 and Hoxc9 were strongly induced at D4 (Figure 1C). Delaying the addition of CHIR to differentiating mESCs until D3 resulted in a concomitant shift in the timing of Hox gene induction (Figure S2A–C). Furthermore, in agreement with studies indicating that RA represses the most posterior Hox genes [41], exposure of cells to FGF/CHIR without subsequent addition of RA induced Hoxc10 characteristic of the lumbar spinal cord (Figure S2E). Finally we passaged NH and NP cells at D5 and allowed them to differentiate until D8, at which point we assayed the expression of genes expressed in motor neurons (MNs). Both NH and NP cells adopted a neuronal morphology and expressed the neuronal marker class III β-tubulin (Tuj1). The NH cells acquired a posterior hindbrain MN identity evident by the coexpression of Hoxb4 and the cranial motor neuron marker Phox2b [42] (Figure 1D). In the case of NP cells however, only a few Hoxb4 expressing cells were detected (Figure S1D) and most of the β-tubulin expressing neurons acquired a Hoxc6 and Hoxc9 identity characteristic of neurons of the brachial and thoracic spinal cord, respectively [43] (Figure 1E,G). Moreover NP cells expressed Olig2, a marker of somatic motor neuron progenitors, as well as the differentiated MN markers Hlxb9 and Islet1/2 [3] (Figure 1E). Taken together these data indicate that similar to the situation in vivo [44] and in embryoid bodies [29] exposure of monolayers of differentiating ESCs to a combination of Wnt, Fgf and RA signalling generates spinal cord cells.
To address how the combination of Wnt and Fgf signalling induces spinal cord identity we examined gene expression in differentiating ESCs at D2.5 and D3 (Figure 2A). ESCs that had been exposed to Fgf/Wnt signalling for 12 h (D2.5) and 24 h (D3) induced the expression of Cdx2 and the mesoderm transcription factors Brachyury and Tbx6 (Figure 2B). Recombinant Wnt3a protein had a similar activity to CHIR in these assays (Figure S2F). By contrast, ESCs cultured in the absence of Wnt agonist, expressed significantly lower levels of these genes (Figure 2B). These data suggest that Wnt signalling, in combination with Fgf, is initiating a mesodermal transcriptional program. This is consistent with the loss of mesoderm in mouse embryos lacking Wnt3a [45] and the induction of Brachyury by β-catenin [28].
It was also noticeable that the levels of Sox2 mRNA were transiently reduced in D2.5 and D3 NP cells treated with FGF/CHIR (Figure 2B). We therefore assayed Sox2 and Brachyury proteins by immunostaining in D3 NP and NA cells. Strikingly, the level of Sox2 protein was similar in NP and NA cells, consistent with the long half-life of Sox2 protein [46]. Moreover, ∼80% of NP cells coexpressed Brachyury and Sox2 (Figure 2C) whereas only a small number of NA cells expressed Brachyury. These data suggest that the exposure to bFgf and Wnt signalling induces a cell identity reminiscent of the dual-fated neuromesodermal progenitors present during axial elongation in the CLE [15],[44] (Figure S4E).
If D2–D3 NP cells represent NMPs, they should form mesoderm. To test this, we transferred cells at D3 into media containing Wnt agonist but lacking bFgf. In these conditions (termed Meso) the expression of Sox1, Sox2 and Brachyury were downregulated and several genes characteristic of paraxial mesoderm, including Tbx6 and Msgn1 [47], were significantly upregulated (Figure 2D). Immunostaining revealed that more than 90% of cells in this condition expressed Tbx6 protein at D5 (Figure 2E). By D8 Desmin, the intermediate filament protein of muscle sarcomeres [48] and the muscle transcription factor MyoD were highly expressed (Figure 2E). Thus the continued exposure of cells to Wnt signalling induces a paraxial mesodermal identity that differentiates to a muscle-like identity. This provides evidence that ESCs exposed to Wnt and bFgf at D2–D3 represent bipotential neuromesodermal cells that can differentiate into either mesoderm or neural tissue.
We next tested the in vivo potential of NMP cells. For these experiments we took advantage of the chick. Cells with NMP-like behaviour have been identified in chick [16] and chick embryos provide an accessible and experimentally tractable vertebrate host for grafts of mouse ESCs [49]. We grafted small groups of DiI labelled D3 NA cells, not exposed to Wnt signalling, or D3 NMPs, exposed to Fgf/Wnt signalling for 24 h, into the caudal lateral epiblast of Hamburger-Hamilton (HH) stage 8–9 chick embryos (Figure 2F). Analysis of embryos 24 h later revealed efficient incorporation and migration of the NMP cells to both the neural tube and the somites (Figure 2G). Transplanted cells from a single graft contributed to multiple anterior-posterior levels and most embryos showed contribution to both spinal cord and somites (Figure 2G–J). In several embryos grafted cells were also observed in the tail bud of the embryo as well as the neural tube and somites (Figure 2J). Contribution to endoderm was not observed. By contrast, transplanted NA cells showed somewhat lower rates of engraftment and contributed only to the neural tube and not to somites (Figure 2K–L). These data confirm the bipotency of the in vitro derived NMP cells and demonstrate that similar to in vivo NMPs [15] [44] they contribute to both neural and paraxial mesoderm lineage. Single cell and clonal analysis, in vivo and in vitro, will be necessary to test the potency of individual cells and to understand the molecular mechanism by which neural and/or mesodermal progeny are generated from NMP.
We took advantage of the in vitro differentiation to analyse the transcriptional programmes that generate each of the neural and mesodermal lineages (Figure 3A). Principal component analysis of the transcriptomes indicated that each differentiation pathway could be clearly distinguished (Figure 3B). Strikingly, the first principal component (PC) appeared to represent developmental time and the second PC the tissue identity of the differentiated cells. The data revealed a set of genes that distinguished NP, NA, NH cells and Meso cells (Figure S3 and Tables S2, S3). These included the upregulation of Mafb and Phox2b in NH samples and the upregulation of posterior Hox genes, notably Hoxc6, Hoxc8 and Hoxc9 in NP samples. By contrast, the induction of genes such as Tbx6, Hes7 and Hoxc8 and Hoxc9 in D5 cells subjected to mesodermal conditions confirmed the posterior paraxial identity of these cells. Moreover, the analysis indicated a bifurcation in the transcriptional programmes that generate anterior neural and brainstem cells from those that produce posterior neural and paraxial mesodermal cells. It was notable that gene expression typical of paraxial mesoderm was evident at D4 of NP differentiation suggesting a gradual separation of neural and mesodermal identity. Together these data provide a molecular correlate to the distinct cellular origins of anterior and posterior neural tissue [15] and identifies the NMP state as the branch point in the developmental trajectories.
We identified genes upregulated in NMPs compared to mESCs at D1 and NA cells at D3. Comparing these to genes induced in D5 neural and mesodermal cells revealed a large intersection. Thus, in part, NMPs have a transcriptional programme that is a combination of neural and mesodermal gene expression. In addition however, a set of ∼240 genes appeared uniquely upregulated in NMP cells (Table S1). These included the transcription factors Brachyury, Nkx1.2 (also known as Sax1), which is expressed in the stem zone of midgestation embryos [50],[51] Mixl1 [52], Wnt3a and Cdx2 which are expressed in the primitive streak and nascent mesoderm [32],[33]. In addition Follistatin, which plays a key role in neural induction by blocking TGFβ signalling [53] and components of the Fgf signalling pathway, which is implicated in mesoderm induction [16], are upregulated in NMPs (Figure 3C). Together these data support the idea that exposure of differentiating ESCs to Fgf/Wnt signalling between D2 and D3 induces a bipotential neuromesodermal population equivalent to that found in vivo in the CLE [15],[16],[18],[22] and that the balance and timing of these two signals influences the further differentiation of these cells into neural or mesodermal tissues.
The activation of Wnt signalling in differentiating mouse epiblast stem cells (EpiSCs) leads to a modest induction of Brachyury/Sox2 coexpressing cells, suggestive of NMP identity [19]. To improve the efficiency of this induction we adapted our mESC protocol to take account of the more advanced developmental state of EpiSC compared to mESCs (Figure 1F). Accordingly, we exposed EpiSCs to a range of CHIR (Wnt) and bFgf concentrations and assayed the expression of Sox2 and Brachyury (Figure S4A). Maximal proportions of Sox2/Brachyury coexpressing cells resulted from 3 µM CHIR and 20 ng/ml bFgf (hereafter referred to as FGF/CHIR) (Figure 4A, Figure S4A).
Assaying a broader panel of genes supported the idea that FGF/CHIR was inducing NMP identity. The expression of the pluripotency factor Nanog was undetectable and the majority of the Sox2 expressing cells expressed minimal levels of Oct4, suggesting that they had exited pluripotency (Figure 4B, Figure S4D). Moreover, the acquisition of Brachyury/Sox2 coexpression coincided with an upregulation of Wnt3a, Cdx2 and Nkx1.2 as well as trunk Hox genes (Figure 4B), characteristic of embryo and mESC derived NMPs. Consistent with this, the paraxial/somitic mesoderm markers Tbx6 and Meox1 and the neural factor Sox1 were expressed in these conditions (Figure S5A). Immunostaining indicated that by D3 of differentiation Tbx6 and Sox2 expression were mutually exclusive (Figure S5B). By contrast the expression of genes characteristic of anterior neural plate (e.g. Otx2 and Six3) and endoderm (Foxa2) [54] were largely absent in FGF/CHIR conditions (Figure S5A). Collectively, these data indicate that, similar to mESCs, stimulation of Wnt and Fgf signalling in mouse EpiSCs leads to the induction of an NMP state.
The developmental potential of differentiated mouse EpiSCs has previously been tested by transplantation into mouse embryos [19],[55]. We therefore grafted EpiSC-derived NMPs constitutively expressing GFP into the NSB of E8.5 embryos. After 48 h in culture, we observed extensive incorporation of GFP expressing cells (15/15 embryos) (Figure 4C–D). Sections from these embryos revealed integration of transplanted cells into the somites and presomitic mesoderm of host embryos (10/10) and neural tube (4/10) (Figure 4D). We did not observe contributions to endoderm or other tissues. Antibody staining for paraxial mesoderm (Tbx6), somite/dermomyotome (Pax3), neural (Sox2) and floor plate (Foxa2) markers confirmed that the engrafted cells had acquired the marker expression of their host environment (Figure 4E). Moreover, examination of the rostral limit of labelling using the somite level as a reference revealed that grafted EpiSC derived NMPs behaved similarly to homotopic grafts of microdissected E8.5 NSB cells [56]. Strikingly, few cells grafted into the node of E7.5 embryos showed any incorporation (2 out of 8 embryos had 8–10 incorporated cells/embryo), suggesting that these conditions produce a population incompatible with gastrulation-stage development. Similarly, cells differentiated for 24 h in FGF/CHIR did not incorporate into the NSB of E8.5 embryos (n = 5) (Figure 4D). Collectively, these results suggest that 48 h treatment of EpiSCs with FGF/CHIR results in coexpression of Brachyury/Sox2 (up to 90%, Figure S4B) and generates NMPs that functionally resemble their in vivo counterparts.
The resemblance of mouse EpiSCs to human embryonic stem cells (hESCs) prompted us to ask whether an analogous FGF/CHIR treatment regimen was sufficient to generate human NMPs. Treatment of three independent hESC lines with CHIR and bFgf from D0–D3 downregulated NANOG and OCT4 and upregulated the suite of NMP expressed genes—BRACHYURY, NKX1.2 and CDX2—similar to mouse ESCs and EpiSCs (Figure 5C, Figure S4D). SOX2 expression was maintained in this population and up to ∼80% of cells co-expressed SOX2 and BRACHYURY (Figure 5B, Figure S4B). We also observed the spontaneous upregulation of paraxial mesoderm/somite markers (TBX6, MSGN1, MEOX1). (Figure S5C). By contrast, the expression of a lateral plate (KDR) and an endoderm (FOXA2) marker were minimal (Figure S5C). Thus FGF/CHIR treated hESCs appear to adopt an NMP identity and are likely to represent the in vitro correlates of the SOX2 and BRACHYURY co-expressing cells found in the caudal epiblast of human embryos [8]. Consequently we dubbed these cells hNMPs.
To test the potency of hNMPs, we treated hESCs with FGF/CHIR for 72 h to drive the generation of BRACHYURY+/SOX2+ cells and then re-plated them for a further 48 h in serum free media without additional factors to promote the induction of spinal cord identity (Figure 5A). We termed these cells NP and compared them to neural cells derived from hESCs using a dual SMAD inhibition protocol involving Nodal and BMP inhibitors (SB/LDN) [57]. Both conditions induced neural identity, exemplified by increased levels of SOX2, TUBB3 and PAX6 (Figure 5D). As expected, neural cells generated using dual SMAD inhibition expressed the anterior marker OTX2 but lacked expression of HOX genes (Figure 5D). By contrast, neural cells derived from NMPs expressed SOX1 and the posterior HOX genes HOXC6, HOXC8 and HOXC9 but not OTX2 (Figure 5D). A similar expression profile was obtained after treatment with RA and dual Shh agonists SAG and purmorphamine (Pur). This also induced expression of the motor neuron progenitor marker OLIG2 (Figure S5D). Antibody staining verified HOXC8 expression in NP conditions and revealed that the majority of HOXC8+ cells co-expressed SOX2, confirming their neural identity (Figure 5E, F). Treatment of neural cells for 48 h with FGF/CHIR following 72 h dual SMAD inhibition did not result in HOXC8 induction suggesting that posteriorisation is necessary before or concomitant with neural induction (Figure S5E). Together these data suggest that neural differentiation of hNMPs generates spinal cord progenitors similar to mNMPs.
We next tested whether hNMPs differentiate into mesoderm by culturing them in the presence of CHIR alone. This resulted in the expression of paraxial/somitic mesoderm markers TBX6, MSGN1 and MEOX1 (Figure S5D), but little if any expression of KDR, a lateral plate mesoderm marker (Figure S5D). Taken together these findings provide evidence of a human NMP population that gives rise to spinal cord and paraxial mesoderm derivatives but not anterior neurectoderm or lateral plate mesoderm. Moreover, a similar set of developmental cues induces and directs NMPs in human and mouse, consistent with a similar ontogeny of trunk tissues in these species.
The ability to generate NMPs in vitro allows experimental investigations of trunk development that are challenging or impossible in vivo. For example, although the requirement for Brachyury in mesoderm formation is well-established [58]–[60], the truncation of embryos lacking Brachyury has complicated analysis of its role in the elaboration of spinal cord identity. In zebrafish, a non-autonomous role for Brachyury orthologues has been identified [60]. It is unclear whether in mammals Brachyury is required directly to maintain NMPs and therefore generate spinal tissue or indirectly via Wnt induction to establish a mesodermal niche that signals to generate or maintain posterior neural tissue. To address this we took advantage of Brachyury null mESCs (BTBR10) derived from embryos lacking Brachyury [61].
Assaying Brachyury null cells at D3 of differentiation indicated that, in contrast to wild-type ESCs, Tbx6 expression was not upregulated by exposure to FGF/CHIR signalling, whereas Cdx2 and Hoxb1 expression were induced (Figure 6B). This is consistent with the lack of posterior mesoderm induction in Brachyury mutant embryos and prompted us to address the fate of Brachyury mutant cells that would normally form mesoderm. In wild type cells exposed to Meso conditions, Tbx6 was highly expressed at D5 (Figure 6D), as were Desmin and MyoD at D8 (Figure 6E). By contrast Brachyury null cells subjected to the same conditions failed to differentiate into paraxial mesoderm as indicated by the absence of Tbx6 (Figure 6D). Instead these cells expressed Sox1, Sox2 and posterior Hox genes (Hoxc6 and Hoxc9) at D5 (Figure 6C) and differentiated into β-Tubulin expressing neurons (Figure 6E). These data indicate that within mouse NMPs, Brachyury not only specifies mesodermal identity via mechanism(s) in addition to its direct stimulation of Wnt signalling, but also represses neural identity. In the absence of Brachyury, NMPs adopt a neural differentiation route. Thus the induction of posterior neural tissue is not dependent on Brachyury. Moreover the data separate the mesoderm inducing and posteriorising activity of Wnt signalling and provide evidence that posteriorisation of the CNS is not dependent on mesoderm derived signals.
What could be responsible for the induction of posterior Hox genes? Analysis of the transcriptome data revealed the induction in NMPs of the Cdx genes Cdx1, 2 and 4, which have been implicated in the regulation of Hox gene expression (Figure 6F) [29],[62]. Induction of both Cdx1 and Cdx2 were detectable within 12 h of FGF/CHIR exposure and the levels of all three genes increased further at D3 and D4 of NP differentiation and at D5 of Meso differentiation. Moreover, the induction of Cdx2 by Fgf/Wnt signalling was maintained in Brachyury null ESCs (Figure 6B). Thus the induction of Cdx proteins by Fgf/Wnt signalling represents a good candidate for the posteriorisation of NMPs. Moreover the temporal accumulation of Cdx levels following Wnt exposure might provide a timing mechanism for the progressive induction of increasingly more posterior Hox genes.
We describe the in vitro generation of bipotential neuromesodermal progenitors from both mouse and human pluripotent stem cells that are capable of producing posterior neural tissue and paraxial mesodermal tissue. This recapitulates the behaviour of NMPs residing in the CLE and NSB [15],[22] (Figure 6G). Moreover, we provide evidence that Wnt signalling has two distinct functions in NMPs, initiating a mesodermal differentiation programme by regulating Brachyury expression and independently posteriorising these cells. It is also likely that Brachyury maintains NMPs during axis elongation by forming a positive feedback loop with Wnt gene expression as has been previously shown [60]. Strikingly, a neuromesodermal precursor is also present in ascidian embryos [63]. Similar to vertebrates, the induction of these cells depends on the timing of Wnt and Fgf signalling [64],[65]. Moreover the mesoderm and posterior nervous system of many arthropods, including short germband insects, arises from a shared progenitor population that is exposed to Wingless signalling and expresses Cdx [66]. Thus molecular and cellular features of the development of the neural and mesodermal components of the trunk appear to be evolutionarily conserved across bilaterian embryos. This emphasizes the distinct developmental origins of cells that form anterior and posterior regions of bilaterian embryos, suggesting an explanation as to why it has proved difficult to generate spinal cells and skeletal muscle from ESCs. More generally, the ability to produce and manipulate NMPs in vitro has the potential to increase the efficiency with which cell types derived from posterior neural and paraxial mesodermal tissue can be generated and analysed.
Animal experiments were performed under the UK Home Office project licenses PPL80/2528 and PPL60/4435, approved by the Animal Welfare and Ethical Review Panel of the MRC-National Institute for Medical Research and MRC Centre for Regenerative Medicine and within the conditions of the Animals (Scientific Procedures) Act 1986. Human Embryonic Stem Cell UK Steering Committee approval has been obtained (ref. SCSC14-09).
The mouse ES cell lines, HM1 [67] and BTBR10 [68] were maintained in ES cell medium [69] with 1000 U/ml LIF (Chemicon) on mitotically inactive primary mouse embryo fibroblasts. To initiate differentiation, ES cells were removed from feeders by dissociation using 0.05% trypsin and then plated onto tissue culture plates for two short successive periods (20–30 mins) to remove feeder layers. To induce differentiation, the cells were plated on CellBINDSurface dishes (Corning) precoated with 0.1% gelatin (Sigma) at a density of 5×103 cells cm−2 in ‘N2B27’ medium. This medium comprised Advanced Dulbecco's Modified Eagle Medium F12 (Gibco) and Neurobasal medium (Gibco) (1∶1), supplemented with 1×N2 (Gibco), 1×B27 (Gibco), 2 mM L-glutamine (Gibco), 40 µg/ml BSA (Sigma), 0.1 mM 2-mercaptoethanol. Cells were grown in N2B27 supplemented with 10 ng/ml bFgf (R&D) for 3 days (D1–D3) and then were transferred into serum free media without bFgf (D3–D5). To induce ventral hindbrain identity NPCs (NH) 100 nM RA (Sigma) and 500 nM SAG (Calbiochem) was added from D3–D5. Spinal cord identity (NP) was induced by the addition of 5 µM CHIR99021 (Axon) or 100 ng/ml Wnt3a (R&D) from D2 to D3 followed by 100 nM RA, 500 nM SAG from D3–D5. To induce mesodermal differentiation the cells were treated with CHIR99021 from D2–D5. To induce terminal differentiation, cells were trypsinised and plated as single-cell suspension on plates coated with Matrigel (BD Biosciences) at a density of 1×105 cells cm−2 in N2B27 medium supplemented with bFgf (10 ng/ml). The next day bFgf was removed and cells were left to differentiate for an additional 3 days.
The mouse EpiSC line R04-GFP [55] was routinely maintained in N2B27 supplemented with Activin A (20 ng/ml; R&D Systems) and bFgf (10 ng/ml; Peprotech) as previously described [70]. For differentiation of EpiSCs into NM progenitors approximately 1500–2000 cells/cm2 were plated on fibronectin (Sigma)-coated wells in N2B27 medium supplemented with CHIR99021 (3 µM; Signal Transduction Division, Dundee) and bFgf (20 ng/ml). For grafting experiments the initial plating density was 2500 cells/cm2 and cells were plated on either fibronectin or gelatin.
Human ESC lines MasterShef 5 and 7 (a gift of Prof. Harry Moore, University of Sheffield) and a Sox2GFP reporter line (a gift of Dr Andrew Smith, University of Edinburgh) were cultured in Essential 8™ medium on Geltrex™-coated plates. For hNMP differentiation cells were pre-treated for 1 h with ROCK inhibitor Y-27632 (10 µM; Calbiochem), dissociated with accutase and plated at approximately 10,000 cells/cm2 (Sox2-GFP hESCs) or 80,000 cells/cm2 (MasterShef5 and 7 hESC lines) on fibronectin-coated wells in N2B27 medium supplemented with 3 µM CHIR99021/20 ng/ml bFgf and Y-27632 (10 µM). The medium was replaced the following day with fresh N2B27 containing the same components minus the ROCK inhibitor. For directed differentiation of hESCs, cultures were differentiated in the presence of CHIR99021/bFgf for 72 h as described above. For neural/spinal cord differentiation 72 h CHIR99021/bFgf-differentiated cells were treated with Accutase (Sigma) and transferred onto Geltrex (Life Technologies)-coated plates either in N2B27 alone or N2B27 supplemented with RA (0.1 µM; Sigma), SAG (0.5 µM; Calbiochem) and purmorphamine (1 µM; Calbiochem) for 48 h. For mesodermal differentiation 72 h CHIR99021/bFgf differentiated cells were cultured in N2B27 supplemented with CHIR99021 (3 µM) for a further 48 h. For dual SMAD inhibition Sox2-GFP hES cells were plated at 10,000 cells/cm2 on Geltrex™-coated wells in N2B27 supplemented with LDN193189 (100 nM; Stemgent) and SB431542 (10 µM; Sigma). This was followed either by re-plating and culture in N2B27 or in N2B27/CHIR99021 (3 µM)/bFgf (20 ng/ml) for a further 48–72 h. All experiments involving hES cells have been approved by the UK Stem Cell Bank steering committee.
To graft NMP and NA cells into the CLE of stage HH8–9 chick embryos, plates of appropriately prepared cells were labelled for 10 mins at 37°C with DiI and washed 3 times with PBS. N2B27 medium was added and cells were incubated for 30 mins. Small clumps of cells were mechanically detached from the plate and transplanted using a manually pulled glass needle. Groups of 100–200 cells were grafted into the caudal lateral epiblast and the eggs were incubated for a further 24 h. Embryos were then fixed in 4% paraformaldehyde (PFA) for 60 mins at 4°C. Fixed embryos were cryoprotected by equilibration in 15% sucrose and then cryosectioned (14 µm). Images were taken using an Apotome2 (Zeiss) and Leica confocal microscope TCS-SP5.
For mouse embryo grafting, r04-GFP EpiSC were flow sorted for GFP expression using a BD FACSAria II sorter and plated overnight in EpiSC conditions followed by Fgf/Wnt for 24 h or 48 h. Mouse embryo grafting (∼10 cells/embryo), culture and imaging were performed as described previously (Huang et al., 2012).
Cells were fixed for 10 minutes at 4°C in 4% paraformaldehyde in phosphate buffer saline (PBS), then washed in PBST (PBS with 0.1% Triton X-100). Blocking was for 1 h in PBST with 3% donkey serum at room temperature. Primary and secondary antibodies were diluted in PBST containing 1% donkey serum. Cells were incubated with primary antibodies overnight at 4°C, with secondary antibodies at room temperature for 2 h, mounted with DAPI containing Prolong Antifade (Molecular Probes), and fluorescent images were taken using an inverted Leica SP5 confocal microscope or an Apotome 2 microscope or an Olympus IX51 inverted microscope (Olympus). Embryo processing and immunohistochemistry on tissue sections was performed as described previously [55]. Whole embryos were imaged using a Nikon NZ100 dissecting microscope, and sections were imaged in an Olympus BX61 fluorescence compound microscope. Nuclear segmentation followed by single cell fluorescence quantification was performed as described previously [70]. The following primary antibodies were used: mouse anti-Hoxc6 (1∶10) (DSHB), mouse anti-Hoxc9 (1∶10) (gift of T. Jessell), mouse anti-Hoxc10 (1∶50) (DSHB), rat anti-Hoxb4 (1∶100) (gift of A. Gould), rabbit anti-Phox2b (1∶200), mouse anti-Tuj1 (1∶1000) (Covance), rabbit anti-Tuj1 (1∶500) (Covance), rabbit anti-Olig2 (1∶500) (Chemicon), mouse anti-Sox2 (1∶200) (ab92494, Abcam), rabbit anti-Sox2 (1∶200) (Millipore), goat anti-Sox2 (1∶100) (R&D), goat anti-Tbx6 (1∶200) (R&D) or rabbit anti-Tbx6 (0.6 µg/ml) (ab38883, Abcam), goat anti-Brachyury (1∶500) (R&D), rabbit anti-RALDH2 (1∶500) (Sigma), rabbit anti-Desmin (1∶500) (Abcam), mouse anti-MyoD1 (1∶200) (DAKO), mouse anti-Islet1 (1∶2000) (gift of T. Jessell), mouse anti-Lim3 (1∶10) (DSHB), mouse anti-HB9 (1∶100) (DSHB), anti-Foxa2 (1 mg/ml) (Santa Cruz; sc-6554), anti-GFP (10 µg/ml) (Abcam; ab13970), anti-Pax3 (1∶20) (DSHB);. anti-Nanog (2.5 µg/ml) (14-5761-80, eBioscience); anti-Oct4 (1 µg/ml) (N-19, Santa Cruz), HoxC8 (5 µg/ml) (Abcam). Secondary antibodies were anti-mouse, anti-rabbit, anti-goat and anti-rat, Alexa's (488, 568, 647) from Molecular Probes.
Total RNA was isolated from cells using the RNeasy kit (Qiagen) according to the manufacturer's instructions and digested with DNase I (Qiagen) to remove genomic DNA. First strand cDNA synthesis was performed with Superscript III system (Invitrogen) using random primers and amplified using Platinum SYBR-Green (Invitrogen). For QPCR the Applied Biosystems 7900HT Fast Real time PCR or the Light Cycler 480 SYBR Green I Master Mix (Roche) systems were used. PCR primers were designed using Primer3 software. All experiments were performed in biological duplicates or triplicates for each time point analysed. Expression values were normalized against the β-actin or the TATA-binding protein (TBP) and standard deviations were calculated and plotted using Prism 6 software (GraphPad). Primer sequences are available upon request.
Total RNA was processed according to the TruSeq protocol (Illumina). Three separate RNA libraries (biological replicates) were barcoded and prepared for each time point. Library size, purity and concentration were determined using Agilent Technologies 2100 Bioanalyzer with a DNA specific chip (Agilent DNA-1000). For sequencing, four samples were loaded per lane of an Illumina Genome Analyzer Hiseq2500. The sequence files generated each contained approximately 30million reads per sample. Reads were aligned to the Ensembl transcriptome mm10 using Bowtie2 and TopHat2 [71]. Per gene counts were collated using HTseq-count [72] and normalized using the DESeq R package [73]. Data analysis, PCA and Biplots were performed using custom scripts in R and MATLAB (MathWorks). RNA-seq data are available in the Array express database (http://www.ebi.ac.uk/arrayexpress) under accession number E-MTAB-2268.
|
10.1371/journal.ppat.1003499 | Mycobacterium tuberculosis Nucleoside Diphosphate Kinase Inactivates Small GTPases Leading to Evasion of Innate Immunity | Defining the mechanisms of Mycobacterium tuberculosis (Mtb) persistence in the host macrophage and identifying mycobacterial factors responsible for it are keys to better understand tuberculosis pathogenesis. The emerging picture from ongoing studies of macrophage deactivation by Mtb suggests that ingested bacilli secrete various virulence determinants that alter phagosome biogenesis, leading to arrest of Mtb vacuole interaction with late endosomes and lysosomes. While most studies focused on Mtb interference with various regulators of the endosomal compartment, little attention was paid to mechanisms by which Mtb neutralizes early macrophage responses such as the NADPH oxidase (NOX2) dependent oxidative burst. Here we applied an antisense strategy to knock down Mtb nucleoside diphosphate kinase (Ndk) and obtained a stable mutant (Mtb Ndk-AS) that displayed attenuated intracellular survival along with reduced persistence in the lungs of infected mice. At the molecular level, pull-down experiments showed that Ndk binds to and inactivates the small GTPase Rac1 in the macrophage. This resulted in the exclusion of the Rac1 binding partner p67phox from phagosomes containing Mtb or Ndk-coated latex beads. Exclusion of p67phox was associated with a defect of both NOX2 assembly and production of reactive oxygen species (ROS) in response to wild type Mtb. In contrast, Mtb Ndk-AS, which lost the capacity to disrupt Rac1-p67phox interaction, induced a strong ROS production. Given the established link between NOX2 activation and apoptosis, the proportion of Annexin V positive cells and levels of intracellular active caspase 3 were significantly higher in cells infected with Mtb Ndk-AS compared to wild type Mtb. Thus, knock down of Ndk converted Mtb into a pro-apoptotic mutant strain that has a phenotype of increased susceptibility to intracellular killing and reduced virulence in vivo. Taken together, our in vitro and in vivo data revealed that Ndk contributes significantly to Mtb virulence via attenuation of NADPH oxidase-mediated host innate immunity.
| Mycobacterium tuberculosis (Mtb) is a very successful intracellular pathogen that infects lung macrophages. Its resistance to intracellular killing has been linked to the development of pulmonary tuberculosis (TB) in humans. Thus, understanding the mechanism by which Mycobacterium tuberculosis (Mtb) persists in the host is a prerequisite for development of efficient strategies to control TB disease. We have previously shown that Mtb nucleoside diphosphate kinase (Ndk) contributes to phagosome maturation arrest via inactivation of Rab5 and Rab7. In this study, we show that Ndk also targets and inactivates the small GTPase Rac1, an essential component of the macrophage NADPH oxidase (NOX2) complex. Ndk-dependent inactivation of Rac1 was associated with reduced NOX2-mediated production of reactive oxygen species (ROS) and ROS-dependent apoptosis. Conversely, disruption of Ndk expression converted Mtb into a mutant strain that induces strong ROS and apoptosis responses. This phenotype was associated with reduced survival of Ndk mutant in vitro and in vivo. Altogether, our findings demonstrate that Ndk contributes significantly to mycobacterial virulence.
| The ability of Mycobacterium tuberculosis (Mtb) to adapt and thrive intracellularly relies on a variety of strategies to alter mechanisms of the host innate immunity. In particular, interference with phagosome biogenesis was highlighted as a significant aspect of Mtb persistence and replication within the macrophage [1], [2]. How Mtb circumvents phagosomal acidity, bactericidal enzymes, and reactive oxygen species (ROS) remains a central question for many cellular microbiologists.
ROS are produced by the phagocyte NADPH oxidase (NOX2) complex and were classified 30 years ago as powerful microbicidal agents against many intracellular pathogens [3]. In vivo evidence for the contribution of NOX2 to the innate immunity arsenal was deduced from field observations of high susceptibility of chronic granulomatous disease patients (CGD) to opportunistic pathogens [4], [5]. Such observations were experimentally confirmed in mouse models of CGD [6], [7]. Recent years have seen a growing body of evidence to suggest a crucial role for ROS in the control of mycobacterial infections [7]. In particular, one group has recently identified Mtb nuoG as a potential virulence factor operating at the level of NOX2 by mechanisms yet to be defined [8].
The NOX2 complex consists of two constitutively associated transmembrane proteins, gp91phox and gp22phox and four cytosolic subunits: p40phox, p47phox, p67phox, and Rac1, a small GTPase [9]. Fully functional NOX2 requires membrane translocation of p40phox, p47phox, active Rac1 (GTP-bound form) and p67phox, and their assembly around gp91phox and gp22phox subunits [10]. NOX2 assembly leads to gp91phox activation to generate superoxide through a redox chain by transferring electrons from cytosolic NADPH to phagosomal oxygen [9]. The production of superoxide is in turn converted into several other microbicidal molecules, such as hydrogen peroxide and hydroxyl radicals, along with peroxynitrite when combined with nitric oxide radicals [9]. While the role of NOX2 in innate immunity is well established, several reports suggested that it might act beyond the control of intracellular infections to trigger macrophage apoptosis [11], [12], a central event that paves the road to adaptive immunity [13]–[15].
Previous results from our laboratory identified Mtb nucleoside diphosphate kinase (Ndk) as a GTPase Activating Protein (GAP) acting on Rab5 and Rab7 GTPases, leading ultimately to reduced phagolysosome fusion [16], [17]. In the present study, we examined whether Ndk GAP activity extends to other GTPases, with a particular focus on Rho GTPases. We found that Mtb Ndk interacts specifically with Rac1 and inactivates it leading to inhibition of NOX2 assembly and activation in the macrophage. We also established a link between Ndk-dependent NOX2 attenuation and inhibition of apoptosis response to Mtb. Consistent with these findings, Ndk knock down significantly reduced Mtb survival in vitro and in vivo.
We recently showed that mycobacterial Ndk plays an essential role in intracellular survival of the attenuated M. bovis BCG strain by a mechanism dependent on phagosome maturation arrest [17]. To examine whether Ndk also contributes to survival of virulent Mtb, we first attempted to generate an Ndk mutant in the Mtb strain H37Rv using various methods, including a gene disruption approach utilizing transducing mycobacteriophages [18]. Unfortunately, ndkA gene disruption affected severely the growth of bacteria. We therefore opted for protein knock down with mRNA antisense, the only approach developed so far to study essential genes in Mtb [19]. To do so, we transformed Mtb with the integrative vector pJAK1.A, previously designed by us [20], to express a stable full length antisense (or sense, control) mRNA sequence to ndkA. Thus, we generated a strain (Mtb Ndk-AS) in which Ndk protein expression was undetectable by western blot, even after many passages in the absence of the selection marker kanamycin, indicating a stable knock-down with the pJAK1.A vector (Fig. 1A). Fortunately, Mtb Ndk-AS displayed a similar growth profile to that of wild type Mtb and the control sense strain (Mtb Ndk-S) in standard culture media (Fig. 1B), as well as in the presence of oxidative stress (H2O2, Fig. S1). Thus potential fitness disadvantage that could be associated with genetic manipulation of Mtb are unlikely. Knock down of Ndk significantly affected Mtb survival in RAW 264.7 macrophages to the extent that at 72 h post infection, numbers of Mtb Ndk-AS dropped by 2 log colony-forming units (CFUs), relative to the wild type or Ndk-S strains (Fig. 1C). These findings suggested that the Ndk protein might contribute to Mtb virulence in vivo. Virulence during the early acute phase of infection is essentially controlled by innate immune responses and can be rapidly assessed in the SCID mouse model where innate immune responses are intact [21], [22]. In this regard, experiments of SCID mice infection by aerosol showed significant reduction (∼70%) of bacterial load in the lungs of Ndk-AS infected animals compared to those infected with wild type and Ndk-S strains (P = 0.002, unpaired t-test, Fig. 1D). Indicators of morbidity were apparent in the mice within 6 weeks with no significant difference between the three infection test groups (Fig. S2). However, when infected subcutaneously, time to death was extended to 12–15 weeks. Under these conditions, Kaplan Meier survival analysis clearly demonstrated that animals inoculated with Mtb Ndk-AS survived significantly longer (∼20 days, P<0.0001) than the control strain expressing Ndk sense mRNA, which caused the death of mice at similar rates seen in mice infected with the wild type strain (Fig. 1E). Taken together, these data demonstrated clearly that Ndk contributes to Mtb survival in the host through mechanisms that we have attempted to elucidate.
Our recent findings that Ndk expresses GAP activity towards Rab5 and Rab7 [17] suggested that this activity might extend to other host GTPases. Therefore, we examined whether Ndk targets macrophage Rho GTPases, known to play essential roles in early events of innate immunity against intracellular pathogens [23], [24]. To do so, macrophages were allowed to ingest Ndk-coated latex beads and then cell lysates were subjected to immunoprecipitation with Ndk antibody. Proteins associated with Ndk were analyzed by western blot with Rac1, Rho, or Cdc42 antibodies. The results obtained showed that only Rac1 was interacting with Ndk within the macrophage (Fig. 2A). Rac1 binding to Ndk was further confirmed with reverse pull down experiments using Rac1 antibody and western blotting with Ndk antibody, which showed clearly a physical association between Mtb Ndk and Rac1 (Fig. 2B). Results obtained with coated latex beads clearly demonstrated the specificity of Ndk-Rac1 interactions within the macrophage. Accordingly, we then examined Ndk-Rac1 interaction in cells infected with the bacterium instead of beads and showed that Rac1 antibodies are able to pull-down Ndk-Rac1 complexes from cells infected with wild type and Ndk-S but not Ndk-AS Mtb (Fig. 2C)
The results shown above (Fig. 2) suggested that Mtb Ndk must cross the phagosomal membrane towards the cytosol to bind to and inactivate Rac1. We first confirmed the hypothesis of cytosolic translocation of Ndk using i) confocal microscopy analysis, which showed diffused staining of Ndk distant from phagosomes containing wild type and Mtb Ndk-S but not from those containing Mtb Ndk-AS (Fig. S3A) and ii) immunogold staining and EM analysis, which clearly demonstrated that Mtb Ndk effectively crosses the phagosomal membrane toward the cytosol (Fig. S3B). We next examined the level of Rac1 activation in infected macrophages with pull down experiments using binding domain derived from Rac1 interacting protein (PAK-1 PBD), which interacts with Cdc42 as well. We also examined levels of Rho activation using Rho interacting protein (Rhotekin RBD). These binding domains interact only with GTP-bound forms of Rho GTPases [25]. Mtb infected RAW cells were exposed to LPS in order to activate the Rho GTPases, then cell lysates were examined for the amount of active Rac1, Rho, or Cdc42. Western blot analyses with Rac1, Cdc42 and Rho antibodies demonstrated that Mtb significantly inhibits the level of LPS-induced Rac1 activation (Fig. 3A, top panel). In contrast, Mtb had no apparent effect on Cdc42 and Rho activation. This GAP activity was also observed in macrophages ingesting Ndk-coated beads, as opposed to Mtb bacilli, demonstrating a specific Ndk GAP activity on Rac1 (Fig. 3A, lower panel).
To further examine Mtb effects on Rac1 and the phagosomal events it regulates, we performed a time-course Rac1 activation assay with macrophages infected by Mtb Ndk-AS and wild type Mtb. The results obtained showed a dramatic reduction of active Rac1 levels 15 min post infection and undetectable levels 1 h later in macrophages infected with wild type Mtb (Fig. 3B, top panel). In contrast, levels of active Rac1 remain unchanged in macrophages infected with Mtb Ndk-AS (Fig. 3B, bottom panel). Taken together, these data clearly demonstrated that Mtb Ndk expresses GAP activity on both basal and induced Rac1-GTP levels in the macrophage.
Active Rac1 (GTP bound form) has been shown to translocate to early phagosomes in order to facilitate recruitment of its binding partner, the NOX2 subunit p67phox [26], [27]. Given that Ndk expresses GAP activity towards Rac1 (GTP into GDP switch), we examined whether Mtb interferes with phagosomal recruitment of p67phox. We first applied intracellular staining and confocal microscopy to estimate the proportion of Rac1 and p67phox positive phagosomes in Mtb-infected RAW cells. Results obtained (Fig. 4A and 4B) showed a substantial reduction of Rac1 and p67phox positive phagosomes (13% and 29% respectively) in cells infected with live Mtb relative to those infected with killed Mtb (86% Rac1 and 88% p67phox positive phagosomes, respectively). In contrast, recruitment of p47phox to live Mtb phagosomes was comparable to that of phagosomes containing killed Mtb. To demonstrate that the NOX2 assembly defect is related to Ndk GAP activity, we applied similar confocal analyses to cells infected with Mtb Ndk-AS. The images (Fig. 4C and 4D) clearly showed a restoration of Rac1 and p67phox recruitment to Mtb Ndk-AS containing phagosomes to a level similar to those observed in cells infected with killed Mtb (∼% and 72% positive phagosomes, respectively). As expected, much lower numbers of Rac1 and p67phox positive phagosomes (13% and 30% respectively) were observed in cells infected with control strain Mtb Ndk-S.
As an alternative approach, a previously developed quantitative FACS analysis method [28] was used to assess the level of NOX2 components on individual phagosomes. To adapt this method to mycobacterial phagosomes, macrophage plasma membrane was stained with CellMask Deep Red (detectable by FL4 channel), and then cells were infected with Mtb strains expressing fluorescent DsRed protein (FL2). Following cell disruption, mycobacteria included in cell membrane-derived vacuoles (double FL2/FL4 positive events) were readily identified by flow cytometry analysis (Fig. S4). Phagosome preparations were then stained with Rac1 or p67phox antibodies and FITC-conjugated secondary antibodies (FL1). Samples were subjected to flow cytometry analysis and mean fluorescence intensities (MFI) were deduced from fluorescence histograms. Results obtained (Fig. 4E) showed higher recruitment of Rac1 and p67phox to phagosomes containing Mtb Ndk-AS (MFI: 49.6 and 42.6 respectively) relative to phagosomes containing Mtb Ndk-S (MFI: 25.3 and 21.8 respectively) or Mtb wild type (MFI: 23.9 and 15.7 respectively). To establish a direct link between Ndk GAP activity and defective NOX2 assembly, additional flow cytometry analyses were applied to phagosomes containing coated beads (Fig. S5) and showed a marked decrease of Rac1 and p67phox recruitment to the Ndk bead phagosomes (MFI: 8.1 and 3.8 respectively) relative to control phagosomes containing BSA-beads (MFI: 14.6 and 7.3 respectively). Taken together, these findings showed for the first time that Mtb uses Ndk GAP activity to disrupt phagosomal assembly of NOX2 via interference with Rac1-dependent recruitment of p67phox.
Previous findings that Rac1 and p67phox subunits are essential for NOX2 assembly and activation of gp91phox to generate superoxide [10] suggested that disruption of Rac1/p67phox translocation to the phagosome by Ndk would affect NOX2-dependent ROS production. To verify this hypothesis, we applied a luminol-dependent chemiluminescence assay to assess ROS production in Mtb infected cells. Luminol is a membrane diffusible reagent commonly used for quantitative detection of superoxide anion radicals and hydrogen peroxide molecules. Bone marrow derived macrophages (BMDM) from C57BL/6 mice were infected with Mtb strains and assayed for kinetics of chemiluminescence production over a period of 60 min. Relative luminescence profiles obtained (Fig. 5A) revealed that cells infected with Mtb Ndk-AS induced significantly higher amounts of ROS production (peak value = 256 RLU) compared to those infected with wild type Mtb or Mtb Ndk-S (peak value ∼120 RLU). Thereafter, we confirmed the apparent inhibitory effect of Ndk with experiments using coated beads (Fig. 5B), which showed minor ROS response to Ndk beads (peak value = 32 RLU) relative to ROS production induced by BSA beads (peak value = 76 RLU). Additionally we applied confocal microscopy to visualize intracellular accumulation of ROS using CM-DCFDA, a cell-permeable probe that is non-fluorescent until oxidized within the cell. Thus, in RAW cells infected with Mtb Ndk-AS, the confocal images showed a strong colocalization of oxidized CM-DCFDA (green fluorescence) with bacterial phagosomes (red fluorescence) indicating accumulation of large amounts of ROS around Mtb Ndk-AS (Fig. 5C and 5D). Conversely, green signal was totally absent in cells infected with either wild type Mtb or Mtb Ndk-S. This effect of Ndk on ROS production was also reproduced when BMDM were used instead (Fig. S6).
Previous studies reported that mitogen-activated protein kinases (MAPKs) play an important role in the signaling pathway of NOX2 activation [29], [30]. To verify whether Ndk also interferes with MAPK activation, macrophages were allowed to ingest Ndk-beads or BSA-beads (control), and then stimulated with PMA or LPS to activate ERK1/2, and p38MAPK respectively. Cell lysates were then examined for the level of phospho-ERK1/2 and phospho-p38MAPK, which reflects kinase activation. The western blot results obtained (Fig. S7) did not reveal any changes in the levels of kinase phosphorylation in cells infected with Ndk-beads relative to those infected with BSA-beads. Therefore, Ndk effect on NOX2 is clearly independent of MAPK inhibition.
Collectively, these experiments demonstrated that the macrophage oxidative response to Mtb is marginal and that knock down of Ndk converts the bacterium into a potent inducer of the ROS response.
Mtb is known to inhibit macrophage apoptosis [14], [15] by mechanisms yet to be clarified. Based upon previous findings that NOX2 activity might extend beyond intracellular killing to induce apoptosis [8], [14], we examined whether Mtb uses Ndk to disrupt the NOX2-apoptosis link. First, we applied Annexin V cell surface staining, a popular approach for detection of phosphatidylserine (PS) translocation to the extracellular membrane leaflet, which reflects early stages of apoptosis events [31]. Adherent RAW cells on coverslips were infected with Mtb strains for 48 h then stained with Alexa Fluor 488 conjugated Annexin V and examined by confocal microscopy. The images showed very low numbers of Annexin V positive cells in samples infected with wild type Mtb and Mtb Ndk-S (5% and 6% positive, respectively). In contrast, a higher number of Annexin V positive cells (44%) was observed in samples infected with Mtb Ndk-AS (Fig. 6A, top panel). To establish a direct link between ROS and apoptosis in infected cells, Annexin V staining was repeated on macrophages exposed to Mtb Ndk-AS in the presence of a specific gp91phox peptide inhibitor (gp91 INH) or its control scrambled version (gp91 SCR) [32]. The results obtained showed clearly that gp91 INH, but not gp91 SCR, reversed completely Mtb Ndk-AS-induced PS translocation to the cell surface (5% Annexin V positive, Fig. 6A, bottom panel). The effect of the gp91phox inhibitor was confirmed with experiments showing that gp91 INH completely inhibited ROS production in cells infected with Mtb Ndk-AS, which was normally elicited in the presence of gp91 SCR (Fig. S8). In a complementary series of experiments, we analyzed caspase 3 activation, which occurs during the final stages of apoptosis [33]. Thus, infected macrophages were subjected to intracellular staining with antibody to cleaved (i.e. active) caspase 3 and Alexa Fluor 647 conjugated secondary antibody, then analyzed by FACS. Results obtained (Fig. 6B) showed higher numbers of apoptotic macrophages in sample tests infected by Mtb Ndk-AS (11.8% positive events) relative to those infected with wild type Mtb or Mtb Ndk-S (∼4.6%). Not surprisingly, the wild type and Ndk-S strains inhibited the spontaneous apoptosis observed in control non-infected cells (7.3% positive cells). As expected, Mtb Ndk-AS-induced caspase 3 cleavage was abolished in the presence of the gp91phox inhibitor.
It is well known that apoptosis is also induced by nitric oxide (NO) in mouse macrophages [34], [35]. Therefore, the effect of Ndk on macrophage apoptosis might be the result of simultaneous inhibition of ROS and NO production. To verify this possibility we examined IFN-γ-induced NO production in cells infected with Ndk-beads and BSA-beads (control) and the results deduced from the Griess assay (Fig. S9) demonstrated that Ndk has no effect on NO production.
Taken together, our data demonstrated that Mtb blocks macrophage apoptosis by a mechanism dependent, at least in part, on Ndk-mediated attenuation of NOX2 activity.
Results presented above (Fig. 6) together with initial experiments showing attenuated Mtb Ndk-AS survival in RAW macrophages (Fig. 1) suggested that Ndk-mediated inhibition of ROS reduces the macrophage killing capability. To verify this hypothesis, we repeated the survival assay using primary murine macrophages in which ROS production was blocked with gp91 INH. At 72 h post-infection, control experiments showed a significant reduction (∼1.5 Log10) in CFU counts when infecting with Mtb Ndk-AS relative to wild type or Ndk-S (Fig. 7A). However, in the presence of gp91 INH, Ndk-AS survival was restored to a level comparable to that of wild type Mtb at every time point measured (Fig. 7B). These observations were validated with assays in the presence of control scrambled peptide, which did not affect Ndk-AS survival. Taken together, these experiments clearly demonstrated that down modulation of ROS production by Ndk contributes significantly to Mtb persistence in the macrophage.
Previous studies from this laboratory showed that Mtb Ndk exhibits GAP activity towards Rab5 and Rab7 leading ultimately to diminished phagosomal recruitment of their respective effectors EEA1 and RILP [16], [17]. Defective recruitment of EEA1 and RILP correlated with reduced maturation of phagosomes containing Ndk mutant M. bovis BCG or Ndk-coated beads [17]. In the current study, we demonstrate that Ndk further enhances Mtb virulence by additional GAP activity towards the Rho GTPase Rac1. We provided direct evidence that Ndk blocks phagosomal recruitment for both Rac1 and its partner molecule p67phox leading ultimately to inhibition of NOX2-mediated ROS production and ROS-mediated apoptosis. A link between Mtb GAP activities and virulence was established with the observation of reduced survival of Ndk mutant Mtb in vitro and in vivo.
Ndk is a ubiquitous small protein (∼15 kDa) found in virtually all organisms, from eukaryotes to prokaryotes. In Mtb, Ndk catalyzes the production of nucleoside triphosphates as precursors for RNA, DNA and polysaccharide synthesis, which are critical for normal bacterial physiology [36]. This possibly explains why our attempts to knock out the ndkA gene in Mtb were unsuccessful, suggesting that Ndk is probably essential for Mtb growth. Contrasting with this hypothesis, an effort to comprehensively identify all genes required for Mtb growth using the transposon site hybridization (TraSH) technique suggested that the Ndk gene is not essential for Mtb growth [37]. However, as cautioned by the authors of that study, TraSH is simply a screening tool and therefore cannot provide a definitive conclusion about gene essentiality. Indeed, several genes known to be essential for Mtb growth, such as ideR [38], rmlD [39] and whiB2 [40] have not been identified as essential by the TraSH approach. Thus, whether or not Ndk is essential is a research question that is still open for further investigation and remains beyond the scope of our current study, which focused instead on deciphering the mechanisms by which Ndk promotes Mtb survival in the host.
Mycobacterial Ndk has been shown to interact with and inactivate recombinant Rho, Cdc42 and Rac1 proteins [41]. Here we found that within the macrophage, both Mtb and recombinant Ndk (delivered on the surface of latex beads) interact with and inactivate native Rac1, but not Rho or Cdc42. This suggests that results obtained from cell free systems do not always reflect host-pathogen interactions in the whole cell system. Not surprisingly, this type of discrepancy has been observed with other pathogens that use GAP activities as a mechanism of virulence. For instance, secreted YopE from Yersinia, and SptP from Salmonella were shown to have GAP activity towards all three Rho GTPases extracellularly. However, YopE acts only on Rac1 and RhoA, [42] whereas SptP inactivates Rac1 and Cdc42, but not RhoA, [43] in cultured cells. In the case of Yersinia, a recent study established a direct link between YopE-mediated inactivation of Rac1 and inhibition of ROS production [44] consistent with our findings that selective Ndk GAP activity towards Rac1 is sufficient to block ROS production in the macrophage. Inhibition of ROS production in nascent phagosomes has also been reported in macrophages infected with the protozoan Leishmania, an intracellular pathogen that is structurally and metabolically distinct from Mtb, which interferes with NOX2 by a mechanism independent of GAP activities [45]. Indeed, Leishmania was shown to use its abundant surface lipophosphoglycan to restrict phagosomal recruitment of both p47phox and p67phox but not Rac1. Conversely, our study showed that Ndk disrupts the recruitment of Rac1 and its binding partner p67phox but not p47phox. The p47phox subunit contains a PRR (Prolin Rich Region) at its C-terminus that binds with high affinity to the C-terminal SH3 domain of p67phox in the cytosol [46], [47]. It also contains a PH (Pleckstrin Homology) domain that interacts specifically with membrane PI[3,4]P2 and phosphatidic acid [48]. While the tail-to-tail association of p47phox and p67phox plays a crucial role in NOX2 assembly [27], recent studies showed that it is rapidly disrupted after membrane translocation [49]. Therefore dissociated p47phox and p67phox would remain separately attached to the phagosome via membrane lipids and Rac1 respectively. This phagosomal configuration of NOX2 subunits is consistent with the specific dissociation of p67phox from Mtb phagosomes as a result of Ndk-mediated Rac1 inactivation.
The overall emerging picture from ongoing studies of phagosome remodelling by Mtb suggests that more than one virulence determinant might act in concert to modulate a single event of phagosome biogenesis. For instance, the cell wall glycolipid lipoarabinomannan, which blocks the Ca2+ signaling pathway [50], synergizes with the acid phosphatase SapM, which hydrolyzes PI[3]P [51], to abolish phagosome maturation processes that are dependent on recruitment of EEA1. Such a synergism appears to also be the case for mycobacterial interference with NOX2 activity on the phagosomal membrane. Indeed, a recent study showed that the NuoG subunit of the type I NADH dehydrogenase also promotes Mtb interference with NOX2 activity, as evidenced by increased levels of ROS on Mtb ΔnuoG phagosomes [8]. However, the finding that NuoG is not secreted raises a question about the mechanistic connection between distant NuoG, contained within the bacterial cytosol, and NOX2 components on the cytosolic face of the phagosome membrane. Conversely several different groups have shown that Mtb Ndk is secreted [52]–[54], suggesting that Ndk could translocate to the cytosolic surface of the vacuole where it interacts with Rac1. In fact, our EM and confocal data revealed that secreted Ndk crosses the phagosomal membrane towards the cytosol. Consistent with these findings, previous studies showed that live Mtb exports a variety of proteins and glycolipids intracellularly [55]–[57], and that many of them cross the phagosomal membrane towards the host cell cytosol to interact with and inhibit critical regulators of phagosome biogenesis [51], [56], [58]. A possible mechanism for the cytosolic translocation of mycobacterial products is the generation of a semi-porous phagosome membrane by the Mtb ESX-1 secretion system [59], which was also shown to play an essential role in Mtb escape from the phagosome in later stages of infection [60], [61]. Therefore it is possible that the ESX-1 secretion system also mediates translocation of Ndk to the cytosol.
A highly relevant finding from the present study is that Ndk knock down converted virulent Mtb into an attenuated strain that lost resistance to the hostile environment of the host cell. Indeed, Mtb Ndk-AS infected cells were able to generate NOX2-dependent ROS production and also to undergo apoptosis thus ensuring maximal restriction of bacterial proliferation. In contrast, virulent Mtb strains were shown to down-modulate apoptosis in favor of necrosis [62], [63], which releases viable intracellular bacilli for further spreading of the infection and tissue damage during active tuberculosis disease. The link between ROS production, apoptosis and intracellular killing demonstrated in our study is consistent with earlier findings that intracellular oxidative stress induces phosphatidylserine externalization and increased caspase 3 activity [64], [65], and that apoptosis induced by the Fas ligand attenuates Mtb survival within the macrophage [66]. In addition to restricting the niche for mycobacterial replication, macrophage apoptosis contributes indirectly to the initiation of adaptive immunity mediated by dendritic cells. Indeed, infected macrophages undergoing apoptosis shed vesicles loaded with bacterial material (or apoptotic blebs) that prime dendritic cells for enhanced presentation of mycobacterial antigen to T cells [13]–[15].
In summary, while the role of Ndk in physiological processes has been intensively investigated, its contribution to Mtb pathogenesis has not been previously addressed. Our recent findings and current investigations have extended the knowledge of the biological effects of Ndk, to include inactivation of GTPase effector functions in the macrophage, therefore highlighting a novel strategy used by Mtb to circumvent host innate immunity
DMEM, Fetal calf serum (FCS), and HBSS were purchased from Gibco Laboratories (Burlington, ON, Canada). Luminol, CM-DCFDA, Annexin V-488, and CellMask Deep Red, were purchased from Invitrogen (Burlington, ON, Canada). Endotoxin-free culture reagents were from StemCell Technologies (Vancouver, BC, Canada). Protease inhibitor mixture, PMSF, and trypsin-EDTA were purchased from Sigma-Aldrich (St. Louis, MO). Protein A-agarose beads were from Bio-Rad laboratories (Hercules, CA). Aldehyde/sulfate latex beads (diameter, 4 µm) were obtained from Interfacial Dynamics (Portland, OR). gp91phox inhibitor peptide and its scrambled version [34] were synthesized by GenScript (Piscataway, NJ).
Rac1, RhoA, Cdc42 antibodies were purchased from Millipore (Temecula, CA). p67phox antibody was purchased from BD Transduction Laboratories (Mississauga, ON, Canada) and 47phox antibody was purchased from Santa Cruz Biotechnology (Santa Cruz, CA). Cleaved caspase-3 (Asp175) antibody was purchased from Cell Signaling (Danvers, MA). Alexa Fluor 647-conjugated anti-rabbit IgG was purchased from Invitrogen. FITC-conjugated anti-rabbit and anti-mouse IgG were purchased from Sigma Aldrich.
M. tuberculosis H37Rv and its derivative strains were grown in Middlebrook 7H9 broth (BD Diagnostic Systems, Mississauga, ON, Canada) supplemented with 10% (v/v) OADC (oleic acid, albumin and dextrose solution; BD Diagnostic Systems) and 0.05% (v/v) Tween 80 (Sigma-Aldrich) at 37°C on standing culture. Mtb Ndk-AS and Ndk-S were generated using our integrative pJAK1.A plasmid (selection marker, kanamycin [17] encoding the full length ndkA gene in sense and anti-sense orientations as described [17]. To generate red-fluorescent bacteria, Ndk-S, Ndk-AS, and wild type Mtb strains were transformed with pSMT3 vector (selection marker, hygromycin) encoding the DsRed protein as described [16].
RAW 264.7 macrophages (ATCC, Manassas, VA) were maintained in 10 cm diameter culture dishes (Corning Inc., Corning, NY) at a density of ∼105 per cm2 in Endotoxin-free DMEM containing 5% FCS and 1% each of L-glutamine, penicillin-streptomycin mixture, HEPES, non-essential amino acids (100× solution, StemCell). Bone marrow derived macrophages (BMDM) were obtained by flushing out femurs and tibias of 6–8 week old female C57BL/6 (Jackson Laboratory, Sacramento, CA) according to protocols approved by the University of British Columbia Animal Care and Use Committees. Cells were then maintained in complete DMEM containing 10 ng/mL M-CSF for 6 days. For macrophage infection, bacteria in mid-log phase were harvested by 5 min centrifugation at 8,000× g. They were subsequently washed twice with 7H9 plus 0.05% tween and passed several times through 25 gauge needles to break bacterial clumps. Thereafter, numbers of bacteria were normalized by optical density (OD600 1.0 = 3×108 bacteria/ml) and adjusted for the desired MOI. Macrophages were then exposed to Mtb strains in complete media without antibiotic for 2 h at 37°C and then washed thrice to remove extracellular bacteria. Cells were reincubated in complete media plus gentamicin (50 ug/ml) for the desired time periods.
Groups of 4- to 6-week old female Fox Chase SCID mice (CB17/Icr-Prkdcscid/IcrCrl) were infected with ∼150 bacteria by inhalation using a Glas-Col inhalation exposure system (Terre Haute, IN). Two mice from each group were processed on day 1 following infection to confirm bacterial deposition in the lung. Remaining animals were monitored for signs of morbidity. Mice were then euthanized and the bacterial load (CFUs) in the lung was determined. Organs were homogenized and serial dilutions plated in duplicate on nutrient 7H10 agar. In other experiments, SCID mice were injected subcutaneously in the scruff of the neck with 106 Mtb strains and then monitored for morbidity over a period of ∼15 weeks.
All animals were maintained in accordance with protocols approved by the Animal Care and Use Committees at the University of British Columbia. Experiments were approved by the Animal Care and Usage Committees and performed according to the Canadian Council on Animal Care Guidelines. The animal assurance welfare number is A11-0247.
Ndk was expressed as a C-terminal 6×His tagged fusion protein in E. coli strain BL21 and purified using Ni-NTA purification resin (Qiagen) as described [17]. The purity of eluted Ndk was confirmed by SDS-PAGE and Coomassie Blue staining (Fig. S10A). Rabbit Ndk antibody was generated by GenScript, using KLH conjugated ELASQHYAEHEGK peptide fragment corresponding to amino acids 44 to 56 of Mtb Ndk. The specificity of Ndk antibody is shown in Fig. S10B. Ndk or the control BSA were non-covalently linked to latex beads as described [17].
Coverslips were mounted on microscope slides and examined by digital confocal microscopy as described [17].
Immunogold staining was conducted at the EM Facility of the James Hogg Research Centre (Saint Paul Hospital, Vancouver, BC, Canada). In brief, Mtb-infected macrophages were fixed with 4% paraformaldehyde, embedded in 4% low melting point agarose and dehydrated in ethanol. Samples were then transferred to LR White resin. After polymerization at 50°C, 60 nm sections were cut with a Leica EM UC6 microtome (Leica Microsystems, Switzerland) and collected on nickel grids. Samples were labelled with Ndk antibodies then F(ab′)2 of ultra-small goat-anti-rabbit IgG (Electron Microscopy Sciences, Hatfield, PA). Sections were then post-fixed in 2% glutaraldehyde and subjected to silver enhancement for gold labeling with Silver R-Gent SE-EM (Aurion, Wageningen, Netherlands). Samples were then washed in distilled water, stained in 2% uranyl acetate, washed again, air dried and examined with a Tecnai 12 electron microscope (FEI Company, Hillsboro, OR).
Confluent RAW cells seeded on 6 cm plates were infected by Mtb strains for 1 h at a MOI of 20∶1. Thereafter, cells were incubated in the presence of 200 ng/ml LPS for 15 min to induce Rho GTPase activation. Subsequently, cells were lysed in cold buffer containing 30 mM HEPES (pH 7.2), 100 mM NaCl, 10% glycerol, 1% Triton X100, 1 mM EDTA, 10 mM MgCl2, and 1 mM PMSF. Soluble protein fractions were analyzed for levels active Rho-GTPases by using a Rac/Rho/Cdc42 activation assay kit (Millipore).
Macrophages were cultured in complete DMEM in 96 well white plates (Corning) at 105 per well. Prior to ROS assay, cell media was replaced with DMEM without phenol red and luminol was added to a final concentration of 50 µM. Wells were then infected with Mtb strains (MOI 10∶1) or coated beads (MOI 5∶1). Thereafter, plates were loaded into a Tropix TR717 microplate luminometer (Applied Biosystems, Bedford, MA) adjusted to 37°C and relative luminescence was then measured at 60 sec intervals over 60 min. Intracellular detection of ROS was achieved by incubating adherent macrophages to cover slips with 10 µM CM-DCFDA for 30 min at 37°C prior to infection with Mtb strains expressing DsRed. Cells were then analyzed by confocal microscopy.
BMDMs, but not RAW cells, were primed with LPS (100 ng/ml, 48 h) prior to ROS assays because expression of fully functional NOX2 complex in BMDMs requires priming with LPS or TNFα [67], [68].
Adherent RAW cells on coverslips were infected with Mtb strains. At 48 h post phagocytosis, cells were washed twice with cold PBS and then incubated in Alexa Fluor 488 Annexin V (1∶20, Invitrogen) in staining buffer containing 10 mM HEPES (pH 7.4), 140 mM NaCl, and 2.5 mM CaCl2, for 20 min at room temperature. Coverslips were then analyzed by confocal microscopy. Alternatively, infected cells were scraped and fixed in PBS plus 2% paraformaldehyde for 15 min at room temperature. Cells were then washed with PBS and incubated with anti-cleaved caspase-3 (1∶250) in permeabilization buffer (0.1% Triton X100 and 1% BSA in PBS) for 20 min at room temperature. Thereafter, cells were washed and stained with Alexa Fluor 647-conjugated goat anti-rabbit IgG (1∶200) for 20 min at room temperature, washed again and analyzed by FACS.
|
10.1371/journal.pgen.0030069 | The tailless Ortholog nhr-67 Regulates Patterning of Gene Expression and Morphogenesis in the C. elegans Vulva | Regulation of spatio-temporal gene expression in diverse cell and tissue types is a critical aspect of development. Progression through Caenorhabditis elegans vulval development leads to the generation of seven distinct vulval cell types (vulA, vulB1, vulB2, vulC, vulD, vulE, and vulF), each with its own unique gene expression profile. The mechanisms that establish the precise spatial patterning of these mature cell types are largely unknown. Dissection of the gene regulatory networks involved in vulval patterning and differentiation would help us understand how cells generate a spatially defined pattern of cell fates during organogenesis. We disrupted the activity of 508 transcription factors via RNAi and assayed the expression of ceh-2, a marker for vulB fate during the L4 stage. From this screen, we identified the tailless ortholog nhr-67 as a novel regulator of gene expression in multiple vulval cell types. We find that one way in which nhr-67 maintains cell identity is by restricting inappropriate cell fusion events in specific vulval cells, namely vulE and vulF. nhr-67 exhibits a dynamic expression pattern in the vulval cells and interacts with three other transcriptional regulators cog-1 (Nkx6.1/6.2), lin-11 (LIM), and egl-38 (Pax2/5/8) to generate the composite expression patterns of their downstream targets. We provide evidence that egl-38 regulates gene expression in vulB1, vulC, vulD, vulE, as well as vulF cells. We demonstrate that the pairwise interactions between these regulatory genes are complex and vary among the seven cell types. We also discovered a striking regulatory circuit that affects a subset of the vulval lineages: cog-1 and nhr-67 inhibit both one another and themselves. We postulate that the differential levels and combinatorial patterns of lin-11, cog-1, and nhr-67 expression are a part of a regulatory code for the mature vulval cell types.
| During development, in which the single-celled egg generates a whole organism, cells become different from each other and form patterns of types of cells. It is these spatially defined fate patterns that underlie the formation of complex organs. Regulatory molecules called transcription factors influence the fate patterns that cells adopt. Understanding the role of these transcription factors and their interactions with other genes could tell us how cells establish a certain pattern of cell fates. This study focuses on studying how the seven cell types of the Caenorhabditis elegans vulva arise. This organ is one of the most intensively studied, and while the signaling network that initiates vulval development and sets the gross pattern of cell differentiation is well understood, the network of transcription factors that specifies the final cell fates is not understood. Here, we identify nhr-67, a new transcription factor that regulates patterning of cell fates in this organ. Transcription factors do not necessarily act alone, and we explore how NHR-67 works with three other regulatory factors (each with human homologs) to specify the different properties of the vulval cells. We also demonstrate that the interconnections of these transcription factors differ between these seven diverse cell types, which may partially account for how these cells acquire a certain pattern of cell fates.
| Complex gene regulatory networks operating in diverse cell types and tissues are crucial for development. Diverse intercellular signals and transcription factor networks control gene expression within individual cell types, acting on cis-regulatory modules of target genes [1]. Understanding such regulation first requires documenting all the regulatory inputs and outputs from each gene [2]. This information allows circuit diagrams to be constructed that provide a global perspective on how diverse cell types acquire their identity. Gene regulatory networks have been well studied in a wide range of biological model systems such as endomesoderm specification in the sea urchin embryo [3], dorso-ventral patterning in the Drosophila embryo [4], and mesoderm specification in Xenopus [5]. The common themes that might emerge from these studies would advance our understanding of organogenesis in vertebrates.
The Caenorhabditis elegans vulva is postembryonically derived from six vulval precursor cells P3.p–P8.p. The central three vulval precursor cells P5.p–P7.p are induced to adopt 1° (primary) and 2° (secondary) vulval fates via epidermal growth factor (EGF) and Notch signaling, whereas the remaining precursors fuse with the hypodermal syncytium hyp7 [6]. The vulva is composed of seven distinct cell types, each with its own set of expressed genes and morphogenetic migrations [7–9]. The P6.p 1° lineages generate the vulE and vulF cells, while the P5.p and P7.p 2° lineages generate the vulA, vulB1, vulB2, vulC, and vulD cells. The signals that induce 1° versus 2° fates in the primordial vulval precursor cells are known. However, the processes that govern patterning and differentiation of the mature vulval cell types are largely unknown [6]. Both Ras and Wnt pathways are required for the precise spatial patterning of the 1° vulE and vulF cells [10], and both Wnt/Ryk and Wnt/Frizzled signaling pathways are necessary for patterning the P7.p 2° vulA–vulD cells [11–13].
Genes expressed in the mature vulval cell types include some with known functions and many others without known physiological roles. lin-3 (EGF) is expressed in vulF and is required to signal from vulF to uterine uv1 cells [14,15]. egl-17 encodes a fibroblast growth factor (FGF)-like protein that is required for migration of the sex myoblasts to their precise final positions [16,17]. egl-17 is initially expressed in the 1° vulval lineages and is shut off during the L4 stage. Expression in vulC and vulD is observed during early L4 and persists throughout adulthood. The vulval expression correlates with the sites of muscle attachment. egl-26 encodes a novel protein that contains an H box/NC domain and is expressed in vulB1, vulB2, vulD, and vulE cells [18,19]. zmp-1 encodes a zinc metalloprotease and is expressed in vulD and vulE during the L4 stage and in vulA in adults [9]. ceh-2 encodes a homeodomain protein that is related to Drosophila empty spiracles and is expressed in vulB1 and vulB2 cells during the L4 stage and in vulC upon entry into L4 lethargus [9,20]. pax-2 is a recent gene duplication of the PAX2/5/8 protein EGL-38 [21] and is expressed exclusively in the vulD cells. zmp-1, ceh-2, egl-26, and pax-2 have no known function in the vulva.
Transcription factor networks in individual vulval cell types somehow generate a spatially precise pattern of cell fates [19]. Several transcription factors that regulate gene expression in the diverse vulval cell types have already been described [19,22–24]. lin-11, a LIM homeobox transcription factor, regulates gene expression in all seven vulval cell types [25,26]. The Nkx6.1/Nkx6.2 homeodomain gene, cog-1, regulates gene expression in vulB, vulC, vulD, vulE, and vulF cells [19,27]. In contrast, egl-38 encodes a PAX2/5/8 protein that appears to be the only known example of a vulval cell type–specific regulatory factor; it promotes expression of certain target genes and restricts expression of other targets exclusively in vulF cells [14,19,28]. Additional regulatory factors need to be identified to elucidate the precise spatial patterning of the mature vulval cell types.
Here, we identify nhr-67 as a component of the gene regulatory networks underlying vulval patterning and differentiation. nhr-67 is required for the accurate patterning of gene expression and regulation of cell fusion in several vulval cell types and is dynamically expressed in the vulva. nhr-67 interacts genetically with cog-1, egl-38, and lin-11 to produce the complex expression patterns of their downstream targets. We demonstrate that the pairwise interactions between these four regulatory genes vary among the diverse vulval cell types. These results indicate that nhr-67, cog-1, lin-11, and egl-38 form a part of a genetic network that generates different patterns of gene expression in each of the seven cell types.
An RNA interference (RNAi) screen of 508 known and putative transcription factors encoded in the C. elegans genome (see Table S1) was conducted in a ceh-2::YFP reporter background. At the time we performed the screen, this was the best available set. ceh-2 encodes a homeodomain protein orthologous to Drosophila Empty Spiracles (EMS) and vertebrate EMX1 and EMX2 and serves as a readout for vulB fate during the L4 stage [20]. Modifiers of ceh-2 expression are good candidates for genes involved in patterning and/or differentiation of 2° vulval descendents. From this screen, we identified nhr-67 as a gene necessary for negative regulation of ceh-2 expression in the 1° vulE and vulF cells (Figure 1A–1B). Reciprocal BLAST searches indicate that nhr-67 encodes an ortholog of the tailless hormone receptor, which consists of an N-terminal transactivation domain, a centrally positioned DNA-binding domain, and a C-terminal ligand-binding domain. The only other positive was the GATA-type transcription factor egl-18, which was previously shown to be involved in vulval development [29–31]. Other genes that should have been positive in the screen (lin-11 and cog-1) were not isolated from the RNAi screen, thus indicating a high false-negative rate. Analysis of the nhr-67 deletion allele ok631 revealed severe defects in early larval development (L1 lethality and/or arrest). In order to bypass this early larval arrest phenotype, we resorted to feeding young L1 larvae with nhr-67 RNAi and assayed for defects in vulval gene expression. nhr-67 was also found to be required for negative regulation of two additional L4-specific markers: egl-26 (wild-type expression in vulB, vulD, and vulE cells) (Figure 1C–1D) and egl-17 (wild-type expression in vulC and vulD cells) in the vulF cells. Thus, nhr-67 activity is necessary for the negative regulation of expression of several 2° lineage-specific genes in the 1°-derived vulval cells during the L4 stage. Consistent with previous reports, nhr-67 RNAi results in a highly penetrant protruding vulva (Pvl) and egg-laying (Egl) defective phenotype [32] (Figure S1). However, other transcription factors exhibiting a Pvl RNAi phenotype, such as fos-1, egl-43, and unc-62, have normal vulval gene expression (unpublished data).
In addition to its negative regulatory role, we also found that nhr-67 is necessary for promoting expression of specific genes. For example, nhr-67 is necessary for zmp-1 expression in vulA during the adult stage (Figure 1E–1F). nhr-67 is also required for vulD-specific expression of pax-2 and egl-17 during the L4 stage (Figure 1G–1J). These examples show that nhr-67 positively regulates gene expression in the secondary vulA and vulD cells. nhr-67 is also required for positively regulating gene expression in the 1° vulval cells, namely vulF-specific expression of lin-3, an EGF-like protein (Figure 1K and 1L). Therefore, nhr-67 regulates gene expression in at least four of the seven vulval cell types.
In the L3 stage, the early 1° and 2° vulval cell fates can be distinguished by the patterns of cell division of their descendents. The 1° fated cell typically gives rise to four granddaughters that divide transversely (left-right axes); whereas a subset of the granddaughters derived from a 2° cell divide longitudinally (anterior-posterior axes). To determine if nhr-67-dependent alterations in gene expression are a consequence of fate transformations in the early 1° and 2° vulval lineages, we monitored the pattern of the vulval cell divisions in an nhr-67 RNAi background. In the absence of nhr-67, the vulval cell lineages appear wild-type in terms of both cell number and orientation of cell division (unpublished data). Thus, the perturbations in gene expression caused by reduced nhr-67 function are not the result of gross abnormalities in the early vulval cell lineages.
During the L4 stage, the seven vulval cell types invaginate cooperatively to assume a characteristic morphology. The similar cell types subsequently fuse, generating toroid rings that line the vulval cavity [8]. We wanted to ascertain if the observed cell fate transformations in nhr-67(RNAi) animals were possibly due to improper fusion events between the wrong cell types. Cell fusion defects can be assayed using ajm-1::GFP (an adherens junction marker) to visualize the cell number and architecture of the vulval toroids. When observing the mid-sagittal plane of wild-type animals, ajm-1::GFP appears as dots between cells. The eight dots on either side correspond to the seven distinct vulval cell types (Figure 2A). Most nhr-67 RNAi–treated animals do not exhibit dramatic defects in cell fusion (Figure 2B). The 2° vulval lineage–derived cells (vulA, vulB1, vulB2, vulC, and vulD) consistently generate mature toroids. However, inappropriate fusion often occurs (65%, n = 17) between the presumptive vulE and vulF cells (indicated by the missing dots at the top of the vulval invagination) (Figure 2C). Since nhr-67 regulates gene expression in vulval cells other than vulE and vulF, improper cell fusion events cannot fully account for all its altered gene expression patterns.
We then wanted to determine if the altered gene expression occurring in the 1° vulval cells was dependent on these improper fusion events. We attempted to address this question using two approaches: (a) by analyzing the effect of nhr-67 RNAi on the expression of egl-17 and ceh-2 transgenes in an eff-1(hy21) background, and (b) by monitoring the vulval expression levels of eff-1 in animals with reduced nhr-67 activity. eff-1 is a type I membrane protein necessary for cell fusion [33]. Disruption of nhr-67 function in an eff-1-deficient background is still sufficient to cause upregulation of both egl-17 (Figure 2D and 2E) and ceh-2 (Figure 2F and 2G) in the 1° vulval cells. Thus, the nhr-67-dependent alterations in gene expression are not dependent on eff-1-mediated cell fusion. We also observed that eff-1 levels (strong expression in vulA and vulC cells, weak expression in vulF cells) are highly elevated in vulD and vulF cells when nhr-67 gene activity is compromised (Figure 2H and 2I). However, we also note that eff-1 is not sufficient to rescue the vulE-vulF fusion defects observed in nhr-67 (RNAi) background (unpublished data). One possibility is that eff-1(hy27) is a temperature-sensitive allele that fails to completely eliminate cell fusion. Another possibility is that in addition to eff-1, nhr-67 negatively regulates other target genes that mediate cell fusion.
Previous work reported that an nhr-67 construct containing 6 kb of the promoter region directs expression in several head neurons [34]. We generated several additional transcriptional reporter constructs that tested the entire nhr-67 coding region, introns and the 3′ noncoding region for enhancer activity using the Δpes-10 basal promoter [35]. An 8-kb fragment that consisted of 1-kb 5′ sequence, the entire coding region and introns, and 2 kb of the 3′ noncoding region yielded expression in the vulva, the hyp7 epidermal syncytium, late stage embryos, and the male tail (Figures 3 and 4A). This nhr-67 construct exhibits a dynamic expression pattern in the vulval cells. During the late L4 stage, nhr-67 is first observed in vulA cells (Figure 3A) (and occasionally in vulB1), and this expression is maintained throughout adulthood. Expression in vulC is only seen upon entry into L4 lethargus and persists in adults (Figure 3B). Strong vulB1 and vulB2 expression (and occasional vulD expression) is observed only in young adults (Figure 3C). A 4.5-kb reporter construct that spans from the fourth intron to the 3′ noncoding region is sufficient to drive expression in the same tissues as seen with the 8-kb fragment (Figure 4B). No expression is seen in the vulC, vulD, vulE, and vulF cells during the L4 stage unless nhr-67 or cog-1 activity is eliminated (see below). Thus, the cis-elements driving the vulval expression of nhr-67 appear to be located in the region spanning the fourth intron to the 3′ noncoding region. We then wanted to confirm if these regulatory elements were capable of interacting with the endogenous promoter of nhr-67 in order to promote its transcription in the vulva. To test this, we generated an nhr-67 transcriptional reporter driven by 1 kb of its native promoter and containing regulatory sequences downstream of the fourth exon in their normal context. The nhr-67 transcriptional construct containing the endogenous promoter recapitulated the vulval and embryonic expression pattern observed with the nhr-67::Δpes-10 constructs (Figure 4C). We also examined whether the upstream regulatory sequences of nhr-67 interact with the downstream regulatory elements to influence its vulval expression. This test was accomplished by coinjecting a transcriptional green fluorescent protein (GFP) construct that contains a 6-kb upstream sequence of nhr-67 (Figure 4D) with the 8-kb nhr-67::Δpes-10 construct (Figure 4A) described above. We find that in the presence of the 6-kb promoter region, the vulval expression is identical to that of the 8-kb nhr-67::Δpes-10 constructs. Besides the previously reported expression in head neurons, we observed expression in the anchor cell (AC) (during mid–late L3 stage) in hermaphrodites and the linker cell in males (Figure S2A and S2B).
We attempted to understand the trans-regulation of vulval expression in the diverse cell types by analyzing the regulation of two target genes in detail: egl-17 and ceh-2. To dissect the trans-regulation of these target genes, we constructed various double and triple mutant/RNAi combinations and assayed for alterations in gene expression in the 1° vulval cells.
During the L4 stage, the egl-17 transcriptional reporter is expressed solely in vulC and vulD, being absent in both vulE and vulF (Figure 5 and Table 1). nhr-67 RNAi in an otherwise wild-type background results in an increase of egl-17 expression in the vulF cells (Figure 5 and Table 1). In those nhr-67 RNAi animals, only one of the four vulF cells exhibits this ectopic egl-17 expression during the L4 stage. egl-17 expression is consistently absent in the vulF cells of cog-1 and egl-38 hypomorphic alleles (Figure 5 and Table 1). In comparison, cog-1 animals treated with nhr-67 RNAi are qualitatively enhanced (i.e., several vulF cells misexpress egl-17), whereas egl-38 animals treated with nhr-67 RNAi displayed a qualitatively and quantitatively higher egl-17 expression in the vulF cells (Figure 5 and Table 1). cog-1 is necessary for negatively regulating egl-17 expression in the vulE cells and acts redundantly with egl-38 to negatively regulate egl-17 in the vulF cells [19] (Figure 5 and Table 1). We also observed frequent egl-17 upregulation in the vulE cells of egl-38; nhr-67 (RNAi) doubly perturbed hermaphrodites (Table 1), which is invariably absent in either singly perturbed background. Our study provides the first example of egl-38 modulating gene expression in the vulE cell type. Hence, egl-38, nhr-67, and cog-1 act together to negatively regulate egl-17 expression in the 1° vulval lineages during the L4 stage. Loss of lin-11 function leads to complete abolition of egl-17 gene expression in all vulval cells [26] (Figure 5 and Table 1). Lastly, the ectopic egl-17 expression visualized in the 1° descendents of cog-1-, egl-38-, and nhr-67-depleted backgrounds is dependent on lin-11 activity (Figure 5 and Table 1). Loss of nhr-67 in combination with lin-11 yields rare egl-17 expression in apparently random vulval cell types (∼4% of animals).
In wild-type L4 hermaphrodites, ceh-2::YFP expression is only observed in the vulB cells and is invariably absent in both vulE and vulF cells. nhr-67 RNAi results in a moderate frequency of ectopic ceh-2 expression in the vulE and vulF cells (Table 2). Eliminating lin-11 function leads to complete loss of ectopic ceh-2 expression in the 1° vulval lineages of nhr-67 RNAi animals (Table 2). ceh-2 expression is consistently absent in the 1° vulF cells of cog-1 and egl-38 single mutants (Table 2). cog-1 mutants exhibit a moderate increase of ceh-2 expression in the vulE cells [19] (Table 2). We also found that 90% of cog-1; egl-38 doubles show increased ceh-2 expression in vulE cells compared to cog-1 (32%) or egl-38 (0%) single mutants (Table 2). Thus, analysis of these double mutants provides us with a second example of egl-38 regulating gene expression in the vulE cells. As with the egl-17 reporter, simultaneous depletion of cog-1 and egl-38 activities results in a high frequency of ceh-2 misexpression in the vulF cells (Table 2). Both cog-1 and egl-38 are thus required for negative regulation of ceh-2 expression in the vulF cells.
cog-1, lin-11, and nhr-67, all of which regulate different aspects of vulval gene expression, exhibit dynamic spatial and temporal expression patterns in the developing vulva [26,27]. egl-38 expression has been observed in the vulF cells [15]. As mentioned previously, nhr-67 expression is primarily restricted to vulA (and occasionally vulB1) cells during L4 stage. Yet numerous perturbations in gene expression are observed in nhr-67 RNAi–treated animals, suggesting that nhr-67 is indeed functional during the L4 stage in other mature vulval cell types besides vulA (Figure 1). A similar observation can be made about cog-1. Wild-type animals occasionally exhibit weak cog-1 expression in vulE cells but none in vulF cells (Table 3). However, cog-1 synergistically interacts with egl-38 and nhr-67 to regulate egl-17 expression in the vulF cells (Figure 5 and Table 1). One attractive hypothesis is that levels of both these transcription factors are maintained under strict spatio-temporal control. We thus set out to investigate the interactions among these regulatory factors by assaying for alterations in the reporter gene expression in various mutant backgrounds.
During the L4 stage, lin-11 is consistently expressed in the 2° vulB, vulC, and vulD lineages, and occasionally in the vulA and vulF cells. Neither cog-1 nor egl-38 mutations alter lin-11 vulval expression [36]. Similarly, reduction of nhr-67 gene activity also does not impact lin-11 expression in the vulva (Table 3).
The cog-1 translational reporter is strongly expressed in vulC and vulD, weakly expressed in vulE, and undetectable in vulF cells during L4 (Figure 6 and Table 3). We found that cog-1 levels are increased in the 1° vulF cells of nhr-67 RNAi–treated hermaphrodites as well as in lin-11 and egl-38 mutants (Figure 6 and Table 3). nhr-67 RNAi–treated animals also showed elevated cog-1 expression in the vulE cells (Table 3). In lin-11 mutants, cog-1 levels in vulD are completely abolished as opposed to the vulC-specific expression, which is only partially affected (∼57% of animals) (Table 3). Overall cog-1 expression levels in lin-11 loss-of-function mutants are noticeably reduced when compared to the wild-type reporter background. The frequency of vulD-specific cog-1 expression is significantly increased in egl-38 mutants (Table 3). cog-1 negatively autoregulates in vulA, vulB1, and vulB2 cells (Table 3).
nhr-67::GFP expression is consistently observed in vulA during the L4 stage (Table 3). lin-11 mutants only partially eliminate the vulA-specific expression of nhr-67 (Figure 7 and Table 3). nhr-67 expression in vulA is completely abolished only in the absence of both lin-11 and its positive autoregulatory activity (Table 3). Overall, nhr-67 expression levels in lin-11 loss-of-function mutants are noticeably reduced when compared to a lin-11(+) background. lin-11 activity is also required for directing the ectopic nhr-67 expression in the 1° lineages when the autoregulatory loop is compromised (Table 3, see below). Also, loss of lin-11 sometimes caused premature vulC expression of nhr-67 during L4 stage, which can be interpreted either as a cell type or a temporal regulatory defect (Figure 7 and Table 3). Reduction of cog-1 function results in increased expression of nhr-67 in vulC and vulD during the L4 stage and vulE and vulF during L4 lethargus (Figure 7 and Table 3). Depletion of both cog-1 and nhr-67 activities leads to a more robust increase in nhr-67 levels in the vulF cells (Table 3). egl-38 mutants sometimes showed ectopic nhr-67 expression in vulC and vulD cells during the L4 stage (Figure 7 and Table 3) and significantly increased its frequency of expression in vulB1 cells (Table 3).
In addition to the cross-inhibitory interactions between cog-1 and nhr-67 in both the 1° vulE and vulF cells, we also discovered that they both negatively autoregulate in the same cell types. Inhibition of nhr-67 by RNAi feeding results in the robust increase of nhr-67::GFP expression levels in both vulE and vulF cells (Figure 7 and Table 3). Elevation of nhr-67 transcriptional levels is also visible in the vulC and vulD lineages of nhr-67 (RNAi) animals during L4 stage. Upregulation of nhr-67 expression in vulC, vulD, vulE, and vulF cells is also visible with the 4.5-kb nhr-67 transcriptional reporter construct (Figure 4B) in an nhr-67(RNAi) background (unpublished data). We used fos-1 RNAi feeding as a control to exclude the possibility that the observed negative autoregulation was a nonspecific effect of inducing RNAi. fos-1 RNAi–treated animals exhibited a strong Pvl phenotype (at least in part due to its AC invasion phenotype) [37] and did not alter nhr-67 levels in the 1° lineages (Figure 7 and Table 3). Similarly, ectopic expression of cog-1::GFP in all 1° vulval descendents is consistently observed when cog-1 activity is compromised (Figure 6 and Table 3). Thus, nhr-67 and cog-1 appear to be activated in all the mature vulval cell types but are then restricted by both autoregulatory and trans-regulatory mechanisms.
nhr-67 encodes a C. elegans ortholog of tailless, a crucial regulator of blastoderm patterning in the terminal pathway of Drosophila embryogenesis as well as neuronal development. We find that nhr-67 activity is required for the regulation of gene expression in several mature vulval cell types and is dynamically expressed in the vulva. For technical reasons, we have been unable to determine whether nhr-67 acts in the vulval cells for these functions. However, the expression of nhr-67 in the vulva and the complexity of the interactions are most consistent with a primarily autonomous action of nhr-67. However, given the expression of nhr-67 in the AC, it is possible that the effects (particularly on the 1° lineage) are nonautonomous. For example, the AC generates EGF and Wnt signals and is required to differentiate vulE and vulF cells, presumably via these signals [10]. Loss of vulF-specific lin-3 expression in an nhr-67 RNAi background is certainly consistent with this model. The AC also promotes 1° over 2° fate [38]. The ectopic expression of 2° lineage-specific genes ceh-2 and egl-17 in the 1° vulval cells is also consistent with this model. However, lineage analysis of nhr-67 (RNAi) hermaphrodites argues that these alterations are not full 1° to 2° cell fate transformations in the early vulval lineages. In addition, the observed effects on pax-2 and zmp-1 expression are inconsistent with this model. It remains a formal possibility that some of nhr-67 effects in the vulva are due to a role in the AC.
Our data are consistent with the function of Drosophila tailless, which facilitates proper gap gene expression at the posterior end of the blastoderm embryo via its dual activator/repressor activity [39–41]. Specifically, tailless blocks segmentation and maintains the identity of the terminal boundaries via repression of Kruppel and knirps activity and promotes hunchback expression, which is necessary for the establishment of terminal-specific structures [42,43]. tailless is also necessary for regulating gene expression during the generation of head segments as well as anterior brain development [44]. We also find that nhr-67 prohibits improper fusion events between related cell lineages, at least partly due to strict spatial regulation of the fusogen eff-1 in certain vulval cell types.
As discussed below, nhr-67 interacts genetically with three other transcriptional regulators, cog-1, egl-38, and lin-11, to produce complex patterns of gene expression, probably through trans-regulation of cell type–specific enhancers (Figure 8).
We have uncovered a novel set of genetic interactions between nhr-67, several transcription factors, and many target genes that contribute to the identity of distinct vulval cell types. For example, nhr-67 appears to be particularly important in the execution of vulF fate and maintaining its cellular identity via regulation of gene expression and fusion events between distinct cell types. Not only does nhr-67 inhibit inappropriate gene expression that is associated with the 2° vulval lineages (Figure 1), but it also promotes gene expression of the EGF protein LIN-3, which is necessary for uv1 fate specification and facilitates proper vulval-uterine connection during development [14]. The functional data obtained from numerous RNAi experiments demonstrates that nhr-67 (like its Drosophila ortholog) is a versatile regulatory gene that operates on at least four of the seven vulval cell types (vulA, vulD, vulE, and vulF). However, we have not tested whether any of these approximately ten interactions are direct.
An interesting feature of the network is our suggestion that both nhr-67 and cog-1 might negatively autoregulate in the same vulE and vulF cells. Drosophila melanogaster tailless does not regulate itself [45], suggesting that nhr-67 autoregulation is a developmental phenomenon unique to nematodes (C. elegans). This apparent divergence in tailless regulation between phyla suggests that a more precise fine-tuning of tailless levels is required for the execution of accurate patterning in the C. elegans vulva. In contrast to their different autoregulatory properties, we find that certain genetic interactions are indeed conserved between the D. melanogaster tailless and C. elegans nhr-67; namely tailless restricts the expression domain of ems in the head segments [44], which is comparable to nhr-67 repressing the worm ems ortholog ceh-2 in the inappropriate vulval cells. Additional tailless targets from other organisms [40,46,47] may also have an impact on vulval patterning. Predictions can also be made in the reciprocal direction and used to elucidate vertebrate development. For example, FGF signaling is required for both vertebrate and inverterbrate heart development [48,49]. The LIM domain protein ISL1 promotes differentiation in a subset of cardiac progenitor cells and transcriptionally activates several FGF genes in mice [50]. Our trans-regulation experiments reveal that both egl-17 and ceh-2 contain cis-regulatory elements that are directly or indirectly dependent on cog-1 (Nkx6.1/6.2), egl-38 (Pax2/5/8), nhr-67 (tll), and lin-11 (LIM) activity. These data may provide further insights into the elaborate regulation of classic developmental genes such as FGF and EMS, both of which have multiple roles in metazoan development.
Previous work demonstrated that patterning of the E and F descendents of the 1° vulval lineage involves both a short-range AC-dependent signal using the Ras pathway as well as lin-17 (Wnt) signaling [10]. In the context of egl-17 gene expression, cog-1 single mutants exhibit increased levels in the vulE cells only. In contrast, nhr-67 RNAi appears to exclusively affect egl-17 expression in the vulF cells. The negative regulatory activities of cog-1 in vulF and nhr-67 in vulE only become apparent in an egl-38 mutant background (which shows no phenotype on its own). This difference suggests that cog-1-mediated negative regulation plays a greater role in vulE cells whereas nhr-67-mediated negative regulation functions primarily in vulF cells. One hypothesis is that vulF cells are biased by proximity to the AC to have higher levels of nhr-67 compared to cog-1 (Figure 9). The genetic regulatory interactions within the vulval network demonstrate that cog-1 levels are negatively regulated in vulF cells via four inputs: lin-11, egl-38, nhr-67, and cog-1. In comparison, nhr-67 expression in vulF cells is modulated by two antagonistic inputs (cog-1 and nhr-67) and one positive input (lin-11), thus possibly resulting in its higher levels. These observations are consistent with a model where nhr-67 acts as the major negative regulator in vulF cells.
nhr-67 and cog-1 cross-inhibit each other's transcriptional activities, specifically in the vulE and vulF cells, implying that a mutually antagonistic feedback loop exists that exclusively affects the cells of the 1° vulval lineages. Both cog-1 and its mammalian ortholog Nkx6.1 have been previously implicated in bistable loops that reinforce one of two possible stable end states [51,52]. The cross-inhibitory interactions between nhr-67 and cog-1 might be relevant in the specification of vulE versus vulF cell fates. The nature of the bistable loop between cog-1 and nhr-67, however, is unknown. In particular, the bistable loop may be a consequence of either direct transcriptional regulation (as implied in Figure 9) or indirect regulation through an unknown intermediate regulatory factor.
However, the above observation does not rule out the possibility that additional regulatory factors might also contribute to proper patterning of 1° lineages. These other inputs could presumably operate via several potential mechanisms such as modulating the balance between cog-1 versus nhr-67 levels, being exclusively active in one 1° cell type, and interacting at distinct cis-regulatory elements of the downstream targets.
Given the complexity of the observed vulval regulatory interactions, we propose that the network operating on each vulval cell type is unique (Figure 9). A single regulatory factor may have differential functions in terms of executing accurate spatio-temporal gene expression in diverse cells. For instance, lin-11 may upregulate cog-1 levels in the 2° vulC and vulD cells while antagonizing them in the 1° vulF cells. A similar argument can be made about the lin-11-dependent regulation of nhr-67. lin-11 may temporally regulate nhr-67 by inhibiting its vulC-specific expression during the L4 stage. In contrast, lin-11 is clearly critical for the positive regulation of nhr-67 expression in both vulE and vulF cells.
Both cog-1 and nhr-67 are present at high levels in a subset of the 2° vulval cells, yet are barely detectable in the 1° vulval cells. Nevertheless, the disruption of either factor yields obvious defects in 1° vulval cell–specific gene expression. A cross inhibition circuit, such as we propose for cog-1 and nhr-67, can be bistable, with stable states that tolerate inherent fluxes in gene expression (i.e., it would not randomly oscillate between states) [53–55]. Negative autoregulatory circuits have been shown to reduce cell–cell fluctuations in the steady-state level of transcription factors [56] and can speed up the response times of transcription networks without incurring the cost of constant protein production and turnover [57]. These two distinct circuits might enable cells to reach a developmental state with built-in flexibility, allowing rapid switching of their fate upon transient inputs (as opposed to sustained inductive inputs that are metabolically costly). In this model, dynamic levels of cog-1 and/or nhr-67 expression could correlate with particular aspects of 1° vulval cell fate execution. This might account for the elaborate autoregulatory and trans-regulatory interactions specifically seen in 1° vulval descendents, as opposed to their 2°-derived counterparts. We postulate that although all the vulval cells appear to use the same regulatory factors, their differential effects on the diverse cell types is what results in accurate gene expression.
During the L4 stage, the gradient of nhr-67 expression is opposite to that of either cog-1 or lin-11. This difference in gene expression domain raises the question of whether the levels of these factors are critical for vulval development. For example, high levels of lin-11 result in misexpression of egl-17 in vulA and abnormal vulval invagination [26]. Different concentrations and combinatorial expression patterns of lin-11, cog-1, and nhr-67 might thus encode mature vulval cell types (Figure 9). For example, differentiation to the 1° vulF cell type may entail low levels of LIN-11 and NHR-67 along with lower levels of COG-1. In contrast, the 1° vulE cells require medium levels of COG-1 along with low doses of LIN-11 and NHR-67. vulA and vulB are similar to each other with respect to maintaining low COG-1 levels. However, vulA cells are characterized by their high NHR-67 levels and medium LIN-11 levels as opposed to the reverse situation in vulB1 and vulB2 cells (medium-low NHR-67, high LIN-11). Lastly, both vulC and vulD have indistinguishably high levels of LIN-11 and COG-1, and we are unable to precisely define what distinguishes these two cell types from each other. One hypothesis is the differential regulation of NHR-67 and COG-1 in both cell types: COG-1 levels are impacted by egl-38 in vulD (but not vulC), whereas NHR-67 levels are negatively regulated by lin-11 in vulC (but not vulD). An obvious limitation of this proposed regulatory code is that it does not take into account other transcription factors that may potentially mediate vulval patterning.
The intricacies of vulval organogenesis can be deconstructed by rigorously elucidating the genomic networks that operate within the seven mature vulval cell types. Deciphering this regulatory code will provide valuable information on network connections and might provide insights into other examples of organogenesis.
Transgenic worms were anesthetized using 3 mM levamisole and observed using Nomarski optics (http://www.nomarski.com). Photographs were taken with a monochrome Hamamatsu digital camera (http://www.hamamatsu.com) and Improvision Openlab 4.0.4 software (http://www.improvision.com). The fluorescent images were overlaid with their respective DIC images using Adobe photoshop 7.0.1 (http://www.adobe.com). The vulval expression patterns for all strains except syIs49 were visualized during the late L4 stage. In the case of syEx716, the vulval expression was also examined during L4 lethargus and adult stage. In syIs49 animals, vulA-specific zmp-1::GFP expression was scored in adults only.
C. elegans strains were cultured at 20 °C using standard protocols (Brenner, 1974). Transgenes used in this study are as follows: syIs54 [ceh-2::GFP], syIs55 [ceh-2::YFP], syIs51 [cdh-3::CFP], syIs49 [zmp-1::GFP], syIs77 [zmp-1::YFP], syIs59 [egl-17::CFP] [9], syIs78 [ajm-1::GFP] [26], syIs107 [lin-3::GFP] [58], ayIs4 [egl-17::GFP] [16], guEx64 [pax-2::GFP] (gift from Chamberlin lab), kuIs36 [egl-26::GFP] [18], syIs63 and syIs64 [cog-1::GFP] [27], syIs80 [lin-11::GFP] [59], syEx716 [8-kb nhr-67Δpes-10::GFP], syEx749 [8-kb nhr-67Δpes-10::GFP], syEx744 [nhr-67 intron4 Δpes-10::GFP], syEx925 [6 kb upstream nhr-67::GFP + 8 kb nhr-67Δpes-10::GFP], syEx865 [nhr-67p::GFP::nhr-67 int4–3′end], and syEx756 [unc-53::GFP]. Alleles used in this study: LGI, lin-11(n389); LGII, cog-1(sy275), eff-1(hy21); LGIII, unc-119(ed4); LGIV, unc-31(e169), egl-38(n578), dpy-4(e1166sd), dpy-20(e1282); LGV, him-5(e1490). A complete list of strains is included in Table S2.
Transgenic lines were generated using standard microinjection protocol that produces high-copy number extrachromosomal arrays [60]. syEx756 was generated by injecting the pNP10 construct [61] into unc-119(ed4); him-5 background using unc-119(+) [62] and pBSK+ (Stratagene, http://www.stratagene.com) as coinjection markers.
A reverse genetics screen was conducted against 508 transcription factors (Table S1) from the Ahringer library (Medical Research Council Geneservice) to assay for alterations in vulval expression patterns for the ceh-2::YFP transgene. RNAi feeding protocol is similar to that previously described [32]. Embryos were harvested by bleaching gravid adults and were placed on a lawn of Escherichia coli strain expressing double-stranded RNA at 20 °C. Animals were scored after 36 h (during the L4 stage) using Nomarski microscopy. We resorted to nhr-67 RNAi feeding for the rest of this study since the nhr-67 deletion allele (ok631) results in L1 lethality and/or arrest (International C. elegans Knockout Consortium). All subsequent nhr-67 RNAi feeding experiments were done as described above. nhr-67 RNAi feeding experiments that entailed the restriction of cell fusion (via a temperature-sensitive allele of eff-1) were conducted at 25 °C.
nhr-67::Δpes-10::GFP reporter gene constructs: The pPD97–78 vector, which includes the Δpes-10 basal promoter driving GFP and the unc-54 3′ UTR (gift from Fire lab), was used as a template to generate 2-kb Δpes-10::GFP products. The primers used for amplification are 5′-GCTTGCATGCCTGCAGGCCTTG-3′ and 5′-AAGGGCCCGTACGGCCGACTAGTAGG-3′. All nhr-67 gene fragments were amplified from the C08F8 cosmid and were stitched together with the Δpes-10::GFP fragment via PCR fusion [63] and were designated as “pdd-1 constructs.” Construct (1) consists of 1-kb promoter sequence, the entire coding region, and introns and 2 kb of the 3′ noncoding region attached to minimal Δpes-10::GFP. The primers used to amplify this template are 5′-CTGCTCAAAACTTTTGCTCC-3′ (forward) and 5′-CAAGGCCTGCAGGCATGCAAGCTTAAAGAACTACTGTAGTTTTTG-3′ (reverse). Construct (2) spans from the fourth intron to the 3′ noncoding region fused to minimal Δpes-10::GFP. This product was generated using the forward primer 5′-GTTCGATCATGGATCCTCTCC-3′ and the same reverse primer as construct (1). Construct (3) is an nhr-67p::GFP reporter that contains 1 kb of the native promoter stitched in-frame with a 700-bp coding fragment of GFP (amplified from the pPD95–69 vector, a gift from Fire lab). The resulting 1.7-kb gene product was subsequently fused to 4.5 kb of nhr-67 regulatory sequences (that span from the fourth intron to the 3′ noncoding region) via PCR. Construct (4) contains 6-kb sequence upstream of the predicted first ATG of nhr-67, appended to minimal Δpes-10::GFP. The following primers were used to amplify this product: 5′-GAACCCGGCGACGTTACGGGGCTTC-3′ and 5′-CAAGGCCTGCAGGCATGCAAGCCATCTGTGAAACCGCAGTCATCAT-3′.
Reporter constructs were injected into unc-119(ed4); him-5 worms using unc-119(+) [62] and pBSK+ (Stratagene) as coinjection markers. lin-11(n389); syEx749 doubles were constructed by injecting the 8-kb nhr-67::Δpes-10::GFP construct into lin-11(n389); unc-119(ed4); him-5 background using unc-119(+) as a rescue marker.
The WormBase Gene IDs (www.wormbase.org) as well as the Refseq accession numbers (www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=Nucleotide) for the genes described in this study are ajm-1:WBGene00000100 (NM_077135; NM_077137; NM_077136; NM_171966); cdh-3:WBGene00000395 (NM_066286); ceh-2:WBGene00000429 (NM_059345); cog-1:WBGene00000584 (NM_182115); eff-1:WBGene00001159 (NM_001026819); egl-17:WBGene00001185( NM_075706); egl-26:WBGene00001193 (NM_061251); egl-38:WBGene00001204 (NM_069435); lin-3:WBGene00002992 (NM_171418;NM_171919;NM_171918); lin-11:WBGene00003000 (NM_060295); nhr-67: WBGene00003657 (NM_069693); pax-2:WBGene00003938 (NM_068112); unc-53: WBGene00006788 (NM_001027000;NM_001026999); and zmp-1:WBGene00006987 (NM_171138). |
10.1371/journal.pntd.0000753 | Evaluation of Chikungunya Diagnostic Assays: Differences in Sensitivity of Serology Assays in Two Independent Outbreaks | The sensitivity and specificity of two in-house MAC-ELISA assays were tested and compared with the performance of commercially-available CTK lateral flow rapid test and EUROIMMUN IFA assays for the detection of anti-Chikungunya virus (CHIKV) IgM. Each MAC-ELISA assay used a whole virus-based antigen derived from genetically distinct CHIKV strains involved in two chikungunya disease outbreaks in Singapore (2008); a January outbreak strain with alanine at amino acid residue 226 of the E1 glycoprotein (CHIKV-A226) and a May-to-September outbreak strain that possessed valine at the same residue (CHIKV-226V). We report differences in IgM detection efficacy of different assays between the two outbreaks. The sensitivities of two PCR protocols were also tested.
For sera from January outbreak, the average detection threshold of CTK lateral flow test, MAC-ELISAs and EUROIMMUN IFA assays was 3.75, 4.38 and 4.88 days post fever onset respectively. In contrast, IgM detection using CTK lateral flow test was delayed to more than 7 days after fever onset in the second outbreak sera. However, MAC-ELISA using CHIKV-226V detected IgM in the second outbreak sera 3.96 days after fever onset, which was approximately one day earlier compared to the same assay using CHIKV-A226 (4.86 days). Specificity was 100% for both commercial assays, and 95.6% for the in-house MAC-ELISAs. For sensitivity determination of the PCR protocols, the probe-based real time RT-PCR method was found to be 10 times more sensitive than one based on SYBR Green.
Our findings suggested that the two strains of CHIKV using variants A226 and 226V resulted in variation in sensitivities of the assays evaluated. We postulated that the observed difference in antigen efficacy could be due to the amino acid substitution differences in viral E1 and E2 envelope proteins, especially the E1-A226V substitution. This evaluation demonstrates the importance of appraisal of different diagnostic assays before their application in clinical and operational settings.
| Chikungunya is a mounting public health concern in many parts of the world. Definitive diagnosis is critical in differentiating the diseases, especially in dengue endemic areas. There are some commercial chikungunya kits and published molecular protocols available, but no comprehensive comparative evaluation of them was performed. Using sera collected in outbreaks caused by two variants of Chikungunya virus (A226 and 226V), we tested 2 commercial IgM tests (CTK lateral flow rapid test and EUROIMMUN IFA) alongside our in-house IgM assays (using both variants of the virus). Sensitivities of 2 published PCR protocols were also evaluated based on RNA standards derived from cell-cultured viruses. The commercial assays had different performances in each outbreak, with CTK's lateral flow test having the best performance in the first outbreak and EUROIMMUN IFA being more sensitive in the second outbreak. Use of the current circulating virus in a test assay improves sensitivity of the MAC-ELISAs. For PCR, a probe-based real time RT-PCR method was found to be 10 times more sensitive than the SYBR Green method. Despite this, the latter protocol is found to be more suitable and cost-effective for our diagnostic laboratory. This evaluation demonstrates the importance of appraisal of commercial kits and published protocols before application of a diagnostic tool in the clinical and operational setting.
| Chikungunya virus (CHIKV) has seen a resurgence in recent years, with outbreaks being described in Republic of Congo in 2000, La R'eunion in 2005, India, Sri Lanka, Malaysia and Gabon in 2006, Italy in 2007, Singapore and Thailand in 2008 [1], [2], [3], [4], [5], [6], [7], [8], [9]. The current pandemic involves a newer CHIKV strain of the East-Central South African (ECSA) genotype. Extensive research and analysis demonstrated the role of a viral mutation, A226V, in the changed epidemiology of the disease. There is no evidence that this particular mutation caused any alteration in virulence of the CHIKV or clinical manifestations of the disease, but the mutation, residing in the viral envelop protein, has been shown to facilitate enhanced transmissibility of the virus by Aedes (Ae.) albopictus. Several sophisticated studies have established that the A226V mutation rendered higher viral replication and dissemination rates in Ae. albopictus, and thus shortening the extrinsic incubation period in the vector [10], [11]. The length of extrinsic incubation period determines the infective life span of a vector, and consequently has great influence on the epidemic potential of the virus-vector partnership.
Since 2006, in response to the outbreaks in the region, the Environmental Health Institute (EHI), a national public health laboratory in Singapore, has initiated laboratory surveillance for CHIKV. Two main outbreaks were detected: the first occurring in January 2008 was a small outbreak with 13 local cases [12], [13]; and the second commenced in May 2008 and peaked two months later, resulting in 231 local cases by the end of September 2009 [13].
Phylogenetic analysis concluded that viruses isolated from these two outbreaks were related to the ECSA genotype [13]. Interestingly, the viruses from the first outbreak showed alanine at amino acid residue 226 (A226) of E1 gene and those from the second outbreak showed valine (226V) at the same codon. While Ae. aegypti was the implicated vector in the first outbreak, Ae. albopictus was the confirmed vector of the second outbreak [13]. Though CHIKV was not isolated from any field caught Ae. aegypti during the first outbreak, entomological investigations in the affected area found only Ae. aegypti adults, and data from routine surveillance (part of Singapore dengue control programme) also showed that Ae. aegypti was the predominant species in the area. On the other hand, CHIKV was isolated from Ae. albopictus caught during the second outbreak.
Laboratory confirmation of CHIKV infection is critical, especially in dengue endemic areas, as clinical symptoms of the two diseases are similar. However the two viruses may be transmitted by different vectors (Ae. aegypti and Ae. albopictus), which require different control strategies. RT-PCR is an excellent tool for the early phase confirmation of CHIKV infections, and many protocols have been established for this purpose [13], [14], [15], [16], [17], [18], [19], [20], [21]. Unfortunately, this viral detection method is limited to the viraemic phase, which is usually one to five days after fever onset. Thereafter, confirmation of CHIKV infection requires serological tests. In recent years, a few commercial CHIKV diagnostic kits have emerged in the market, but there are very few reports on the systematic and comparative evaluation of these commercial products. The most recent CHIKV diagnostic assay on the market is the indirect immunofluorescence assay (IFA) from EUROIMMUN AG (Lübeck, Germany), whose IgM assay presented a specificity of 98.3% and a sensitivity of 96.9% [22]. This assay, together with an IgM lateral flow rapid test kit by CTK Biotech Inc (San Diego, USA), was evaluated alongside an in-house IgM Capture Enzyme-Linked Immunosorbant Assay (MAC-ELISA). Sensitivities, specificities and approximate time antibodies first became detectable in an infected patient were determined. A comparison between the sensitivity of two PCR protocols was also performed using RNA standards derived from cell-cultured viruses. The evaluation, using samples from two independent and epidemiologically distinct CHIKV outbreaks in Singapore, was done to establish diagnostic capability in the laboratory. The IgM titres were not determined, and thus the kinetics of the antibody has not been established. However, the availability of a reliable assay allows antibodies to be titred for each sample, and thus facilitating an ongoing antibody kinetics study, that will be reported later.
Environmental Health Institute is a public health laboratory that functions as a licensed diagnostic laboratory, with an ISO9001 accreditation. It served as the national reference laboratory during the CHIKV outbreaks in 2008. The three diagnostic techniques were evaluated using three characterized panels of sera.
Persons with an acute febrile illness, signs or symptoms compatible with chikungunya fever (fever, joint pain, or rash) were tested with CHIKV RT-PCR [12], a routine test which has been offered under EHI's quality assured programme as required for the national license. Sera panel A and B were multiple consecutive samples collected from RT-PCR CHIKV confirmed patients, during the first and second CHIKV outbreaks respectively (See supporting information “Supporting Data S1”). The patients were warded at the Tan Tock Seng Hospital Communicable Disease Centre (TTSH CDC). The first samples were collected on the day of first medical consultation and subsequently, more samples were collected as the disease progressed, till convalescence. The daily samples were used to determine the sensitivity of IgM serology on each day of illness.
Panel A, comprising residual blood from eight CHIKV-confirmed patients from the first outbreak (January 2008), were collected for clinical management and to determine when the patient could be discharged. Six to 11 samples were collected from each patient, resulting in a total of 60 samples (See supporting information “Supporting Figure S1”).
Panel B, from the second outbreak (May to September 2008) was collected from 28 CHIKV-confirmed patients in August 2008 prospectively. Each patient had five to 12 samples collected, leading to a total of 225 samples.
All sera samples were kept at 4°C after phlebotomy, transported on ice, and reached the laboratory within 24 hours. The samples were either tested on the same day, or stored at −80°C until testing.
Panel C sera were used for specificity tests and consisted of 45 flavivirus-confirmed sera (44 Dengue, one Japanese Encephalitis) and five non-CHIKV alphavirus-confirmed sera (two Barmah Forest, three Ross River).
Analysts of the serology tests were not blinded to the RT-PCR results of the first samples. However, they were blinded to the serology results derived from the other tests.
Panels A and C were residual samples of sera sent to EHI for diagnosis. Use of residual samples for evaluation of diagnostic assays to establish in-house capability is exempted from internal review by the National Environment Agency Bioethics Review Committee. Use of sera panel B, which was collected for a larger study, was approved by the National Healthcare Group Domain Specific Review Board, and written informed consent was obtained from the study participants.
To determine the sensitivities of PCR protocols, two previously published protocols were tested. The first was a one-step SYBR Green based RT-PCR, from Hasebe et al. 2002, for the detection of a fragment of the non-structural protein 1 (nsP1) gene of CHIKV [15]. The PCR conditions were described in Hapuarachchi et al. 2010. The other assay was a taqman probe-based RT-PCR protocol adapted from Pastorino et al. 2005, with slight modifications to the primers which target the E1 region (Table 1) and included following modifications to the PCR assay: PCR assay was performed using the Qiagen QuantiTect Probe RT-PCR kit, in a final reaction volume of 20 µl containing 5 µl of template, 1× of buffer mix, 0.2 µl of RT enzyme mix, 0.25 µM of probe, 0.25 µM and 0.4 µM of forward and reverse primers respectively. The amplification cycles were extended to 50 with denaturation at 94°C for 5 sec and annealing/extension step at 60°C for 1 min. The analytical sensitivity and reproducibility of both assays were determined using 10-fold dilutions of cultured CHIKV strains D67Y08 (A226) and D1225Y08 (226V) (2.7×10−1 to 2.7×108 pfu/ml). CHIKV RNA was extracted from the dilutions using QIAamp viral RNA mini kit (Qiagen, Hilden, Germany).
A total of three analysts were involved in the evaluations. The two commercial IgM assays were performed by one analyst, and the in-house ELISA assays were performed by another. The third analyst performed the PCR sensitivity tests. All analysts were trained in-house, certified by the Director of the diagnostic laboratory, and regularly passed the external (RCPA) and internal proficiency tests, under the EHI quality assurance programme.
The sensitivity and specificity of assays were calculated in Microsoft Excel 2007. ANOVA, to test for variance amongst results obtained by the serological assays, and Student t-tests to determine any significant differences between the different assays were calculated using SPSS 13.0 software.
The commercial lateral flow rapid test and IFA, along with in-house MAC-ELISAs using both D67Y08 (226A) and D1225Y08 (226V), were evaluated using three panels of sera. Sera panels A and B were collected during the two outbreaks in January and May to September in 2008 respectively, and were from CHIKV RT-PCR-confirmed patients (See supporting information “Supporting Data S1”). Average IgM detection threshold, according to day after fever onset was determined. During the first outbreak, the lateral flow (CTK) kit enabled the detection of IgM on an average of 3.75 days after fever onset. IFA (EUROIMMUN) and in-house MAC-ELISA detected IgM on 4.88 day and 4.38 day respectively. The use of CHIKV-A226 or -226V did not alter the effectiveness of the in-house ELISA assays (Table 2) during the January outbreak.
Though CTK's lateral flow assay had the best performance in the January outbreak, its performance was not repeated in the second outbreak. Among the first 10 CHIKV- confirmed patient sera (total of 74 samples from Panel B), none had detectable IgM within seven days after fever onset. The earliest IgM detection attained by the lateral flow assay, was day nine after the onset of fever (n = 1) and eight of the 10 patients did not show seroconversion even after 14 days. To investigate if the drop in performance was due to batch variability of the CTK kit, 30 CHIKV-A226 IgM positive and 10 CHIKV negative samples from the first outbreak were retested with the second batch of CTK kits. No difference in results interpretation was observed. As the ineffectiveness of the kit was clearly demonstrated by the 74 samples from the 10 patients, and both batches of kits were no longer available, the rest of the samples from the second outbreak (151 samples from 18 patients) were not tested with the CTK assay.
Using IFA (EUROIMMUN) and MAC-ELISA (A226) on the panel collected from the second outbreak, the average day of IgM detection was 4.86 after fever onset. Interestingly, the use of CHIKV-226V as antigen in MAC-ELISA increased the sensitivity to 3.96 day after onset of fever (p<0.0001).
Overall, the sensitivity of assays increased along with the progression of the disease (Figure 1). Sensitivities of all assays were very low from day zero to day four of the disease, ranging from 0 to 66.7% (Figure 1a). Sensitivity improved from day five, when MAC-ELISA (226V) fared the best at 93.94%; followed by MAC-ELISA (A226) at 84.85%; and IFA (EUROIMMUN) at 75.76%. Lateral flow (CTK) remained insensitive at 12.12%. By the sixth day, 100% sensitivity was attained by MAC-ELISA (226V), and by day seven, MAC-ELISA (A226) and IFA (EUROIMMUN) also achieved 100%.
In view of the possible variation in sensitivity between the two outbreaks, the samples from the CHIKV-A226 and CHIK-226V outbreaks were analysed separately (Figure 1b). In the first outbreak involving CHIKV-A226, lateral flow (CTK) test was positive in two out of five samples that were collected one day after fever onset. The sensitivity steadily increased to 100% by seven days. The other three assays started to register sensitivities greater than 50% on day five after fever onset.
In the second outbreak, involving CHIKV-226V, MAC-ELISA (226V) had the highest sensitivity of 75% at four days after fever onset and increased to 100% by day six (Figure 1c). IFA (EUROIMMUN) and MAC-ELISA (A226) detected less than 50% of samples on day four, and attained 100% only on day seven. It appeared that in the second outbreak, MAC-ELISA (226V), which utilized the virus isolated in the same outbreak, was the most sensitive.
Panel C comprising of non-CHIKV sera was utilized to determine the specificity that turned out to be 100% for both commercial assays. The sensitivity of in-house MAC-ELISA was 95.6%. The latter picked up two dengue confirmed sera which were paired samples from a single patient collected 14 days apart. Though MAC-ELISA yielded positive results with the two samples, no increase in titre was observed and negative CHIKV PRNT results were obtained (data not shown).
Two real-time PCR protocols were evaluated on RNAs extracted from a serial dilution of each of the cell-cultured viruses, D67Y08 (A226) and D1225Y08 (226V), isolated from the two outbreaks. As relative sensitivity of PCR protocols can be validated using RNA from isolated virus, no patient samples were used for this purpose. The sensitivities of each protocol for both viruses were equivalent. However, the sensitivity of the taqman probe-based protocol was found to be 10 times (2.17×100 pfu/ml) more than the SYBR Green assay (2.17×101 pfu/ml) (Table 3). The inverse relationship of Cp values and virus titres showed very good linear correlation between the detection of CHIKV-A226 and -226V when either assays were used (Figure 2).
As dengue and chikungunya infections elicit similar symptoms and can be present in the same locations, clinical differentiation may be difficult. In Singapore, it was found that the major chikungunya outbreak in the second half of 2008 was transmitted by Ae. albopictus, an outdoor mosquito. Control of chikungunya fever was thus different from the strategy employed for control of dengue fever, which is transmitted by Ae. aegypti, a peri-domesticated mosquito. It is thus important to ascertain the cause of a cluster of febrile illness. This study was carried out to ensure that accurate and robust diagnostic tools were used to diagnose chikungunya fever in Singapore.
Our findings suggested that two variants of CHIKV, A226 and 226V, had resulted in variation in sensitivities of the assays evaluated. Though CTK's lateral flow rapid test was found to be a reliable kit in the first outbreak (January 2008), it was ineffective in the second ((May to September 2008). Retesting a panel of CHIKV characterized samples from the first outbreak showed that the inconsistency was not a result of batch variation. We postulated that the inconsistency may be due to the different CHIKV variants involved in two outbreaks. The CTK rapid test kit used a recombinant antigen covering the 226 residue of E1 gene derived from the CHIKV-A226 [personal comm.], which was similar to the strain involved in the first outbreak. It is thus suggested that CHIKV-A226 derived recombinant antigen was specific for recognition by antibodies elicited by the A226 virus circulating during the first outbreak, but not for those elicited against the 226V virus of the second outbreak. Therefore, it is highly likely that the sensitivity of the rapid test can be improved by including the recombinant antigen derived from the variant virus.
The variation in sensitivity of an assay due to different antigens used was also demonstrated by our in-house MAC-ELISA, where CHIKV-226V antigens yielded higher sensitivity than CHIKV-A226, when tested on sera obtained from patients infected with the CHIKV-226V strain. Similarly, the decrease in sensitivity of the EUROIMMUN IFA was probably attributed to the use of CHIKV-A226 [22], [24]. The less striking sensitivity differences in these assays may be attributed to the use of whole viruses that offer more epitopes for recognition. Interestingly, both A226 and 226V viral strains offered the same sensitivity among samples collected from the first outbreak (A226). Notwithstanding the latter observation, our results indicated that sensitivity of a test could be improved by using the circulating virus isolated during a particular outbreak. It is unlikely that the difference in sensitivity was due to differences in quality of the antigen, as a single batch of antigen was prepared from each virus and yet, sera from the two outbreaks were giving different results for the A226 antigen (but not for the 226V antigen). Both antigens were prepared in the same way, and all sera were tested in one experiment using the same reagents and controls. Our results suggest the disadvantage of using a recombinant antigen that is too specific.
The commercial assays displayed excellent specificity, but the in-house ELISAs picked up two paired dengue IgM samples. These samples did not demonstrate increase in CHIKV IgM titres, and were CHIKV PRNT negative. As such, these were not Dengue and Chikungunya co-infected samples, rather a false positive due to non-specific IgM reactions. In our experience with dengue diagnosis, this phenomenon is not unknown, especially among adults who suffer from conditions such as Systemic Lupus Erythematosus or other immunological conditions. Though the phenomenon is poorly understood, we believe that false positives in IgM serology could be due to other immunological factors, and this may also be the case for CHIKV infection. The cross reaction may not be due to Dengue virus cross- reacting with the CHIKV assay. Our investigation and analysis in Singapore using geographical information system had also revealed that Singapore's major outbreak, due to CHIKV-226V and Aedes albopictus, did not overlap spatially with dengue fever, which is transmitted by Ae. aegypti. The likelihood of co-infection in Singapore was thus assessed to be very low.
A comparison of the amino acid sequences of the non-structural and structural polyproteins of D67Y08 (A226) and D1225Y08 (226V) isolates revealed 5 amino acid substitutions in the non-structural polyprotein and 4 in the structural polyprotein (Table 4). The A226V substitution was the only variation in the E1 envelope protein and the remaining 3 amino acid substitutions in the structural polyprotein were in E2 envelope protein. At the same time, among the 162 epitopes predicted by the Kolaskar & Tongaonkar algorithm in JEMBOSS (ANTIGENIC) version 1.5 [25], only 2 epitopes coincided with the amino acid differences observed in the structural polyprotein between DS67Y08 (CHIKV-A226) and DS1225Y08 (CHIKV-226V) isolates. Those amino acid substitutions were at residues 677 (E2 envelope protein) and 1035 (A226V in E1 envelope protein) of the structural polyprotein. Though the epitope was similar for both isolates at residue 677, the programme predicted different configurations for the two viral variants at residue 1035 (A226V). CHIKV-226V had a single linear, 34 amino acids long (amino acid positions 1019–1052) epitope, while CHIKV-A226 had 2 short epitopes flanking the same region: a 15 amino acid epitope (amino acid positions 1019–1033) and a 18 amino acid epitope (amino acid positions 1035–1052). Based on these observations, we hypothesized that the structural differences due to the A226V variation may have resulted in 226V antigen being more specific to the paratope of IgM of sera infected with the variant (226V) virus. In the absence of other recognition epitopes, a recombinant E1 antigen with A226 could have much reduced affinity to antibodies produced against the 226V variant, thus rendering a test that relies on an inappropriate A226 E1 gene ineffective during an outbreak involving CHIKV-226V strain.
Using whole virus as antigen (as in the case of MAC-ELISA and EUROIMMUN IFA) offers more antibody recognition sites. As a result, the difference in sensitivity affected by CHIKV-A226 and 226V as antigens was not very prominent. However, as E1 and E2 envelope proteins exist as a heterodimer on the alphavirus surface, contribution of amino acid substitutions in the E2 protein to the observed differences between two antigens could not be underestimated and will be of future interest.
The sensitivity of each RT-PCR protocol was not altered by the virus used. However, the probe-based PCR protocol was at least 10 times more sensitive than the SYBR Green assay. Nevertheless, the SYBR Green method was maintained as the routine test at EHI, due to the following considerations: 1) the cost of the SYBR Green assay was half of that of the probe-based; 2) the SYBR Green assay took 30 minutes, while the probe-based assay required 2.5 hours; and 3) the SYBR Green assay was sensitive enough for routine diagnosis of acute cases. Our previous study has shown that the SYBR Green method was able to detect viral RNA after resolution of fever in 30% of cases [12]. The method also detected three asymptomatic viraemic cases, one day prior to their onset of fever [13]. Taken together, it was concluded that the SYBR method was a cost effective tool for the diagnosis and surveillance of chikungunya fever.
The kinetics of viraemia in patient samples were previously examined and high levels of viraemia were observed during the first 5 days of illness [12]. Combining previous molecular findings [12] and current serology findings, medical practitioners in Singapore have been encouraged to request for PCR-based assays for patients who present within five days of fever and IgM assays for those with fever for more than 5 days. We have found that both EUROIMMUN IFA and MAC-ELISA assays were suitable for outbreaks involving both A226 and 226V variant viruses. For IgM test, MAC-ELISA has the advantage of being cost effective and easy to perform, whereas commercial EUROIMMUN IFA is suitable in laboratories with limited capacity for setting up in-house ELISA systems. An improved rapid test would benefit the community too. For PCR, the SYBR Green protocol was cost effective for the diagnosis of acute patients. This study demonstrates the importance of evaluation of commercial kits and published protocols before application of a diagnostic tool in the clinical and operational settings. With a cost effective and reliable in-house ELISA assay, as demonstrated in this study, the time course of IgM in CHIKV infected individuals is currently being investigated.
|
10.1371/journal.pntd.0002893 | Filarial Excretory-Secretory Products Induce Human Monocytes to Produce Lymphangiogenic Mediators | The nematodes Wuchereria bancrofti and Brugia spp. infect over 120 million people worldwide, causing lymphedema, elephantiasis and hydrocele, collectively known as lymphatic filariasis. Most infected individuals appear to be asymptomatic, but many exhibit sub-clinical manifestations including the lymphangiectasia that likely contributes to the development of lymphedema and elephantiasis. As adult worm excretory-secretory products (ES) do not directly activate lymphatic endothelial cells (LEC), we investigated the role of monocyte/macrophage-derived soluble factors in the development of filarial lymphatic pathology. We analyzed the production of IL-8, IL-6 and VEGF-A by peripheral blood mononuclear cells (PBMC) from naïve donors following stimulation with filarial ES products. ES-stimulated PBMCs produced significantly more IL-8, IL-6 and VEGF-A compared to cells cultured in medium alone; CD14+ monocytes appear to be the primary producers of IL-8 and VEGF-A, but not IL-6. Furthermore, IL-8, IL-6 and VEGF-A induced in vitro tubule formation in LEC Matrigel cultures. Matrigel plugs supplemented with IL-8, IL-6, VEGF-A, or with supernatants from ES-stimulated PBMCs and implanted in vivo stimulated lymphangiogenesis. Collectively, these data support the hypothesis that monocytes/macrophages exposed to filarial ES products may modulate lymphatic function through the secretion of soluble factors that stimulate the vessel growth associated with the pathogenesis of filarial disease.
| Lymphatic filariasis is caused by parasitic worms with approximately 120 million people infected worldwide and over 1 billion people at risk. The adult worms reside in host lymphatic vessels (LV) but most infected individuals do not present with overt clinical symptoms. Individuals exhibiting lymphedema, a common form of the disease, are often antigen negative; however, infected individuals, though often asymptomatic, have dilated LVs suggesting that early damage to the lymphatic architecture may lead to lymphedema in these infected individuals. In the LVs, adult worms release excretory-secretory (ES) products. Filarial ES products do not directly activate lymphatic endothelial cells (LEC), so we hypothesized that accessory cells may activate LECs indirectly and contribute to the development of disease. Here, we show that adult filarial ES products induce human blood cells, specifically monocytes, to produce lymphangiogenic factors such as IL-8 and VEGF-A and that these factors induce the formation of LVs in vivo. These results support a role for filarial ES products in altering the lymphatic architecture in filarial-infected individuals and this may contribute to LV pathology and the development of lymphedema.
| Lymphatic vessels (LVs) are important components of a system vital to the body's maintenance that includes immune surveillance and fat absorption; the primary function of these vessels is to drain excess interstitial fluids and to prevent tissue swelling [1]. Lymphangiectasia is a condition in which LVs are abnormally dilated and this pathology is often associated with the development of lymphedema, when lymphatic fluid becomes stagnant and leaks back into the surrounding interstitium. Lymphatic dilation may result from a variety of causes including trauma, cancer-related treatment regimes such as lymphadenectomy, and genetic mutations in FOXC2 or VEGFR-3. However, the majority of lymphatic pathology seen worldwide is associated with the filarial nematode parasites, Wuchereria bancrofti and Brugia malayi which cause lymphedema in millions of individuals.
An estimated 120 million people worldwide are infected by filarial parasites [2]. Lymphatic filariasis is an infection with varying degrees of clinical disease, where infected individuals can exhibit overt clinical symptoms such as lymphedema and hydrocele or be asymptomatic yet with microfilaremia. Although these asymptomatic microfilaremic individuals do not display any overt clinical manifestations, they do present with hidden sub-clinical complications [2], [3] such as dilated and tortuous lymphatics [4], [5], and scrotal lymphangiectasia in men [6], [7]. Ultrasonographic examination of the scrotal region of 14 asymptomatic Brazilians revealed that 50% of microfilaremic individuals demonstrated lymphatic dilation and tortuosity [8]. In microfilaremic individuals, abnormal lymphatics are present in 69% of limbs by static lymphoscintigraphy and in 100% of limbs by dynamic flow lymphoscintigraphy, which are sensitive indicators of lymphatic dysfunction [4], [5], [9]. In addition, studies of superficial skin punch biopsies have revealed that 78% and 68% of limbs from patients with clinical disease and asymptomatic microfilaremia, respectively, contained LVs that were abnormally dilated [5], [10]. More recently, it was also demonstrated that children as young as three years of age can present with lymphangiectasia as measured by lymphoscintigraphy suggesting that sub-clinical pathology can occur at a very early age [11].
The causes for the lymphatic dilation in filarial-infected individuals remain unknown, but lymphangiectasia is seen in SCID mice infected with Brugia suggesting that the worm and/or innate mechanisms, and not the host's adaptive immune system, are involved in the development of lymphatic dilation [12], [13]. Furthermore, the dilation can be reversed in nude mice by removing or killing the adult worms [14], [15]. An important finding was made by Shenoy et al. who showed that there is a reduction in lymphatic dilation following worm death induced by DEC treatment [11], [16].
The molecules involved in the proliferation and maintenance of endothelial cells (EC) are a family of growth factors known as the vascular endothelial growth factors (VEGF) as well as cytokines such as IL-3, IL-6, IL-7 and IL-8 [17]–[25]. VEGF-A, VEGF-C and VEGF-D, and their corresponding receptors, have all been shown to support lymphatic endothelial cell (LEC) proliferation, migration, survival and tubule formation; thus these molecules are potent regulators of lymphangiogenesis [17], [26]. Several studies have shown that plasma levels of these lymphangiogenic factors, including VEGF-A, VEGF-C, VEGF-D and angiopoietins, are significantly elevated in filarial-infected individuals including those with filarial lymphedema compared to endemic normal control subjects [27], [28]. Elevated plasma levels of VEGF-A were also seen in individuals with hydrocele [29]. Furthermore, human infection with the filarid, Onchocerca volvulus, induces lymphangiogenesis in parasite-containing nodules [30] and this neovascularization is associated with the expression of lymphangiogenic molecules such as VEGF-C [31].
Monocytes/macrophages appear to be the predominant producers of the VEGFs and the presence of monocytes/macrophages has been correlated with lymphangiogenesis [32]–[36]. In human onchocercal nodules, some mononuclear cells expressed both the macrophage marker, MAC-1, and the lymphatic-specific marker, LYVE-1, and these double-positive cells were integrated into the lymphatic endothelium [30], [31]. Thus in this present study we have addressed the role of monocytes/macrophages contributing to the production of lymphangiogenic mediators in response to filarial ES and their influence on lymphatic ECs.
For human studies, informed consent was obtained from all human subjects and approved by the Institutional Review Board at CDC. For animal studies August rats, imported from the MRC London, UK and bred locally at Western Michigan University, and maintained in standard animal laboratory housing conditions, were used. All animals were anesthetized using isoflurane gaseous equipment (Summit Medical Equipment Company, Foster City, CA). The animals were housed individually for the course of the experiment in the Animal Facilities of Western Michigan University. All the animal procedures were approved by the Western Michigan University Animal Use and Care Committee (IACUC) under project 10-01-07 before the project was begun. The study conformed to the Guide for the Care and Use of Research Animals published by the National Research Council.
Brugia malayi adult female worms were collected from the peritoneal cavity of infected jirds, Meriones unguiculatus, obtained from the NIAID Filariasis Research Reagent Repository at the University of Georgia (Athens, GA). Worms were isolated 4–12 months post infection from jirds and some of the adult females will have been gravid at this point. For the collection of ES products, 50 live adult female worms were cultured in vitro for 7 days at 37°C in 10 mL serum-free RPMI 1640 media (GIBCO) supplemented with 2 mM L-glutamine and antibiotics (100 U/mL penicillin and 100 µg/mL streptomycin). Supernatants were collected and fresh medium added daily. Supernatants containing the ES products were centrifuged at 1000× g for 10 min to remove the microfilariae and the microfilariae were resuspended in PBS and counted to ensure worm viability. Supernatants were then concentrated with a Centricon filter (Millipore, Bedford, MA) to a volume of ∼300 µL. This process resulted in ∼670 ng/mL of worm protein. ES products were stored at 4°C until further use. Male worms were not used because they do not secrete the same quantity of protein material as females (unpublished observations). Prior to cell stimulations with ES products, ES products were filtered using 0.45 µm Millex-HA syringe filters (Millipore, Carrigtwohill, Ireland) and used in a dose-dependent (diluted at 1∶10, 1∶50) manner across various replicates and batches. A batch is defined as a specimen containing the concentrated ES products from 50 female worms over one week pooled together. All batches of ES products were tested for endotoxin activity using the Limulus Amebocyte Lysate QCL-1000 assay (Lonza, Walkersville, MD) and ES products were only used for experiments when endotoxin concentrations were ≤0.1 EU/mL.
Human PBMCs were isolated using lymphocyte separation media (MP Biomedicals, Solon, OH) as directed by the manufacturer. In brief, blood was collected from normal healthy donors by venipuncture in 10 mL EDTA Vacutainer tubes (Becton Dickinson, Franklin Lakes, NJ). After centrifugation the buffy coat was removed, resuspended in RPMI 1640 media supplemented with 10% FBS (Atlas Biologicals, Fort Collins, CO), 2 mM L-glutamine and antibiotics and layered over lymphocyte separation media. Cells were centrifuged for 30 min at 1000× g at 4°C, the buffy coat was removed, washed and cells were counted using a hemocytometer.
Human CD14+ monocytes were enriched by positive selection from PBMCs using CD14+ MACS technology (Miltenyi Biotec, Auburn, CA) as directed by manufacturer. CD14+ monocyte isolation was confirmed by flow cytometry using mouse anti-human CD14+ PE (BD Pharmingen, San Jose, CA) and CD14+ cells were routinely enriched to a purity of 94–98%.
Human dermal lymphatic microvascular endothelial cells (HMVEC-dLy) were purchased from Clonetics (Lonza) and maintained in EBM-2 basal media supplemented with EGM-2 MV SingleQuots (Lonza) according to manufacturer's instructions. Cells were used from passages 4-8.
Cells were plated at 1×106 PBMCs or 5×105 CD14+ cells in 500 µL RPMI 1640 media supplemented with 10% FBS, 2 mM L-glutamine and antibiotics and stimulated with or without 100 ng/mL LPS or ES diluted at 1∶10 and 1∶50 for 72 h. The final concentration of the ES in the dilutions used to stimulate the human cells was approximately 10 to 67 ng/mL. Cell culture supernatants were collected and analyzed for IL-3, IL-6, IL-7, IL-8 and VEGF-A using the Bio-Plex Pro multiplex suspension array system (Bio-Rad, Hercules, CA) according to the manufacturer's instructions. Data was obtained using low PMT voltage settings and analyzed by the Bio-Plex Manager software version 4.1.1 and concentrations were calculated based on a standard curve derived from a recombinant cytokine standard. If the cytokine level in the sample was higher than the highest value on the standard curve, which occurred in many of the LPS stimulations, the highest value of the standard curve was reported for that data point. All samples were stimulated in parallel with ES products diluted at 1∶10 and 1∶50, but only results from the ES concentration which generated optimal stimulation were reported. VEGF-C and VEGF-D production were analyzed by the Quantikine Immunoassay kits (R&D, Minneapolis, MN) as directed by the manufacturer.
LECs were released from the flask by gentle trypsinization (Lonza), washed, counted and 1×105 LECs were stimulated in 200 µL EGM-2 MV SingleQuot media devoid of VEGF and spiked with 10 ng/mL IL-6 (R&D), 10 ng/mL IL-8 (Sigma, St. Louis, MO) or 1 ng/mL VEGF-A (R&D) for 10 min at 37°C before seeding. Cells were plated onto 100 µL Growth Factor-reduced Matrigel Matrix (BD Biosciences, Bedford, MA) coating a 24 well plate using the thin gel method as per manufacturer's instructions. After 24 h, 5 randomized fields per well were photographed at 5× magnification on a Zeiss AxioVert 200M microscope (Carl Zeiss, Thornwood, NY). The images were opened and analyzed in AxioVision release 4.7.2. At a scaling ratio of 1∶1 image analysis was performed; the total number of tubules was counted and the length of each tubule measured.
These experiments were carried out in August rats as previous work using this strain of rat demonstrated that the most appropriate time to sample for dermal vascular growth is 9 days [37]. Rat carrier-free recombinant proteins including IL-8, IL-6, VEGF164 were purchased from R&D Systems. Growth Factor-reduced Matrigel was injected into rats with or without 10 ng/mL IL-8, 10 ng/mL IL-6 or 10 ng/mL VEGF-A as directed by the manufacturer. For the injections of the recombinant proteins, 3024 µL of liquid Matrigel was mixed with 576 µL of the recombinant rat lymphangiogenic proteins yielding a final concentration of 10 ng/mL for each protein with each animal receiving a 0.5 mL injection. In addition, we collected supernatants from PBMCs stimulated with or without worm ES products (1∶10) as previously mentioned. Supernatants from 5 different individuals were pooled and 576 µL of the pooled supernatants was mixed with 3024 µL of liquid Matrigel and 0.5 mL of this mix was injected into each rat [38]. These supernatants were analyzed by luminex bead technology using the Bio-Plex 8-plex kit (IL-2, IL-4, IL-6, IL-8, IL-10, GM-CSF, IFNγ, TNFα) as well as IL-5, IL-13 and VEGF (Bio-Rad) according to the manufacturer's instructions. Regardless of Matrigel dilution with either recombinant proteins or PBMC supernatants, the Matrigel concentration was kept constant across all parameters and animals at 6.64 mg/mL.
Six rats were used per group and the various Matrigel-test agent samples were placed in the sub-dermal tissue of the flank using an 18G needle; the same position was used on each anesthetized animal, with only one plug being injected in a single rat. All injections were made by the one person and care taken with each injection to maintain a constant injection pressure and to produce a uniformly distributed plug of the material in the tissues. All rats were observed at least twice daily during the course of the experiment; they tolerated the procedures without any difficulty and did not interfere in any way with the sites where the Matrigel plugs were located.
Animals were sacrificed at day 9 as described above. During plug excision, the skin and underlying tissues/body wall were carefully dissected to observe the status of the plug and the surrounding tissues noting color, presence of scar tissue, vasculature and presence of any abnormal host tissue reaction.
The plugs and adjacent dermal tissues were removed intact from the animals and cut three times through their longest axis to provide three relatively equal slices before fixing. The plugs from each animal were photographed in situ and after removal and sectioning. Tissue sections were then taken from the cut faces of these portions to provide three different areas of each plug for histological preparation and for quantitative assessment. All tissues were fixed in 3.7% buffered formalin for a maximum of 24 h whilst maintaining that the fixation solution remained clear. Tissues were stored for processing in 60% ethanol. Specimens were then processed, embedded in paraffin, and sectioned on a rotary microtome at 4–6 µm. Sections were placed on slides coated with 2% 3-amino-propyl-tri-ethoxysilane and dried at 56°C overnight. Following de-paraffinizing in xylene and hydrating through descending concentrations of ethyl alcohol to distilled water (DW), the slides were placed in Tris-buffered saline (TBS) pH 7.5 (Scytek Labs, Logan, UT) for 5 min for pH adjustment.
Following TBS, the podoplanin and IgG test slides underwent heat-induced epitope retrieval utilizing citrate buffer pH 6.0 (Scytek) in a vegetable steamer for 30 min at 100°C, allowed to cool on the counter at RT for 10 min and rinsed in several changes of DW. Von Willebrand Factor VIII (vWF) slides underwent enzyme-induced epitope retrieval utilizing 0.03% Pronase E in TBS for 10 min at 37°C followed by running tap and DW rinses. Prior to test antibody (Ab) use, the sections were subjected to an endogenous peroxidase blocking step (3% hydrogen peroxide/methanol bath for 30 min followed by running tap water and DW rinses), a nonspecific protein blocking step for 30 min (normal horse serum) (Vector Labs, Burlingame, CA), and finally an avidin/biotin blocking system (avidin, Vector Labs; biotin, Sigma Chemicals, St. Louis, MO) for 15 min. Following pretreatment, avidin-biotin complex staining steps were performed at RT on the Dako Autostainer (Dako North America, Inc., Carpinteria, CA). All staining steps were followed by two rinses in TBS+Tween 20 (Scytek). After the sections were rinsed in TBS/Tween20, they were incubated at various times (usually 40–60 min) and various concentrations (from 1 in 40 to 1 in 200) with the various test (primary) Abs. The optimal procedures for each test Ab were determined following assessment under the microscope. Abs were diluted with Normal Antibody Diluent (NAD) (Scytek, Logan, UT). Primary Ab slides were incubated for 60 min with the monoclonal mouse anti-rat podoplanin (ReliaTech/Angio-Proteomie, Boston, MA) diluted 1∶400, or the biotin-conjugated polyclonal rabbit anti-rat IgG (Novus Biologicals, Littleton, CO) diluted 1∶100 in NAD. Slides were rinsed in 2 changes of TBS/Tween20, and then incubated in appropriate biotinylated secondary Ab for the host species of the primary Ab (biotinylated anti-rat, anti-goat, and anti-mouse from Vector Labs) at 10–11.0 µg/mL in NAD incubated for 30 min. The slides were then again rinsed in TBS/Tween20, and then R.T.U. Vectastain Elite ABC Reagent (Vector Labs) was applied for 30 min. The slides were rinsed with TBS/Tween20 and developed using NovaRED peroxidase substrate kit (Vector Labs) for 15 min. After a rinsing in DW, they were finally counterstained using Gill 2 (Lerner) hematoxylin (Thermo Fisher, Kalamazoo, MI), differentiated in 1% aqueous glacial acetic acid, and rinsed in tap water. Slides were then dehydrated, cleared with xylene, and mounted using Flotex permanent mounting media (Thermo Shandon, Pittsburg, PA).
During the establishment of optimal immunostaining for Matrigel sections we compared three antibodies as markers of lymphatic endothelia: a monoclonal mouse anti-rat podoplanin (ReliaTech/Angio-Proteomie, Boston, MA), a monoclonal mouse anti-human D2-40 (DakoCytomation) and a rabbit polyclonal anti-LYVE-1 Ab (Angiobio, Del Mar, CA). From this pre-study we selected the first of these reagents as being the most suitable for quantification. Control tissues used in these studies included rat lymph nodes and dermal neoplasia, and the staining controls included omitting the primary antibody. Routine hematoxylin and eosin staining (H&E) was also employed to examine the tissues and establish the most suitable area for quantification.
To avoid any complication from the natural host cellular response to Matrigel, the areas used for cellular assessment for vascular invasion were those central areas free of any overt host cellular response to Matrigel itself; from an examination of all the samples this free area was seen to cover a minimum of 4.0 mm2 centered around the mid point of the plug. This central area of the Matrigel plug was quantitatively assessed for the presence of anti-podoplanin and anti-vWF positivity in serial sections. Photographs were taken using bright field and Differential Interference Contrast Microscopy (DIC).
The Chalkley Point Array random sampling technique (The Graticules Ltd. Chalkley Point Array - Model NG52) was used to quantify the immuno-positive staining elements present in the Matrigel plug and thus a relationship to the proportion of the two vascular components present in different groups; the number of points lying over a positively stained entity is statistically proportional to the area occupied by that component [39]. Three areas in each section taken from the three slices of each individual plug were quantified and the number of immuno-positive components was recorded. This provided nine counts for each sample, and thus 54 counts for each treatment group.
In addition, confirmatory values were obtained using a commercial image analysis system - Image-Pro (MediaCybernetics, Bethesda, MD) and Image J (NIH - rsbweb.nih.gov/ij/) and examining the central 4.0 mm2 test area of each Matrigel slice. The immunohistochemical staining intensity was standardized for each section by setting the positivity limit for each marker using the respective cell or tissue component in the dermal tissue surrounding the plug in each section. Each area was measured using a pixel color gate (i.e. for marker positive cells), and subtracting the background using collagen tissue as the negative. Triplicate runs were made with the pixel number calculations for each assessment area.
The Signed Rank Test was used in the Statistical Analysis Software (SAS) version 9.1 to compare median cytokine and growth factor production by PBMCs in stimulated and control supernatants. The Signed Rank Test was also used to compare the production of these factors by CD14+ monocytes compared to non-CD14 cells. GraphPad Prism 5 software (San Diego, CA) was used to carry out additional statistical analyses to compare the number of tubules per microscopic field in response to stimuli. It was determined from a standard power calculation, a minimum number of animals in each test group to obtain an acceptable significant result was 5; therefore we used 6 animals in each group. The animals were injected in random order and the tissues were assessed blinded to minimize bias. Student's t test and ANOVA were used to assess the results.
We evaluated the ability of Brugia ES products to induce the secretion of molecules known to exhibit lymphangiogenic potential in other in vivo and in vitro settings. Human PBMCs were isolated from healthy volunteers and cultured with or without filarial ES products for 72 h. The supernatant fluids were collected and analyzed for the production of the potentially lymphangiogenic molecules IL-3, IL-6, IL-7, IL-8 and VEGF-A by luminex technology. PBMCs cultured in media alone supplemented with 10% FBS served as a negative control. Cells cultured with filarial ES products secreted significantly higher levels of IL-8, IL-6 and VEGF-A compared to cells cultured in media alone (Fig. 1). We did not detect IL-3 or IL-7 in any of our supernatants. We also attempted to measure VEGF-C and VEGF-D by ELISA, but these were below the limit of detection. Taken together, these data suggest that Brugia ES products are capable of inducing the secretion of lymphangiogenic molecules by circulating PBMCs.
Monocytes/macrophages have been shown to play an important role in the production of VEGFs in tumors and inflammation, so we hypothesized monocytes could be the PBMC in the periphery contributing to the production of IL-8, IL-6 and VEGF-A seen in response to worm ES products. We carried out CD14 fractionation experiments using MACS technology to isolate CD14+ monocytes from total PBMCs. As seen in Fig. 2A and 2C, CD14+ monocytes secreted significantly higher amounts of IL-8 and VEGF-A compared to CD14-depleted cells in response to filarial ES products. However, CD14-enriched and depleted cell populations produced similar levels of IL-6 (Fig. 2B). CD14+ monocytes produced significantly more IL-8 and VEGF-A spontaneously compared to CD14-depleted cells (Fig. 2A and 2C). CD14+ monocytes stimulated with Brugia ES products also secreted significantly higher levels of IL-8 and IL-6 compared to CD14+ cells cultured in media alone. LPS was used as a positive control for the production of IL-8 and IL-6 and robust IL-8 and IL-6 responses were seen following LPS stimulation. Taken together, these data suggest that CD14+ monocytes are the primary producers of the lymphangiogenic molecules IL-8 and VEGF-A in response to worm ES products, but CD14+ monocytes are not the major cell type contributing to the production of IL-6 in response to worm ES products.
Since we were able to demonstrate the production of lymphangiogenic molecules by PBMCs in response to Brugia ES products, we examined the ability of these mediators detected following ES stimulation to alter LEC function as measured by tubule formation. LECs were layered on Matrigel cultures and stimulated with concentrations of IL-8, IL-6 and VEGF-A comparable to the amounts detected in supernatants of ES-stimulated PBMCs. After 24 h, LECs cultured in the presence of IL-8, IL-6 and VEGF-A formed a more elaborate tubule network compared to cells cultured in media alone (Fig. 3A). Using image analysis software used to quantify tubule formation, cells cultured in the presence of IL-8, IL-6 or VEGF-A formed a greater number of tubules per microscopic field compared to LECs cultured without stimulus (Fig. 3B).
Given that mediators produced by PBMCs in response to filarial ES stimulation such as IL-8, IL-6 and VEGF-A induced LEC tubule formation in vitro, we hypothesized these molecules could also promote LV formation in vivo. To determine if the soluble mediators present in ES-induced supernatants could induce vessel formation in vivo, we injected rats with Matrigel containing supernatants from PBMCs (collected from 5 different individuals) that were stimulated with ES products or cultured in media alone. Characterization of the pooled PBMC supernatants which included measurable concentrations of IL-2, IL-6, IL-8 and VEGF is seen in Table 1. In parallel rats were injected with Matrigel containing rat recombinant IL-8, IL-6 or VEGF-A in case the human mediators released by PBMCs in response to filarial ES did not induce a cross species effect and stimulate vessel formation in rats. Given that Matrigel contains a variety of basement membrane proteins including laminin and collagen, Matrigel alone was used as a non-specific protein negative control.
After 9 days, the plugs were excised and subjected to gross inspection for vessel infiltration (Fig. 4A and 4B). Surprisingly, even upon initial gross examination in situ, the Matrigel plugs displayed an overt difference between treated groups and controls. Animals given ES-stimulated PBMC supernatants had increased redness in the plug denoting blood vessel infiltration compared to supernatants from unstimulated PBMCs. Furthermore, rats injected with lymphangiogenic cytokines also had an increased redness compared to Matrigel alone control plugs. The plugs in situ were generally uniform in size and shape. All except one had formed a distinct flattened oval shaped plug; one of six samples from experimental Group 2 was not clearly a round elliptical entity and was dispersed over a wide and indistinct area in the dermis; this was discarded. There was quite considerable variation in color, ranging from yellow-brown to deep pink/red. The control animals showed the yellow-brown end of the spectrum while those in groups receiving lymphangiogenic factors were generally a deeper red color (Fig. 4).
The Matrigel plugs were first examined histologically with H&E staining to identify and quantify the cellular infiltration into the central area of the plugs. Different degrees of cellular infiltration were seen in the specific quantification sites of the plugs in different test groups (Fig. 5). The principle cellular elements present were vascular; other cellular elements such as lymphocytes and monocytes were only seen within these vascular elements and not independently in the extra-vascular areas. The presentation of the vascular elements varied from tubular formations (Fig. 5B and 5C) to distinct elongated vessels (Fig. 5D). The number of cells present in the examined areas of the Matrigel plugs varied between the groups, although there was consistency in form and amount within each treatment group. Immunohistochemical staining for the presence of vWF and podoplanin was carried out to identify blood and lymphatic vessels, respectively (Fig. 5E and 5F).
Overall, staining against podoplanin which identifies the lymphatic endothelium was more prevalent in the Matrigel plugs from all groups when compared to anti-vWF staining which identifies the blood vascular endothelium. When comparing different treatments for the presence of lymphatic endothelial elements, Groups 1 (Matrigel alone) and 2 (Unstimulated PBMCs alone) were not significantly different, whereas Groups 3–6, or those containing the ES-stimulated supernatants and lymphangiogenic mediators, had significantly more lymphatic vascular elements than either Group 1 or 2 (Table 2). Plugs from Groups 3–6 had significantly more blood vascular elements than either the control Matrigel alone (Group 1) or unstimulated PBMC Matrigel (Group 2). Assessment of the color intensity by pixel enumeration with either podoplanin or vWF also showed similar significant differences between the groups VEGF-A, IL-8 and IL-6 compared to control samples and a significant difference between the ES-PBMC group compared to the unstimulated PBMC supernatant group (Table S1).
Lymphangiectasia, or the dilation of LVs, and lymphangiogenesis are subclinical features of filarial infection. LVs containing adult worms from infected individuals are characterized as distended, dilated, tortuous and highly indented [40]–[42]. In dilated lymphatics, flow is impaired leading to improper drainage of interstitial fluids. The progression of mild lymphangiectasia to clinical lymphedema may be due to the accumulation of lymphatic fluid in the tissues over time following damage to the LVs. Lymphangiectasia is not restricted to the site of the worm nest, but is found along the length of the infected vessel [8] arguing that a soluble factor secreted by the worm, that can travel the length of the vessel, is responsible for the altered lymphatic pathology. Additionally, lymphangiectasia is greatest near the worm nest and the removal or killing of worms can reduce lymphatic dilation [14]–[16], [40] suggesting living adult worms and their ES products have the strongest biological effects locally and are associated with altering lymphatic pathology.
A number of factors may play a role in the development of lymphangiectasia and our data suggest that parasite products are central in this process. Since no direct effects of ES products on LECs were detected, we hypothesized that ES products activate the lymphatic endothelium indirectly through an accessory cell [43]. Here, we have demonstrated that Brugia ES products stimulate host cells to produce lymphangiogenic mediators such as IL-8, IL-6 and VEGF-A. Autocrine stimulation by these molecules on the PBMCs themselves may have also amplified the response in our system. Next, we demonstrated these same mediators altered LEC phenotypes. Moreover, the mediators tested in this study not only induced LV formation in vivo using a Matrigel plug model, but these mediators also induced angiogenesis. Therefore, the production of these molecules could contribute to the development of lymphangiectasia in filarial-infected individuals.
Other studies have supported the role of parasite molecules in lymphangiogenesis and lymphangiectasia. Bennuru et al. showed microfilariae stimulate LEC proliferation and alter LEC junction adherence pathways which could contribute to lymphatic dilation [44]. Microfilariae may also contribute to the development of lymphatic disease as this stage is released simultaneously with adult ES products and microfilarial ES is found in our adult worm ES. Others have proposed that parasite endosymbiont Wolbachia is responsible for elevated lymphangiogenic mediators, but Bennuru et al. elegantly demonstrated that the levels of VEGF-A, VEGF-C and VEGF-D pre- and post-DEC treatment did not change suggesting a minimal role for Wolbachia [27]. Bacterial infection, including Wolbachia, has been linked with IL-8 production [45], so the levels of other lymphangiogenic mediators such as IL-8 and IL-6 will also need to be examined in this setting. Furthermore, human ECs exposed to live intact microfilariae either carrying or free of Wolbachia or not, only induced a limited number of cytokines and angiogenic mediators suggesting Wolbachia is not a strong stimuli altering the EC phenotype [46].
In this present study, we aimed to mimic the relationship between the living adult worm and the lymphatic endothelium, and not the changes associated with dead worms, thus we used Brugia ES products rather than adult worm or microfilariae extracts. Crude extracts would be more representative of stimuli associated with worm death, a different scenario. Upon worm death, there is an immense inflammatory reaction which is distinct from the lack of inflammation associated with the presence of the living worm. Responses to living worms differ histopathologically from the granulomatous responses seen with dead worms (Mackenzie, unpublished observations). Monocytes/macrophages appear to be central in both responses, although they may be acting differently in each situation. Filarial ES products are generally thought to be immunosuppressive but here ES induced PBMCs to produce IL-8 and IL-6 which can lead to a massive recruitment of inflammatory cells. However, the lack of inflammation adjacent to living worms suggests IL-8 and IL-6 production does not lead to a massive inflammatory reaction in vivo. In contrast, worm death either by drug treatment or natural attrition may exacerbate the development of lymphatic pathology if the acute inflammatory reaction provides a stimulus for downstream processes leading to lymphatic insufficiencies. Future studies will be needed to compare the production of lymphangiogenic mediators and the induction of LVs in vivo in response to ES products versus crude extracts.
Even though the expression of lymphangiogenic mediators is generally perceived to be beneficial for the formation of new LVs and to reverse malfunctioning LVs [47]–[49], the over-expression of lymphangiogenic molecules over an extended period of time has been shown to be detrimental and to impair lymphatic function. A massive expansion of the lymphatic network can lead to defective LVs and thus decreased drainage and lymphedema. For example, VEGF-A and VEGF-C over-expression results in structurally and functionally abnormal and dilated lymphatics [50]–[52]. ES-stimulated host cells may compromise lymphatic function by secreting lymphangiogenic factors over many years throughout the duration of worm infection. It is important to note that a worm infection can last five years or more so the kinetics and molecular mechanisms associated with altering lymphatic pathology may differ from those involved in acute infection and may be cumulative over time. The cumulative amounts/effects of these soluble mediators may parallel those observed in over-expression model systems leading to defective lymphatics. For instance, elevated plasma levels of VEGF-C have been found in microfilaremic individuals compared to endemic normal individuals [28] suggesting the same VEGF and cytokine molecules involved in lymphangiogenesis and lymphangiectasia in other models are also present in filarial infection. These lymphangiogenic cytokines and growth factors may be binding their receptors which are expressed on LECs lining the vessel [20], [53], [54]. Besides the chronicity of filarial infections, worm infections, and specifically worm ES products, are also associated with a down regulation of the immune response so future experiments will also need to address how a chronic infection alters the formation of LVs in the presence of a dampened proinflammatory response.
Even though we did see the production of VEGF-A by PBMCs in response to worm ES, we did not see the production of VEGF-C that was previously shown to be elevated in filarial-infected individuals [28], [31]. We also did not detect elevated levels of VEGF-D or lymphangiogenic cytokines IL-3 or IL-7. The lack of detection of VEGF-C, VEGF-D, IL-3 or IL-7 may be because we were examining the production of these molecules by PBMCs which may not be the cellular source; these molecules may be produced by a cell found focally at the infection site. VEGF-C and VEGF-D signaling through VEGFR-3 is the primary and most well-characterized mechanism contributing to lymphangiogenesis, but there is also an emerging role for VEGF-A in lymphangiogenesis [52], [55]–[57], so it is possible that this molecule may be playing an important role in filarial-induced lymphatic pathologies. In addition to potential systemic versus local differences in lymphangiogenic mediators, differences between individual responses were also noted. The variability in lymphangiogenic mediators, especially for IL-8, produced by PBMCs basally and after ES stimulation made control experiments injecting supernatants from unstimulated PBMCs of paramount importance. Regardless, supernatants from ES-stimulated PBMCs induced significantly more podoplanin and vWF staining compared to supernatants from unstimulated PBMCs. Furthermore, supernatants from unstimulated PBMCs induced more vessel formation than Matrigel alone confirming the basal production of these mediators and providing an important baseline control beyond Matrigel alone.
Monocytes and macrophages play a major role in supporting lymphangiogenesis. They can produce lymphangiogenic factors such as VEGFs and cytokines which induce LEC proliferation, survival, migration and tubule formation [33], [34]. In this present study monocytes were primarily responsible for the production of IL-8 and VEGF-A in response to ES products; however we did not identify the cell type responsible for the production of IL-6, so future experiments need to identify the source of IL-6.
Monocytes and macrophages may play a role in the lymphatic pathology associated with filarial infection. Typically, LVs from infected individuals are thought to be devoid of an inflammatory response [41]; however, some have noted small lymph thrombi composed of mononuclear cells and multinucleated giant cells within the lumen [42]. Here, we defined CD14+ cells as the primary producer of IL-8 and VEGF-A in response to Brugia ES products and others have also reported the presence of monocytes/macrophages in regions of lymphangiectasia and lymphangiogenesis in O. volvulus infection [30], [31]. In nodules isolated from humans infected with O. volvulus, the predominant cell type associated with the worms was the macrophage and many macrophages stained positive for the lymphatic-specific marker LYVE-1 [30]. Additionally, some LYVE-1+ macrophages were integrating into the lymphatic endothelium [31]. Taken together, these data suggest that monocytes/macrophages are important in lymphangiectasia and lymphangiogenesis in filarial infections and future research is needed to define the role of these cells in lymphatic filariasis.
One could speculate that the worm induces lymphangiogenesis and lymphangiectasia for many reasons. The worm may increase vessel diameter to provide a larger space for habitation; increasing the vessel diameter also slows lymphatic flow and increases the availability of nutrients and resources. The worm may stimulate expansion of the lymphatic network by inducing host production of VEGFs and cytokines to increase LEC proliferation and differentiation as a mechanism of LV dilation. We also demonstrated tubule formation in response to ES-stimulated mediators. Filarial worms may induce the formation of new LVs to expand their biological niche, to maintain flow through a collateral network, or to increase the likelihood that their microfilariae reach the periphery for transmission.
In this study we have begun to dissect the molecular mechanisms involved in the development of lymphangiectasia and lymphangiogenesis; however, similar studies must be carried out in cells isolated from endemic populations to confirm that the same molecules and cell types occur in filarial-infected individuals. Given that parasite products induce the production of lymphangiogenic molecules and that infected persons exhibit lymphangiectasia, we hypothesize that these molecules are elevated in infected individuals. We are currently examining the production of VEGFs and cytokines by microfilaremic individuals, endemic normals and those with lymphedema in response to ES products. Since many infected individuals exhibit lymphangiectasia, which may progress to a lymphedema, we need to define the initial molecular mechanisms responsible for the development of disease. Given many of the lymphangiogenic mediators identified in this study are expressed in a variety of inflammatory settings, we hypothesize that lymphangiogenesis is a hallmark of inflammation. Therefore, understanding the pathogenesis of lymphatic filariasis may identify potential molecular targets for preventing disease initiation and progression as well as a greater understanding of the molecular mechanisms associated with lymphatic pathologies from cancer and inflammation.
|
10.1371/journal.pntd.0000932 | Subacute Sclerosing Panencephalitis in Papua New Guinean Children: The Cost of Continuing Inadequate Measles Vaccine Coverage | Subacute sclerosing panencephalitis (SSPE) is a late, rare and usually fatal complication of measles infection. Although a very high incidence of SSPE in Papua New Guinea (PNG) was first recognized 20 years ago, estimated measles vaccine coverage has remained at ≤70% since and a large measles epidemic occurred in 2002. We report a series of 22 SSPE cases presenting between November 2007 and July 2009 in Madang Province, PNG, including localized clusters with the highest ever reported annual incidence.
As part of a prospective observational study of severe childhood illness at Modilon Hospital, the provincial referral center, children presenting with evidence of meningo-encephalitis were assessed in detail including lumbar puncture in most cases. A diagnosis of SSPE was based on clinical features and presence of measles-specific IgG in cerebrospinal fluid and/or plasma. The estimated annual SSPE incidence in Madang province was 54/million population aged <20 years, but four sub-districts had an incidence >100/million/year. The distribution of year of birth of the 22 children with SSPE closely matched the reported annual measles incidence in PNG, including a peak in 2002.
SSPE follows measles infections in very young PNG children. Because PNG children have known low seroconversion rates to the first measles vaccine given at 6 months of age, efforts such as supplementary measles immunisation programs should continue in order to reduce the pool of non-immune people surrounding the youngest and most vulnerable members of PNG communities.
| Subacute sclerosing panencephalitis (SSPE) is a disabling and usually fatal brain disorder that typically occurs 3–10 years after acute measles infection. Papua New Guinea (PNG) has particularly high rates of SSPE. We report 22 cases of PNG children presenting to the provincial referral hospital in Madang Province who probably contracted acute measles when <12 months of age during a national epidemic in 2002 and who developed SSPE 5–7 years later. Based on these cases, the estimated annual SSPE incidence in Madang province in 2007–2009 was 54/million population aged <20 years. Four sub-districts had an annual incidence >100/million population aged <20 years, the highest rates ever reported. Young PNG children do not respond well to measles vaccine. Because of this, efforts such as supplementary measles immunisation programs should continue in order to reduce the pool of non-immune older people surrounding the youngest and most vulnerable members of PNG communities.
| Despite a declining incidence in developed countries, acute measles infection is still responsible for an estimated 164,000 deaths/year and is therefore a major vaccine-preventable cause of death worldwide [1], [2]. Subacute sclerosing panencephalitis (SSPE) is a rare but usually fatal late complication which presents 3–10 years after the acute infection. SSPE is a distinctive clinical entity characterized by behavioural changes and myoclonic jerks, followed by motor dysfunction and profound global cognitive impairment, and then death within a few years of presentation in most cases. The diagnosis is made by the presence of characteristic clinical signs and, if available, electroencephalographic (EEG) findings in conjunction with elevated measles-specific antibodies in serum and cerebrospinal fluid (CSF) [3].
The incidence of SSPE in most countries is <5 per million population <20 years of age, although this figure can be higher in the developing world where vaccination programs are not fully established [4]. The first reports of an unusually high incidence in Papua New Guinea (PNG) were published in the early 1990's [5], with rates between 1988 and 1999 that varied from 13 [5] to 98 [6] per million population <20 years of age. However, these data need to be interpreted against fluctuations in the incidence of acute measles infection over the preceding decade, and should take into account background vaccination coverage and the possibility that localized clusters may contribute disproportionately to overall incidence rates estimated at provincial or country level. In addition, published PNG data to date have come from highland areas which may not be representative of the country as a whole.
We report a series of children presenting to a coastal PNG provincial referral hospital with clinical and laboratory features typical of SSPE. Using available local demographic data, as well as retrospective vaccination and disease surveillance, we have estimated the annual incidence of SSPE in Madang Province and interpreted this figure in relation to prior national measles vaccination coverage and acute measles incidence, as well as the regional distribution of cases.
Approval for the study was provided by the PNG Institute of Medical Research Institutional Review Board and the Medical Research Advisory Committee of the PNG Health Department. Written informed consent for participation was obtained from parent(s)/guardian(s). The risks and benefits of lumbar puncture (LP) were explained to parent(s)/guardian(s) by the attending ward pediatrician who carried out the procedure with regard for conventional indications (suspicion of meningitis, subarachnoid hemorrhage or central nervous system disease) and contraindications (such as increased intracranial pressure or coagulopathy) [7].
Madang Province on the North Coast of PNG has an estimated population of approximately 450,000 people, 54% of whom are <20 years old [8]. Modilon Hospital is the provincial referral hospital and the only health care facility in the province that offers diagnostic and treatment facilities for severely ill patients. A longitudinal detailed observational study of severe illness in all children aged 6 months to 10 years was started at Modilon Hospital at the end of 2006. Prior to this initiative, documentation of cases was insufficient to allow epidemiologic analyses of specific diseases. In November 2007, the first child with symptoms and signs of SSPE was admitted to the present study. There was a subsequent increase in the numbers of similar cases before a decline after 12 months. Data collection was continued until July 2009, at which time relatively few such cases were being admitted.
Measles immunization was started in PNG in 1982. A modified two-dose schedule at six and nine months of age was used with the aim of providing partial coverage for young infants at high risk of pneumonia and SSPE [9]. However, subsequent available national data indicate that coverage has remained low (see Figure 1). In a recent study of 2007 data, for example, 58% of eligible children received the first dose and 47% the second dose [10]. Cyclical measles epidemics have continued to occur, the last in 2002 (see Figure 2) [11], [12], [13]. Supplementary immunisation activities (SIA) for children aged 6 months to 7 years have been deployed since 2004, with a reported coverage of 79% in 2008 [13].
The measles vaccine coverage recorded in the health diaries of children in Madang Province is similar to that reported elsewhere in PNG, with 41% of children <10 years of age surveyed at two sites within a 20 km radius of Madang town between September 2007 and June 2008 having received at least one dose [10]. Nevertheless, an increasing seroprevalence with age (60% and 79% for children 1–4 years and 5–9 years old, respectively) may indicate that wild measles virus remains prevalent in the community [10] and that there is under-reporting of cases as found in other epidemiologic settings [14], [15].
After recruitment, a standardized case report form was completed detailing demographic information, medical history and history of the current illness. Vaccination history was identified from the health record book held by the parent(s)/guardian(s) of each child where this was available. Since there is no local or central vaccination register, it was assumed that children without such documentation were unvaccinated. Standardized physical assessment included nutritional status assessed by calculating a weight-for-height Z-score [16], with a value <2 considered to indicate malnutrition. We defined severe illness as the presence of one or more of the following features: i) impaired consciousness or coma (Blantyre Coma Score (BCS) <5 [17]), ii) prostration (inability to sit or stand unaided), iii) multiple seizures, iv) hyperlactatemia (blood lactate >5 mmol/L), v) severe anemia (hemoglobin <50 g/L), vi) dark urine, vii) hypoglycemia (blood glucose <2.2 mmol/L), viii) jaundice, or xi) respiratory distress. These criteria are consistent with the World Health Organisation definition for severe malaria [18].
Children with clinical evidence of SSPE, including myoclonic jerks, behavioural changes, and/or speech and motor deficits, underwent detailed neurologic examination by study clinicians (LM, ML). Level of consciousness was graded according to Blantyre Coma Score [17]. Upper motor neuron signs were considered to be present if the child had i) extensor plantar responses, ii) increased muscle tone of either upper or lower limbs, iii) sustained clonus, iv) hyperreflexia, and/or v) pyramidal tract muscle weakness of either upper or lower limbs. In children whose parents/guardians provided informed consent and who had no contraindications, LP was performed. All children were examined daily until discharge at which time a basic assessment of performance status was made. Moderate disability was defined as that requiring considerable assistance with self-care and severe disability as that requiring special assistance with all self-care, categories that are consistent with Karnovsky's performance scores of 50% and <50%, respectively [19].
CSF was examined macroscopically for turbidity, blood staining and clots. We used the Neubauer Improved counting chamber (BoeCo, Germany) to obtain total and differential CSF white cell counts (WCC). Semi-quantitative measures of CSF glucose and protein were performed using dipsticks (Acon Laboratories, San Diego, USA). Specific measles IgG in CSF and serum was measured using a standard indirect immunofluorescence antibody assay (IFA). Serial two-fold dilutions of patient samples were added to separate wells of glass slides to which were fixed measles virus-infected Vero cells. After incubation and washing, anti-human IgG fluorescein isothiocyanate conjugate was then added and, following further incubation and washing, slides were examined under an ultra-violet microscope. Fluorescence was scored as 1+ to 4+, with levels of ≥1+ regarded as positive. This test was performed in an accredited laboratory and had been assessed and approved by the Australian National Association of Testing Authorities in accordance with requirements of the Australian National Pathology Accreditation Advisory Council. Details of other laboratory tests including malaria microscopy, plasma biochemistry and bacterial culture have been published elsewhere [20].
Confirmed SSPE was defined as clinical features of SSPE and the presence of measles-specific IgG in CSF, regardless of titer. Probable SSPE was defined as clinical features of SSPE and negative measles-specific IgG in CSF or when no LP was performed.
The calculation of SSPE incidence was based on PNG Census data for the year 2000 [8] which includes population structure at provincial, district and local-level government (LLG, sub-district) level. The 2008 population was estimated by applying an annual growth rate of 2.6% (Dr Bryant Allen, Australian National University, Canberra, Australia; personal communication). Using this approach, the total population for Madang Province was estimated to be 448,330 with 241,165 (53.8%) <20 years of age. The Global Positioning System co-ordinates of each child's home village were obtained to facilitate LLG incidence estimates [21]. All SSPE incidence rates were expressed per million population <20 years of age which ranged from 5,545 in Iabu Rural LLG to 28,066 in Amenob Rural LLG with an inter-quartile range of 9,880 to 21,267. Reported annual rates of measles vaccination coverage [11], [12] and cases of acute measles infection reported to the PNG Department of Health [12], [13] were obtained from World Health Organization sources. Statistical testing was by means of parametric or non-parametric tests using PASW Statistics (version 17; SPSS Inc. Chicago, Ill) and a level of significance of 0.05.
Baseline, clinical and laboratory data relating to cases of SSPE identified during the 19-month surveillance period are summarized in Table 1. These 22 children (16 confirmed and 6 probable cases; see below) were a subset of 671 admitted with severe illness during the study period. Although the median duration of illness prior to admission reported by the parent(s)/guardian(s) was 60 (range 1 to 1,000) days, the data provided were not sufficient to allow an accurate estimate of the age of each child at symptom onset.
Two children had documentation or parental knowledge of a past history of acute measles infection, one at six months and the other at two years of age. Neither had a documented history of measles vaccination. There were 14 (64%) children in whom the first dose of measles vaccine had been given and all but one of these (59%) had subsequently received the second dose. In a contemporaneous sample of 44 children hospitalized with other severe non-SSPE illness matched 2∶1 by age and sex with the SSPE cases, the equivalent percentages were 67% and 67% respectively (P>0.55 by Chi-squared test). Two children diagnosed with SSPE within a few months of each other were first cousins.
Sixteen children had characteristic myoclonic jerks on admission and four had a clear prior history of myoclonic jerks obtained from the child's parents. One child presented with a short (two-week) history of severe involuntary muscle spasms and died soon after admission, while another presented with complex involuntary dyskinetic movements of upper and lower limbs. The majority of children had additional neurologic findings such as impaired consciousness, difficulty walking and impairment of speech.
LP was performed in 18 of the 22 children. Sixteen of these (89%), including the two with atypical non-myoclonic features, had high titre measles-specific antibodies in both serum and CSF and were therefore confirmed cases of SSPE. Of the probable cases, four did not undergo LP but each had high serum titres of measles-specific antibodies. The remaining two children presented with clinical features consistent with SSPE (myoclonus, motor and speech deficits) with negative CSF measles IgG titers but elevated serum titres at 1∶2048 and 1∶16, respectively. The latter child had no history of measles vaccination.
In all six probable SSPE cases, no other cause of encephalopathy was identified. Normal plasma electrolytes and hepatorenal function excluded metabolic, renal and hepatic encephalopathy. Giemsa-stained thick blood films were negative for malaria parasites and plasma C-reactive protein, blood lactate, white cell count and blood culture results did not suggest an acute infective aetiology. The absence of a CSF pleocytosis in the two children with probable SSPE in whom LP was performed made tuberculous meningitis or cryptococcal meningitis unlikely. CSF from both children was negative by PCR for enteroviruses, Japanese encephalitis virus, Murray Valley encephalitis virus, West Nile virus (including Kunjin) and dengue virus, and serum was negative for the presence of IgM to flaviviruses. Based on clinical presentation and course, serum measles antibody titres and the exclusion of other causes of an encephalopathy, the six children with probable SSPE were included in estimates of SSPE incidence.
Although only one child died in hospital, the remaining children were discharged in line with usual management of SSPE in PNG. These children had moderate or severe disability requiring assistance with most or all activities of daily living. An examination of post-discharge outcome was beyond the scope of the present study.
Figures 1 and 2 show PNG national vaccine coverage and acute measles cases since 1997 [11], [12], [13], and the year of birth of the present 22 SSPE cases is shown in Figure 3. Despite relatively stable vaccination coverage between 50% and 65% from 1997 to 2008, there was a substantial increase in the numbers of reported acute measles cases in 2002 with a smaller prior peak in 1999 and 2000. There is a close concordance between the distribution of the years of birth of the SSPE cases and that for acute measles nationally (Spearman r = 0.88, P = 0.002).
The location of the home village for each child with SSPE and the annual incidence of SSPE in the 13 districts in Madang Province are shown in Figure 4. The majority of the children were from remote rural districts with very limited health care access. The overall estimated annual incidence for Madang province was 29 (95% confidence intervals [18 to 45])/million total population or 54/million population <20 years of age. In Josephstaal, Yawar, Astrolabe Bay and Bundi LLGs, the estimated annual incidence was 296 [96 to 691], 194 [78 to 400], 122 [15 to 442] and 119 [3 to 660]/million, respectively. There were no reported SSPE cases from 4 districts. Three of these have no roads and are only accessible by air, river or foot.
The present study conducted in coastal Madang Province confirms the relatively high incidence of SSPE in PNG shown previously in several highland provincial surveys conducted during the 12 years up to 1999 [5], [6], [22]. However, our data also show that such incidence rates must be interpreted in the light of prior measles epidemiology. There was a clear association between the year of birth of our SSPE cases and national figures for acute measles infection that included a substantial increase in cases in the year 2002. This relationship suggests that, despite the possibility of under-reporting [10], [14], [15], temporal trends in measles cases in PNG are relatively accurate. Without equivalent antecedent data, it is difficult to interpret prior reports [5], [6], [22] in which a high SSPE incidence may have simply reflected peaks in measles cases 3–10 years beforehand. The decline in reported acute measles in PNG since 2002, including very few cases over the last 5 years [13], should herald a substantial reduction in SSPE incidence in PNG over the next few years. Nevertheless, a rising seroprevalence during childhood which exceeds that associated with vaccination coverage and SIA may mean that continued local measles transmission will sustain future low-level presentation of new cases [10]. Although our study captured the delayed peak in SSPE incidence attributable to the 2002 measles epidemic, there have been two further children admitted to Modilon Hospital with a clinical diagnosis of SSPE in the 12 months since recruitment to the present study finished.
The demographic features and clinical course of our patients were similar to those of published series from PNG and other countries. Consistent with previously-reported studies [23], we could not always determine age of onset of symptoms accurately, but the median age at the time of admission in our children (7.3 years) and the male∶female ratio (1.4∶1) were similar to those in SSPE cases from the PNG highlands a decade ago (7.9 years and 1.2∶1, respectively) [4]. Although a male excess is usual, there has been a large age range at presentation, from <5 years in one of the first PNG studies [22] to >10 years in Europid populations [23], [24]. This is likely to reflect population-specific differences in contributing factors such as persistence of maternal antibodies and vaccination policies.
The present study is the first to have had access to LLG population data to facilitate an assessment of SSPE epidemiology at a sub-provincial level in PNG. The incidence of SSPE exceeded 100 per million population <20 years old in four LLGs of Madang Province. Although there were relatively few cases in some of sub-districts, this is the highest rate yet recorded. Only half of PNG children receive both doses of measles vaccine before their first birthday [10], [13], but the prior measles vaccination rate documented for the children from these districts did not differ significantly from that of the non-SSPE severely ill control children nor from national coverage at the time of likely measles infection. This suggests that factors other than vaccine delivery were responsible. There are known continuing difficulties with ensuring a reliable vaccine cold chain in PNG [25], [26], but it is also possible that post-measles vaccination seroconversion rates were unusually low in these areas or that the acute measles incidence was particularly high. Alternatively, the children in these communities have an increased susceptibility to SSPE.
Vaccine seroconversion is highly age-dependent. Only 36% of Melanesian children will develop protective measles immunity after their first vaccination at 6 months [27] while recent data from Madang also indicate low rates of protective immunity to measles in children who had received one or both doses of measles vaccine before one year of age [10]. This reflects, in part, persistence of low-level interfering passive maternal antibodies for up to 12 months [28], especially when maternal immunity has been acquired by natural infection rather than vaccination [29]. The weight of epidemiologic evidence suggests that SSPE is more likely to occur when measles infects a child in the first year of life [23], [30], [31]. Indeed, the clear relationship between year of birth of our children with SSPE and nationally reported measles incidence implies that our cases were very young when they encountered measles virus for the first time. Given that most young children are vulnerable to measles, even if vaccinated, it is likely that differences in the numbers of SSPE cases between districts reflect similar local differences in acute measles incidence between 1998 and 2003.
To ensure adequate herd immunity to measles, countries must achieve 92–95% vaccination coverage that includes two separate doses of the vaccine [32]. Given that the expanded programme for immunisation started in 1982 in PNG and the fact that measles vaccine coverage in PNG has remained ≤70% for at least the last 10 years [11], it is likely that herd immunity was very low in the more remote communities of Madang Province early in the millennium and that acute measles cases were correspondingly high. The fact that four very remote districts were not represented in our series suggests that either they were isolated from the increase in acute measles cases at that time or that children with SSPE were not brought to Modilon Hospital because of the logistic issues involved with patient transfer.
Prior to widespread vaccination, the incidence of SSPE was between 1.2–6.7 per million population <20 years of age in countries where valid data were available [33], [34], [35], [36]. However, incidence rates up to 43 per million population <20 years of age have been estimated in some developing countries [6]. Furthermore, even within closely located communities, the incidence of SSPE is not uniform. For example, Ashkenazi Jews in Israel have a lower rate of SSPE than Separdic Jews (0.5 vs 3.4 cases per million population, respectively) [36]. Our children are from a comparatively homogenous Melanesian group but there were two first cousins of similar age diagnosed within a year of one another. Although from a single set of close-living relatives, SSPE has been described previously in sibling and twin pairs implying at least some familial predisposition, but a clear genetic basis for susceptibility has yet to be defined [37], [38], [39], [40], [41]. Single nucleotide polymorphisms in a number of immunity-related genes have been found to be associated with SSPE in Japanese [42], [43] and Turkish [44] patients, but not in other ethnic groups [45]. The search for genetic associations is difficult in an uncommon disease like SSPE, even in high incidence settings such as PNG, and they would have to account for socio-cultural factors that might promote measles infection at an early age in relatively non-immune, unvaccinated populations [46].
SSPE is diagnosed on clinical grounds alone in resource poor, high incidence settings similar to that of the present study where there are no brain imaging or EEG facilities. Serologic testing for measles is only available as a research or epidemiologic tool. In our case series, six children had probable SSPE without confirmatory CSF serology. Four of these children did not have LP performed but had very high serum titres of measles-specific IgG. The remaining two had negative CSF serology but characteristic clinical features. Because relatively comprehensive clinical and laboratory investigations excluded other likely causes of encephalopathy, the fact that measles CSF and serum titres in SSPE cases can overlap those of controls [3], and given the high pre-test probability of SSPE, we believe that these two latter children represent part of the spectrum of the disease. Our series also included two children with atypical clinical features. One presented with muscle spasms and rapidly progressed to death within two weeks of illness onset. The other had a subacute history and complex dyskinetic movements. Atypical presentations have been described previously and are sometimes accompanied by radiologic evidence of extensive brainstem as well as cortical involvement [47]. It is likely that, had cerebral imaging been available, there would have been similar radiologic findings in these two children.
The PNG Pediatric Guidelines recommend that the first measles vaccine be given at 6 months of age and the second at 9 months of age [7]. The reason for this policy is the increased morbidity and mortality from acute measles in younger infants rather than high rates of SSPE [48]. However, the low seroconversion rates in this age-group argue for a delay in the age of the first measles vaccine to 9 months of age followed by a second dose at 12–15 months. This could be reconsidered if there were to be an outbreak of measles, but this has not happened in PNG for the last 8 years.
The present study extends past published data suggesting that PNG has the highest reported incidence of SSPE globally. However, this high incidence is related to prior measles epidemics. We have also shown that the incidence of SSPE varies between communities. This could reflect localized failure of the vaccine cold chain, community-specific factors that increase measles transmission, variability in public health surveillance and/or differences in genetic susceptibility to SSPE. Young PNG children do not respond well to measles vaccine. Because of this, efforts such as SIA should continue in order to reduce the pool of non-immune older people surrounding the youngest and most vulnerable members of PNG communities. SSPE has a high mortality, but most children with SSPE require prolonged care because of profound disabilities. Such dependence comes at substantial cost for caregivers.
|
10.1371/journal.ppat.1007983 | The heterogeneous nuclear ribonucleoprotein hnRNPM inhibits RNA virus-triggered innate immunity by antagonizing RNA sensing of RIG-I-like receptors | Recognition of viral RNA by the retinoic acid-inducible gene-I (RIG-I)-like receptors (RLRs), including RIG-I and MDA5, initiates innate antiviral responses. Although regulation of RLR-mediated signal transduction has been extensively investigated, how the recognition of viral RNA by RLRs is regulated remains enigmatic. In this study, we identified heterogeneous nuclear ribonucleoprotein M (hnRNPM) as a negative regulator of RLR-mediated signaling. Overexpression of hnRNPM markedly inhibited RNA virus-triggered innate immune responses. Conversely, hnRNPM-deficiency increased viral RNA-triggered innate immune responses and inhibited replication of RNA viruses. Viral infection caused translocation of hnRNPM from the nucleus to the cytoplasm. hnRNPM interacted with RIG-I and MDA5, and impaired the binding of the RLRs to viral RNA, leading to inhibition of innate antiviral response. Our findings suggest that hnRNPM acts as an important decoy for excessive innate antiviral immune response.
| Infection by virus, such as the RNA virus Sendai virus, induces the host cells to express proteins that mediate antiviral immune responses. Upon infections, the retinoic acid-inducible gene-I (RIG-I)-like receptors (RLRs) detects the intracellular viral RNA and initiates innate immune responses. Although the regulation of RLR-mediated signal transduction has been extensively investigated, how the recognition of viral RNA by RLRs is regulated remains enigmatic. In this study, we found that a protein called hnRNPM plays an important role in the process of antiviral immune response. hnRNPM does this by impairing the binding of the RLRs to viral RNA. Our results suggest that hnRNPM is an inhibitor of RNA virus-induced signaling which provides a critical control mechanism of viral RNA sensing for the host to avoid excessive and harmful immune response.
| Innate immune response provides the first line of host defense against invading microbial pathogens [1]. Upon infection, the conserved microbial components called pathogen-associated molecular patterns (PAMPs) are sensed by cellular pattern recognition receptors (PRRs). This leads to induction of type I interferons (IFNs), pro-inflammatory cytokines, and other downstream effector genes. These downstream effector proteins mediate innate immune and inflammatory responses to inhibit microbial replication and clear infected cells [1, 2].
Viral nucleic acids are major PAMPs that are sensed by the host cells after viral infection. Extracellular viral RNA is recognized by transmembrane and endosomal Toll-like receptor 3 (TLR3), which is expressed mostly in immune cells [3], whereas intracellular viral RNA is detected by the retinoic acid-inducible gene-I (RIG-I)-like receptors (RLRs), including RIG-I and MDA5[4]. Genetic studies have demonstrated that RIG-I and MDA5 play crucial roles in innate immune response to different types of RNA viruses [1] [5]. RIG-I and MDA5 utilize similar signaling pathways to induce downstream antiviral genes. Upon binding to viral RNA, RIG-I or MDA5 undergoes conformational changes and is recruited to the mitochondrial membrane-located adaptor protein VISA (also called MAVS, IPS-1 and Cardif) [6–9]. This triggers the formation of large prion-like VISA polymers, which in turn serve as platforms for recruitment of TRAF2/3/5/6 through its TRAF-binding motifs [10, 11]. The TRAF proteins further recruit TBK1 and the IKK complex to phosphorylate IRF3 and IκBα respectively, leading to activation of IRF3 and NF-κB and induction of downstream antiviral effectors.
Both RIG-I and MDA5 contain two tandem caspase-recruitment domains (CARDs) at their N terminus, which mediate downstream signaling; a central DExD/H helicase domain with an ATP-binding motif; and a C-terminal RNA-binding domain [5]. Although RIG-I and MDA5 share similar signaling features and structural homology, various studies have demonstrated that the two helicases may discriminate among different ligands to trigger innate immune response. It has been demonstrated that RIG-I preferably recognizes viral 5’-ppp double-strand (ds) RNA and relatively short (approximately 300 bp) dsRNA, while MDA5 has higher affinity to long dsRNA [12–14]. Various studies have shown that RIG-I is essential for induction of downstream antiviral effector genes in response to RNA viruses including Sendai virus (SeV), vesicular stomatitis virus (VSV), Newcastle disease virus (NDV), influenza virus and Japanese encephalitis virus (JEV), whereas MDA5 is critical for the detection of picornaviruses, such as encephalomyocarditis virus (EMCV) [15, 16].
RLR-mediated innate antiviral responses are regulated by distinct mechanisms. For examples, TRIM25, TRIM4, Riplet (also known as RNF135), TRIM13, USP4, USP3, USP15, USP21, CKII, PP1α/γ, and TRIM38 have been reported to regulate the post-translational modifications of RLRs [17–27]. MEX3C in stress granules enhances the affinity between viral RNA and RIG-I [28]. RAVER1 regulates MDA5- but not RIG-I-mediated antiviral immune response by promoting the binding of MDA5 to viral RNA [29]. More recently, it has been shown that ZCCHC3 acts as a co-receptor for the binding of RIG-I and MDA5 to viral RNA [30]. However, how RLR activation is monitored to prevent excessive innate antiviral response is unclear.
The heterogeneous nuclear ribonucleoprotein M (hnRNPM) contains three RNA recognition motif (RRM) domains. It has been shown that hnRNPM is involved in pre-mRNA splicing and diverse aspects of RNA metabolism, including translational control, telomere biogenesis, mRNA stability, and trafficking[31, 32]. Here we identified hnRNPM as a decoy of innate antiviral response. Viral infection led to export of hnRNPM from the nucleus to the cytoplasm, at where it impaired the binding of RLRs to viral RNA and subsequent innate antiviral response. Our findings reveal a mechanism on how the sensing of viral RNA by RLRs is properly regulated.
To identify candidate molecules involved in viral RNA-triggered innate immune response, we screened ~10,000 independent human cDNA clones for their abilities to regulate IFN-β promoter activity by reporter assays and identified hnRNPM as a candidate protein. As shown in Fig 1A, overexpression of hnRNPM inhibited SeV-triggered activation of the IFN-β promoter, ISRE and NF-κB. Conversely, knockdown of hnRNPM facilitated SeV- and EMCV-induced transcription of the IFNB1, ISG56 and CXCL10 genes (Fig 1B), but not IFN-β-induced transcription of the ISG15 gene (Fig 1C). Consistently, knockdown of hnRNPM enhanced SeV-induced phosphorylation of IRF3, TBK1, STAT1 and IκBα (Fig 1D). These results suggest that hnRNPM negatively regulates RNA virus-triggered induction of antiviral genes.
To investigate whether endogenous hnRNPM is required for innate immune response to RNA virus, we generated hnRNPM-deficient HEK293 individual clones by the CRISPR-Cas9 method. We found that transcription of the IFNB1, ISG56, and CXCL10 genes induced by SeV or VSV were markedly increased in hnRNPM-deficient cells (Fig 2A). Consistently, transcription of the IFNB1, ISG56, and CXCL10 genes induced by cytoplasmic transfected high- or low-molecular-weight (HMW or LMW) poly(I:C) was markedly increased in hnRNPM-deficient cells (Fig 2B). Consistently, phosphorylation of TBK1, IRF3, IκBα, p65 and STAT1 induced by SeV was markedly increased in hnRNPM-deficient cells in comparison to control cells (Fig 2C).
To determine whether the effects of hnRNPM-deficiency are cell type-specific, we generated hnRNPM-deficient THP1 cells. We found that transcription of the IFNB1, ISG56, and CXCL10 genes induced by SeV (Fig 2D) and transfected LMW-poly(I:C) (Fig 2E) was markedly increased in hnRNPM-deficient THP1 cells in comparison to control cells. In similar experiments, hnRNPM-deficiency had no marked effects on transcription of these downstream genes induced by transfected DNA mimics such as HSV120 (a synthetic 120-mer dsDNA representing the genomes of HSV-1), HT-DNA (herring testis DNA) and VACV70 (70-mer dsDNA representing the genomes of VACV) (Fig 2F and S1B Fig). Interestingly, we found that knockout of hnRNPM increased transcription of downstream genes induced by DNA virus HSV-1 (S1A Fig), which indicate that hnRNPM regulates HSV-1-triggered innate immune responses in a viral DNA-independent manner. These results suggest that hnRNPM negatively regulates viral RNA- but not DNA-triggered induction of downstream effector genes.
Since hnRNPM inhibits RLR-mediated signaling, we examined the roles of hnRNPM in cellular antiviral response. Previous studies have demonstrated that RIG-I is essential for the antiviral innate response to NDV and VSV, whereas MDA5 is critical for the detection of EMCV. We found that the replication of GFP-tagged NDV and VSV was markedly inhibited in hnRNPM-deficient cells compared with control cells as monitored by GFP expression (Fig 3A & 3B). Plaque assays showed that viral titers of VSV and EMCV were much lower in hnRNPM-deficient cells compared with in control cells (Fig 3C). These results suggest that hnRNPM-deficiency inhibits replication of RNA virus.
Previous studies have shown that hnRNPM is mainly located in the nucleoplasm and plasma membrane, and barely detected in the cytoplasm [33–35]. We further examined the cellular localizations of hnRNPM before and after viral infection. Confocal microscopy revealed that hnRNPM was mostly localized in the nucleoplasm in uninfected cells. However, SeV or EMCV infection resulted export of hnRNPM from the nucleus to cytoplasm (Fig 4A). Similarly, transfection of poly(I:C) also induced export of hnRNPM from the nucleoplasm to cytoplasm (Fig 4B). Consistent with confocal microscopy, subcellular fractionation analysis showed that hnRNPM was enriched in the cytosol after SeV infection. Notably, recombinant IFN-β also induced hnRNPM translocation (Fig 4C). In addition, the hnRNPM translocation after viral infection was inhibited in RIG-I knockdown cells (Fig 4D). Antiviral stress granules are reported to be a platform for the detection of viruses [36, 37]. Although it seems hnRNPM formed puncta structures in cytoplasm after virus infection, confocal microscopy experiments showed that hnRNPM was not co-localized with G3BP1 (as a marker of stress granules) (Fig 4E). These results suggest that RLR signaling activation causes export of hnRNPM from the nucleus to cytoplasm.
We next investigated the molecular mechanisms that are responsible for the roles of hnRNPM in innate immune response to RNA virus. Results of reporter assays showed that overexpression of hnRNPM had no marked effects on RIG-I- and MDA5-mediated activation of the IFN-β promoter (Fig 5A). Furthermore, knockdown of RIG-I inhibited the induction of IFNB mRNA by SeV in hnRNPM-deficient cells (Fig 5B). In transient transfection and co-immunoprecipitation experiments, hnRNPM interacted with RIG-I and MDA5, but not with MITA, VISA, TBK1 and IRF3 (Fig 5C). Overexpression of hnRNPM had no effects on interaction between RIG or MDA5 with VISA (S2A Fig). Endogenous co-immunoprecipitation experiments indicated that hnRNPM interacted with RIG-I/MDA5 in a viral-infection-dependent manner (Fig 5D). We have further produced recombinant hnRNPM, RIG-I and MDA5 (280–1,025). Vitro protein pull-down analysis showed that hnRNPM can directly interact with RIG-I or MDA5 in RNA-free condition (Fig 5E). These results suggest that hnRNPM acts at the level of RLRs.
Both RIG-I and MDA5 contain two N-terminal CARD domains, a middle helicase domain, and a CTD, whereas hnRNPM contains three RNA recognition motif (RRM) domains (Fig 5F). Domain mapping experiments indicated that the CARD and the helicase-CTD of MDA5 could independently interact with hnRNPM, while the helicase-CTD but not CARD of RIG-I was responsible for its interaction with hnRNPM (Fig 5F and S2B Fig). Furthermore, we found that deletion of an individual RRM domain of hnRNPM had no marked effects on its interaction with RIG-I or MDA5. However, all other examined deletion mutants of hnRNPM failed to interact with RIG-I or MDA5 (Fig 5F and S2C Fig). Reporter assays showed that wild-type hnRNPM and its mutants that interacted with RIG-I or MDA5 but not the other mutants inhibited SeV-induced activation of the IFN-β promoter (Fig 5G). These results suggest that the association of hnRNPM with RIG-I or MDA5 mediates its inhibition of RLR-mediated signaling.
Since hnRNPM is a heterogeneous nuclear ribonucleoprotein that contains three RRMs [38], we determined whether hnRNPM binds to viral RNA similarly as RIG-I and MDA5. Previously, it has been shown that the CTD of RIG-I binds to 5’ppp-ssRNA, 5’ppp-dsRNA, and short blunt-ended dsRNA, with significantly higher affinity for 5’ppp-dsRNA [39], whereas the CTD of MDA5 has higher affinity to long dsRNA such as synthetic poly(I:C). Pull-down experiments indicated that ectopically-expressed hnRNPM could bind to 5’ppp-dsRNA and poly(I:C) (Fig 6A). We also examined whether hnRNPM binds to viral RNA in infected cells by ‘‘footprint” experiments [30], [40]. After SeV infection (which is recognized by both RIG-I and MDA5), we immunoprecipitated hnRNPM and the immunoprecipitates were treated with RNase I. The protein-protected viral RNA was detected by RT-PCR with primers targeting various regions of SeV RNA. The results showed that hnRNPM could bind to naturally infected viral RNA similar as RIG-I and MDA5 (Fig 6B, S3 & S4 Figs). Interestingly, hnRNPM appeared preferly to bind to the 5’- terminus of SeV RNA (Fig 6B, S3 & S4 Figs) and had a higher affinity with viral RNA in the late phase of infection (S4 Fig). Confocal microscopy analysis confirmed the colocation between poly(I:C) and hnRNPM (Fig 6D). These experiments suggest that hnRNPM can directly bind to synthetic and viral RNA. Furthermore, pull-down experiments revealed that the individual RRM-deleted mutants of hnRNPM as well as hnRNPM (aa281-653) and hnRNPM (aa204-729) could bind to 5’ppp-dsRNA or SeV RNA (Fig 6C). These results indicate that RRM regions that bind cellular and viral RNA are different. Notably, although hnRNPM (aa281-653) and hnRNPM (aa204-729) could bind to viral RNA, they lost the ability to inhibit SeV-induced activation of the IFN-β promoter (Fig 5G). These findings suggest that the binding of hnRNPM to viral RNA is insufficient for regulating RLR-mediated signaling.
Finally, we investigated whether hnRNPM regulates sensing of viral RNA by RIG-I and MDA5. The ‘‘footprint” experiments showed that overexpression of hnRNPM inhibited the binding of RIG-I to SeV RNA (Fig 7A and S5A Fig). Conversely, deficiency of hnRNPM enhanced the binding of RIG-I to SeV RNA (Fig 7B and S5B Fig). Furthermore, pull-down experiments also indicated that the binding of RIG-I to 5’ppp-dsRNA or MDA5 to poly(I:C) was enhanced in hnRNPM-deficient cells in comparison to control cells (Fig 7C). We further produced recombinant hnRNPM and MDA5(280–1,025) in bacteria and immunoprecipitated Flag-tagged RIG-I from transfected HEK 293 cells for microscale thermophoresis technology (MST) experiments in vitro. The results showed that hnRNPM bound to dsRNA (25 bp) with an affinity of Kd = 22.1 ± 0.439 nM, which was higher than that of MDA5 (Kd = 131 ± 4.41 nM) and RIG-I (Kd = 69 ± 3.29 nM) with dsRNA (Fig 7D and S5C Fig). Interestingly, recombinant hnRNPM caused approximately 10- and 3-fold decrease of the affinities of RIG-I and MDA5 to dsRNA respectively (Fig 7D and S5C Fig). These results suggest that hnRNPM impairs the binding of RIG-I and MDA5 to viral RNA.
Recognition of viral RNA by RLRs is essential for the initiation of innate antiviral response initiation. Although the regulation of RLR-mediated downstream signaling have been extensively investigated, little is known about the regulatory mechanisms of the recognition of viral RNA by RLRs. Recently, we identified ZCCHC3 as a co-receptor for RIG-I and MDA5 by facilitating the sensing of viral RNA [30]. In addition, MEX3C and RAVER1 have been reported to facilitate the recognition of viral RNA by RIG-I and MDA5 respectively [28, 29]. However, how these processes are negatively regulated remains enigmatic. In the current study, we identified hnRNPM as an inhibitor of RLR-mediated innate immune response by impairing the binding of RIG-I and MDA5 to viral RNA.
Overexpression of hnRNPM inhibited SeV-triggered activation of ISRE, NF-κB, and the IFN-β promoter, while deficiency of hnRNPM had the opposite effects. In addition, the replication of RNA virus was decreased in hnRNPM-deficient cells compared with control cells. These data established a critical role for hnRNPM in innate immune response to viral RNA.
We found that hnRNPM underwent re-distribution between nucleus and cytoplasm during RNA virus infection. hnRNPM was mostly localized in the nucleus in rest cells. Following infection, hnRNPM was translocated from the nucleus to cytoplasm, which was impaired in RIG-I knockdown cells. Furthermore, hnRNPM was also translocated from the nucleus to cytoplasm in IFN-β-treated cells. These results suggest that hnRNPM translocation is dependent on RLRs.
Several evidences suggest that hnRNPM inhibits RNA virus-triggered innate immunity by antagonizing RNA sensing of RLRs. Firstly, viral infection caused export of hnRNPM from the nucleus to cytoplasm, indicating a cytoplasmic role of hnRNPM after viral infection. Second, hnRNPM was associated with RIG-I and MDA5 and their interactions were important for the functions of hnRNPM. Third, overexpression of hnRNPM inhibited the binding of RIG-I to SeV RNA. Conversely, deficiency of hnRNPM enhanced the binding of RIG-I to SeV RNA. Recombinant hnRNPM caused approximately 10- and 3-fold decrease of the affinities of RIG-I and MDA5 to dsRNA respectively. These results collectively suggest that hnRNPM is an inhibitor for RLR-mediated innate immune response by impairing viral RNA sensing by RIG-I and MDA5.
Our results revealed that the classical cellular mRNA binding regions (RRM) of hnRNPM differs from its viral RNA binding regions. The mutants containing partial RRM regions such as 1–149, 1–281 and 71–281 are lack of viral RNA binding activity while the mutant 281–653 being lack of all three RRM domains had a stronger affinity to viral RNA.
Although hnRNPM binds to dsRNA, the binding is insufficient for regulating RLR-mediated signaling. Several truncations of hnRNPM, aa281-653 and aa204-729, that could bind to RNA but failed to interact with RIG-I and MDA5, had no effects on SeV-triggered activation of the IFNB promoter. hnRNPM appeared to prefer to bind to the 5’- terminus of SeV RNA, but inhibited the binding of RIG-I to diverse regions of SeV RNA. The regions of RIG-I that interact with hnRNPM overlapped with its viral RNA binding regions. These results collectively suggest the functions of hnRNPM on RLR-mediated innate immune response are mostly dependent on its association with RLRs instead of RNA binding.
Interestingly, we also found that hnRNPM deficiency enhanced DNA virus -triggered expression of downstream genes. However, hnRNPM deficiency had no significant effects on the induction of downstream genes by transfected DNA mimics such as HSV120, HT-DNA and VACA70. These results suggest that hnRNPM regulates HSV-1-triggered innate immune responses in a viral DNA-independent manner. Previous studies have implicated that RIG-I also detected dsRNA produced by herpesviruses [41–45]. We propose a possibility that hnRNPM is partly involved in HSV-1-triggered innate immune signaling by regulating the sensing of HSV-1-derived RNA by RIG-I.
In conclusion, our study suggests that hnRNPM is a decoy of innate antiviral response by impairing the sensing of viral RNA by RLRs, which provides a critical control mechanism of viral RNA sensing for the host to avoid excessive immune response.
Poly(dA:dT), poly(I:C)-HWM, poly(I:C)-LWM, poly(I:C)-fluorescein, 5’ppp-dsRNA, RNase inhibitor, M-MLV and Lipofectamine 2000 (Invivogen); HT-DNA (Sigma); DTT (Fermentas); RNase I (Ambion); polybrene (Millipore); recombinant IFN-β (PeproTech); SYBR (Bio-Rad); dual-specific luciferase assay kit (Promega); puromycin and EZ-link psoralen-PEG3-biotin and streptavidin agarose resin (Thermo); protein G sepharose (GE Healthcare); anti-Flag affinity gel (Biomake, B73102); RNAiso plus (Takara); and recombinant IFN-β (R&D systems) were purchased from the indicated companies.
Anti-Flag (F3165), anti-β-actin (A2228) and anti-β-tubulin (T8328) were from Sigma-Aldrich. Anti-phospho-IκBα (Ser32/36) (5A5), anti-phospho-IRF3 (Ser396) (4D4G), anti-phospho-STAT1 (Tyr701) (58D6) and anti-phospho-p65 (Ser536) were from Cell Signaling Technology. Anti-LMNB1 (12987-1-AP) was from ProteinTech. Anti-G3BP1(210-323aa) (611126) was from BD Biosciences. Anti-HA (16B12) was from Covance. Anti-TBK1 (ab109735), anti-phospho-TBK1 (Ser172) (ab109272) were from Abcam. Anti-IRF3 (FL-425), anti-p65 (C-20) (sc-372), anti-STAT1 (C-111) were from Santa Cruz Biotechnology. Goat anti-mouse IgG (R37116) or donkey anti–rabbit (R37119) conjugated to Alexa Fluor 594 and goat anti-mouse IgG (A-10684) conjugated to Alexa Fluor 488 were purchased from Thermo Fisher. Mouse antisera against hnRNPM, IκBα, RIG-I and MDA5 were raised against purified recombinant human hnRNPM (1-729aa), IκBα, RIG-I (1-200aa) and MDA5 (1-200aa). Rabbit antisera against SeV was raised against purified SeV.
Human embryonic kidney 293 (HEK293, CRL-1573), Human acute monocytic leukemia cell line (THP-1) and Henrietta Lacks (HeLa, CCL-2) cells were purchased from American Type Culture Collection, and Vero cells were purchased from China Center for Type Culture Collection (Wuhan, China). HEK293T cells were originally provided by G. Johnson (National Jewish Health, Denver, CO). The strains (BL21 and DH5α) were purchased from ATCC.
SeV, VSV (Indiana Strain), NDV and HSV-1 were previously described [46, 47]. EMCV was provided by Dr. H. Yang (China Agricultural University).
The following oligonucleotides were used to stimulate cells:
VACV70: 5’-CCATCAGAAAGAGGTTTAATATTTTTGTGAGACCATGGA
AGAGAGAAAGAGATAAAACTTTTTTACGACT-3’;
HSV120: 5’-AGACGGTATATTTTTGCGTTATCACTGTCCCGGATTGGAC
ACGGTCTTGTGGGATAGGCATGCCCAGAAGGCATATTGGGTTAACCCCT
TTTTATTTGTGGCGGGTTTTTTGGAGGACTT-3’.
RIG-I or MDA5 and their mutants, Flag or HA-tagged hnRNPM and their truncation mutants, pGEX6p-1-GST-hnRNPM and pGEX6p-1-GST-MDA5 (280–1025) were constructed by standard molecular biology techniques. The other expression and reporter plasmids were previously described [9].
HEK293 cells were transfected by standard calcium phosphate precipitation method. To normalize for transfection efficiency, 0.01 μg of pRL-TK (Renilla luciferase) reporter plasmid was added to each transfection. Luciferase assays were performed using a dual-specific luciferase assay kit. Firefly luciferase activities were normalized on the basis of Renilla luciferase activities.
Double-stranded oligonucleotides corresponding to the target sequences were cloned into the pSuper. Retro-RNAi plasmid (Oligoengine).
The following sequences were targeted for hnRNPM mRNA.
The hnRNPM-shRNA#1 targeting sequence: 5’-GGCATAGGATTTGGAATAA-3’
The hnRNPM-shRNA#2 targeting sequence: 5’-GCAATCGCTTTGAGCCATA-3’.
The RIG-I-shRNA targeting sequence: 5’-CCGTGATTCCACTTTCCTG-3’
HEK293 or THP-1 cells were transduced with plentiCRISPRv2-hnRNPM-sgRNA viruses for five days. Puromycin-resistant individual clones were selected and analyzed by immunoblotting to determine the efficiency of hnRNPM knockout.
The hnRNPM-gRNA#1 targeting sequence: 5’-GGCGACGGAGATCAAAATGG-3’.
The hnRNPM-gRNA#2 targeting sequence: 5’-GGCGGCGACGGAGATCAAAA-3’.
Total RNA was isolated for qPCR analysis to measure mRNA abundance of the indicated genes. Data shown are the relative abundance of the indicated mRNA derived from human cells normalized to that GAPDH respectively. Gene-specific primer sequences were as described [46, 47].
Q-PCR primers for HNRNPM:
Forward Sequence: TCCTGAACGCCCACAGCAACTT
Reverse Sequence: TGCCTTTGCTCAGATGGTTGGC
Cells were lysed in NP-40 lysis buffer (20 mM Tris-HCl (pH 7.4), 150 mM NaCl, 1mM EDTA, 1% NonidetP-40, 10 μg/ml aprotinin, 10 μg/ml leupeptin, and 1 mM phenyl-methylsulfonyl fluoride). The lysates were subjected to immunoprecipitation and immunoblotting analysis with the indicated antibodies.
Nuclear and cytoplasmic fractions extraction of HEK293 cells were generated according to the instruction of Nuclear and Cytoplasmic Protein Extraction Kit (PPLYGEN, P1201).
Poly(I:C) and 5’ppp-dsRNA was conjugated to biotin by UV (365 nm wave-length) cross-linking for 1 hour. HEK293 cells transfected with the indicated plasmids were lysed in Pre-lysis buffer pre-treated with DEPC. Lysates were incubated with biotinylated-5’ppp-dsRNA or biotinylated-poly(I:C) for 2 hours at room temperature, and then incubated with streptavidin beads for another 1 hour at room temperature. The beads were washed four times with lysis buffer and analyzed by immunoblots with the indicated antibodies.
The pGEX-6p-1-GST plasmids encoding hnRNPM and MDA5(280–1025) were transformed into BL21 competent cells. Expression of the proteins was induced with 0.1 mM IPTG at 16°C for 24 hours. The proteins were purified with GST resins and eluted with elution buffer (PBS, 100 mM Tris-HCl pH 8.8, 40mM reduced glutathione). To obtain purified RIG-I, Flag-RIG-I plasmid was transfected into HEK293 cells by standard calcium phosphate precipitation method. The expressed Flag-RIG-I protein was immunoprecipitated with anti-Flag affinity gel and eluted with 3×Flag peptides.
Purified recombinant protein hnRNPM and MDA5/RIG-I were incubated in PBS with RNase I at 4°C for 3 hours. Lysates were respectively immunoprecipitated with anti-MDA5 and protein G beads (50 ul) or anti-Flag affinity gel at 4°C for another 2 hours. GST was used as negative control. Bound proteins were analyzed by immunoblots with the indicated antibodies.
MST analysis was performed using a NanoTemper Monolith NT.115 instrument (NanoTemper Technologies GmbH). For detecting affinity between dsRNA and GST-hnRNPM, Flag-RIG-I or GST-MDA5, 20 nM Cy5-labeled 25 bp dsRNA (Sangon Biotech, China) was mixed with different concentrations of proteins in PBS with 100 mM Tris-HCl (pH 8.8). GST was used as negative control. Samples were loaded into Premium Coated Capillaries and MST measurements were performed using 20% MST power and 40% LED power at 25°C. Laser-on and -off times were 30 and 5 s respectively. NanoTemper Analysis 1.2.20 software was used to fit the data and to determine the apparent Kd values.
HEK293 cells were transfected with the indicated HA- or Flag-tagged plasmids for 20 hours, then infected with SeV for 1 hour, washed with medium and cultured for 2 more hours. Cell lysates were immunoprecipitated with IgG or anti-HA (2 μg) and protein G beads (50 μl) at 4°C for 3 hours. The immunoprecipitates were treated with diluted RNase I (1:25 in PBS) at 37°C for 5 min. The bead-bound immunoprecipitates were washed for 3 times with lysis buffer containing RNase inhibitors. The protein and RNA complexes were eluted with 200 μL TE buffer containing 10 mM DTT at 37°C for 30 min. The RNA was extracted using Trizol reagent before qPCR analysis for SeV RNA. The SeV genome primer sequences were described in S1 Table.
Confocal microscopy was performed as previously described [48]. Briefly, cells infected with virus or transfected poly(I:C) for the indicated times were fixed with 4% paraformaldehyde for 10 min at 25°C and then permeabilized and stained with indicated antibodies by standard protocols. The stained cells were observed with a ZEISS confocal microscope under a 100× oil objective.
Host cells (5×105) cultured in 12-well plates were infected with viruses at the respective MOI for 1 hour, then washed with PBS and cultured with 1 ml fresh complete medium. The plates were incubated for 36 hours post-infection at 37°C, 10% CO2. The media were collected and used for plaque assays on monolayers of Vero cells seeded in 24-well plates. The Vero cells were infected by incubation for 1 hour at 37°C with serial dilutions of collected media. After infection for 1 hour, the cells were overlaid with 1.5% methylcellulose and then incubated for about 48 hours. The overlay was removed, and cells were fixed with 4% paraformaldehyde for 15 min and stained with 1% crystal violet for 30 min before plaque counting.
Unpaired Student’s t test was used for statistical analysis with GraphPad Prism Software. p < 0.05 was considered significant.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.